ARGOS MCP DEV

mcp.staticpipe.com · 178 tools registered

What's new RSS

"#2487 new sedar family (SEDAR+ & Canada): sedar_issuer, sedar_filings, canada_feed, mining_projects, sedar_document"

sedar_issuer sedar_filings canada_feed mining_projects sedar_document mining_company_reserves

"#2437 catalyst_calendar: past-date lock-up queries by ticker return rows"

catalyst_calendar

"#2451 sign-in: approve each app on a consent page; stricter app registration"

"#2350 mars_company_resolve: all-caps names like NVIDIA resolve"

mars_company_resolve

"#2340 #2290 supply_chain whats_new: recent news on existing links (dev only)"

supply_chain

"#2315 cascade signal tools retired; get_company_profile is now a short identity + fundamentals card"

get_company_profile company_filings screen_companies compare_companies get_sector_snapshot

"#2300 company_proxy_votes: director support per nominee, withheld votes shown"

company_proxy_votes

"#2280 search_activist_positions: 13D swaps and derivatives, economic exposure"

search_activist_positions

"#2293 run_script: call ARGOS tools from a Python script, get back only the result (dev only)"

run_script

"#2290 supply_chain: semiconductor and AI supply chain (dev only)"

supply_chain

"#2287 follow-up: adviser CIK in the Form ADV and proxy-voting tools; funds_managed cites each fund's N-PORT"

adv_custodians adv_manager_disciplinary_detail adv_manager_fees adv_manager_people adv_manager_profile adv_manager_raum_trend adv_private_funds adv_raum_decliners adv_screen_managers manager_proxy_voting funds_managed

"#2287 fund tools: each adviser's SEC CIK beside its CRD"

funds_managed fund_equity_holders

"#2281 activism tools no longer fail on curly quotes in 13D text (XPOF)"

activism_campaigns search_activist_positions

"#2255 citations, batch 2: blocks that lacked a source now cite it"

screen_fundamentals search_activist_positions get_company_profile bio_cash_runway manager_proxy_voting credit_bdc_loan_book credit_borrower_financing_history donor_person_profile donor_person_track_record capital_structure holder_profile chart_events sector_detail bio_catalyst_dossier

"#2255 citations, batch 1: 34 more tools cite the source filing of every figure"

screen_companies compare_companies thirteenf_portfolio thirteenf_portfolio_full thirteenf_who_holds thirteenf_overlap thirteenf_search_managers thirteenf_position_history adv_manager_profile adv_manager_raum_trend adv_screen_managers adv_manager_fees adv_raum_decliners adv_manager_disciplinary_detail adv_manager_people adv_custodians bio_crl_outcomes tranche_bond_holders fund_equity_holders fund_equity_book funds_managed fund_vote_matrix company_fund_votes bdc_summary mars_mergers_screen mars_mergers_for_company mars_trending_acquirers company_address_history insiders_of_company exec_compensation donor_liquidity_events insider_history insider_signal insider_holdings holders_combined

"sedar_search: search Canadian SEDAR+ filings again"

sedar_search

"#2189 ATM usage: is the company selling stock through its at-the-market program right now?"

capital_structure bio_cash_runway company_dilution

"#2168 13F: securities joining the 13F list no longer read as a wave of buying"

thirteenf_smart_money_consensus thirteenf_screen thirteenf_cohort_flow thirteenf_holders_full thirteenf_holders thirteenf_position_history holder_profile

"get_fundamentals: roic on every row, beta vs SPY"

get_fundamentals

"#2224 resolve_people: email domains reach the public parent"

resolve_people

"#2218 resolve_people: resolve a list of people in one call; resolve shows company and title"

resolve_people resolve

"#2219 funds_managed and other CIK-input adviser tools: ADV CIK fallback"

funds_managed manager_proxy_voting adv_custodians holder_profile

"#2193 catalyst_calendar: unresolved PDUFAs, held AdComs and AdCom votes"

catalyst_calendar

"#2187 13F holders: biotech specialist flag"

thirteenf_holders thirteenf_holders_full holders_combined deal_work

"get_fundamentals: total assets and total liabilities"

get_fundamentals

"#2170 sec_power_rate_cases: as_of leaves out non-rate-case rows"

sec_power_rate_cases

"#2165 mars_search_articles: filter news by company"

mars_search_articles

"#2164 get_fundamentals: R&D, current assets and liabilities, and annual rows"

get_fundamentals

"Notable-fund figures on past quarters: look-ahead stated"

thirteenf_smart_money_consensus thirteenf_screen

"Replay past quarters and dates on the 13F and activist tools"

thirteenf_screen thirteenf_smart_money_consensus search_activist_positions

"Fund type on every fund: ETF, mutual fund, closed-end, money market, interval"

funds_managed fund_equity_holders fund_equity_book

"Faster first calls: discover_tools and holder_profile"

discover_tools holder_profile

"Sign-in button responds at once and sends one email"

"Company results no longer include government bodies or placeholder names"

mars_company_resolve resolve graph_resolve mars_mergers_for_company mars_funding_history company_private_capital mars_trending_acquirers mars_funding_search mars_sector_funding_flow mars_investor_portfolio credit_borrower_financing_history donor_person_track_record

"Deal Desk, Work this deal, the merger votes calendar and Holder Profile (REST API parity)"

deal_desk deal_work merger_votes_calendar holder_profile

"Derivatives on a stock, derivative counterparties, ADV custodians and the fund vote matrix (REST API parity)"

company_derivative_exposure derivative_counterparties counterparty_exposure adv_custodians fund_vote_matrix

"Tool errors no longer show raw database messages"

all

"company_fund_votes: both layouts count a vote once, with Console's key"

company_fund_votes

"company_fund_votes matrix counts a repeated vote row once"

company_fund_votes

"The full N-PORT book: fund_derivatives, non-CUSIP positions and a fund's asset mix"

fund_derivatives tranche_fund_portfolio

"company_fund_votes rollup='family' names families by brand, as Console does"

company_fund_votes

"company_fund_votes matrix follows Console's rules: every N-PX row counts, All nominees column, minor columns flagged"

company_fund_votes

"Company and filings tool descriptions in plain language"

get_company_profile screen_companies company_filings company_red_flags discover_tools earnings_surprise_history upcoming_earnings catalyst_calendar mars_company_resolve resolve_company_name sec_comment_letters sections_filing_change_screen sections_ipo_risk_factors sections_ipo_offering sections_ipo_financials sections_ipo_screen get_filing resolve resolve_company_profile

"resolve matches person name variants in any letter case"

resolve

"company_fund_votes no longer counts a fund vote twice"

company_fund_votes

"company_fund_votes adds a vote matrix: each fund adviser's vote on every ballot item and director nominee"

company_fund_votes

"People tools return paging at the top level; superseded resolvers unlisted; plainer holdings text"

names_search names_profile names_transactions names_affiliations names_vesting_calendar names_990_boards insider_holdings resolve discover_tools mars_company_resolve resolve_company_name names_lookup

"resolve finds people by name"

resolve graph_resolve

"names_search returns people only by default; names_vesting_calendar takes a CIK"

names_search names_vesting_calendar

"People tools no longer return internal backend details"

names_search names_profile names_transactions names_affiliations names_vesting_calendar names_990_boards names_lookup

Responses from the people (names_*) tools are cleaned before they reach you (#1730). Internal hostnames and IPs, search-index details and source table names are removed from _meta. A per-source error inside _meta now reads unavailable instead of the raw backend message. The data rows are unchanged.

names_990_boards now rejects a parameter it doesn't support (for example cik) with a bad_request that lists the supported ones. Before, it dropped the parameter and returned the unfiltered result.

"Activism and holders tools count genuine activists (is_engaged_activist), not every 13D"

sector_monitor sector_detail search_activist_positions chart_events holders_combined

A 13D says a holder is above 5% and not passive. It does not say they are an activist: control holders, founders and insiders, PE sponsors and strategics file 13Ds too. smart_money now classifies each position (is_engaged_activist, activist_type), and Console counts on that flag. The MCP now does the same (#1525):

"activism_campaigns: a proxy fight takes its company's contested election result"

activism_campaigns

A campaign's status still comes only from a withdrawal, a settlement / cooperation agreement or a contested election result, never from another event's outcome text (#1619). A contested election decides every slate on the ballot, though, and mars files the result under one activist. A proxy fight with no outcome of its own now takes its company's election result when that result falls within the fight or up to 120 days after it. outcome_source gives the result's date, accession and the activist it was filed under. Disney / Blackwells now reads the 2024-04-03 election result, where Disney's slate beat Trian's and Blackwells'.

Activist-mode lookups (activist=...) also no longer time out on a cold cache.

"sector_monitor / sector_detail: Form D and activism windows include their first day"

sector_monitor sector_detail

The Form D window now includes its start date, as Console's does: a 90-day window run on 2026-10-22 starts at 2026-07-24 inclusive. Information technology's 90-day Form D now matches Console exactly, at 808 filings and $10,534,775,831 (#1525).

"sector_monitor / sector_detail: Form D, activism and movers now use Console's definitions"

sector_monitor sector_detail

Three blocks now count exactly the way Console does (#1525):

The window now changes only insider flow and Form D.

"mars_reit_fundamentals adds re-leasing spreads and reads verified rows"

mars_reit_fundamentals

mars_reit_fundamentals now returns leasing_spread_cash_pct and leasing_spread_gaap_pct for each period. These are the cash and GAAP re-leasing spreads on new and renewal leases, mostly reported by office, retail and industrial REITs. The per-REIT coverage block already says whether a REIT reports them (leasing_spread). The default view also includes verified rows alongside vouched, so rows reit later promotes to verified stay visible. This aligns the tool with reit's spec v1.1 (#1487).

"sector_monitor and sector_detail: Console's Sector Monitor and sector / industry pages"

sector_monitor sector_detail

Two new tools in the intelligence family give the same numbers as Console's /sectors pages (#1525). Both read Console's nightly sectors base, so the universe, SIC-based sectors, returns and valuation match Console exactly:

Information technology over 1Y is +44.07% cap-weighted across 701 companies, the same as Console.

"mars_search_articles moves from overview to its own news family"

mars_search_articles

mars_search_articles now sits in a new news family instead of overview (#1516). News is gated until it is priced, so callers without the news family no longer see it. Full-catalog callers still see it.

"earnings_surprise_history accepts a CIK, so delisted companies are reachable"

earnings_surprise_history

Since earnings' 10-21 refresh (#1605), each ticker maps to exactly one company, and rows for delisted or acquired companies carry no ticker. Those companies were unreachable by ticker, so earnings_surprise_history now also takes cik (a numeric value in ticker works too). For a dead company, ticker comes back null and former_tickers lists the symbols it traded under (Digimarc, CIK 1438231 → DMRC). Ticker lookups are unchanged.

"Insider dollar totals and counts no longer double-count joint Form 4 refilings"

chart_events insider_sector_pulse insider_history insider_flow insider_cluster_screen

When a trade is reported in several Form 4s (a fund and its GP, Magnetar-style joint refilings), every insider tool now counts it once (#1555):

"New `research` family (Research & Provenance); `filings` folds into `overview`; company_people moves to `people`"

search_classifiers get_sentences semantic_search sections_filing_change_screen company_red_flags screen_companies sections_ipo_screen sections_ipo_risk_factors sections_ipo_offering sections_ipo_financials company_filings sec_comment_letters mars_search_articles company_people

Tool families now match the pricing v3 module list (#1516):

No tool names or inputs changed.

"company_private_capital reports Form D sold amounts; comps / capital_structure explain missing TTM figures"

company_private_capital comps capital_structure

"activism_campaigns: activist campaign timelines for a company or an activist"

activism_campaigns

New tool activism_campaigns (Kscope #1526, Console Corporate Activism, activism family).

"company_private_capital: a company's Form D offerings, funding rounds and Reg CF in one call"

company_private_capital

New tool company_private_capital(company | ticker_or_cik | company_id) (Kscope #1524, Console Private Capital, private family). Works for private companies with no ticker.

"comps: trailing valuation and operating metrics against peers"

comps capital_structure

New tool comps(ticker_or_cik, peers?, n_peers?, as_of?) (Kscope #1523, Console Comps, fundamentals family).

capital_structure's TTM block now includes dividends_paid.

"sec_comment_letters: SEC staff reviews with the company's answers, its counsel, and a market screen"

sec_comment_letters

New tool sec_comment_letters (Kscope #1518, Console Research, filings family).

Every letter carries an accession to open with get_filing. data_through shows the latest staff letter loaded, currently 2026-08-27.

"company_people: current officers and directors, rebuilt from filings"

company_people

New tool company_people(ticker_or_cik, include_former?, as_of?) (Kscope #1519, Console MGMT, overview family). It does not trust stored "current" flags. It rebuilds the roster from the latest proxy board slate (independence, chair / lead independent, committees), insiders with a Form 4 in the last 13 months, and the proxy's named executive officers. Every later 8-K Item 5.02 appointment or departure is then applied in date order, so HON shows the 2026-06-13 spin-off resignations and AAPL shows Ternus as CEO with Cook as executive chair. Each person has titles, director-since, age, latest-year pay, Form 4 shares and sources with accessions. Announced changes that are not yet effective are listed under pending. include_former=true lists departures from the last 3 years.

"New: geographic_revenue gives revenue by country and region as filed, with overlaps removed"

geographic_revenue

geographic_revenue(ticker_or_cik) returns the same data as Console's Geographic Revenue report. It is in the fundamentals family. It reads the geographic revenue companies tag in their 10-K, 10-Q and 20-F XBRL, per fiscal year (default 5) or per quarter (period="quarter"). Nothing is estimated or gap-filled.

"New: capital_structure shows market cap to EV, TTM earnings, leverage and coverage, with 8 quarters of history"

capital_structure

capital_structure(ticker_or_cik) returns the same data as Console's Capital Structure report. It is in the fundamentals family.

"New: material_contracts lists a company's filed agreements with parties, terms and clause types"

material_contracts

material_contracts(ticker_or_cik) returns the same data as Console's Material Contracts report. It is in the fundamentals family. It lists every EX-10 material contract and EX-2.1 M&A agreement the company filed or signed as a party, newest first.

"New: get_filing opens any filing by accession — main text, an exhibit, a section, or matching passages"

get_filing

Several tools (company_deals, company_red_flags, company_dilution, insider_liquidity, company_proxy_votes) return filings by accession and say "open one with get_filing". That tool now exists, in the overview family. It reads Kaleidoscope's own filing store, never sec.gov.

"company_deals: role+status filters no longer error, Closed always dated, sales shown as seller, 8-Ks only from the deal"

company_deals

Four fixes to company_deals:

"earnings_surprise_history: era_ticker per quarter + former_tickers (FB rows under META)"

earnings_surprise_history

earnings_surprise_history still looks rows up by the current ticker, so META returns every quarter, including the Facebook years. Each quarter now also carries era_ticker, the symbol traded at that quarter's end date ("FB" on Meta's rows through 2022-03-31). The response lists earlier symbols in former_tickers. The fields appear once earnings serves the new column (their 2026-10-21 09:30 UTC run); the tool checks for it every 10 minutes, with no restart needed. Kscope #1506 follow-up.

"thirteenf_screen no longer times out on the first run after a holdings table rebuild"

thirteenf_screen

thirteenf_screen could return internal_error when it was the first call after smart_money swapped in a rebuilt holdings table while proc2 was busy. Its count and ranking queries now get a 10-second database timeout instead of 3 seconds, like the thirteenf_smart_money_consensus leaderboard (#1496). Kscope #1465 follow-up.

"thirteenf_smart_money_consensus outflow no longer times out; thirteenf_holders counts lower-case CUSIP filers"

thirteenf_smart_money_consensus thirteenf_screen thirteenf_holders thirteenf_holders_full holders_combined

Leaderboard timeouts (#1496). Right after smart_money swaps in a rebuilt table, the first thirteenf_smart_money_consensus leaderboard calls (outflow at limit 50 or 60) could pass the 3-second database timeout and return internal_error. The fund and new-security counts now run as one query and fail soft: if it times out, the counts come back null with a warning, and the filters still apply. The leaderboard query itself gets a 10-second timeout. thirteenf_screen counts the same way.

Lower-case CUSIPs (#1497). A ticker lookup in thirteenf_holders / thirteenf_holders_full now includes filers who reported the CUSIP in lower case, as a CUSIP lookup already did. COST Q2 2026 lists 4,259 holders, matching thirteenf_screen (was 4,216).

"13F leaderboards exclude funds and new securities by default (is_fund / is_new_security), with counts and include_* switches"

thirteenf_smart_money_consensus thirteenf_screen

The thirteenf_smart_money_consensus leaderboard and thirteenf_screen now leave out funds by default: ETPs, open-end, closed-end and funds-of-funds, using smart_money's new holdings_company_quarterly.is_fund. Untyped securities (is_fund NULL, mostly foreign common stock) stay in. Each response reports n_funds_filtered, and include_funds=true brings funds back. exclude_funds still works: exclude_funds=false is read as include_funds=true. In thirteenf_screen, a ticker_in list isn't fund-filtered unless you pass include_funds=false, so asking for SPY still returns SPY. Securities flagged is_new_security (IPOs and CUSIP successions, where every holder reads as a new buyer) are excluded by default too, reported as n_new_securities_filtered, with include_new_securities=true to rank them. Q2 2026: VUG, IVV, IWF, SPCX and HONA no longer top the buys. Kscope #1465.

"names_transactions: date_start/date_end now filter the trade date (inclusive), date_field switch for filing date"

names_transactions

names_transactions date_start/date_end used to filter the Form 4 filing date, so a one-day window on the trade date (BBD 2026-09-18, filed 09-21/22) returned nothing. They now filter the trade date (transaction_date), both ends inclusive. date_field='filing' keeps the old filing-date behaviour, and the response echoes date_field. Kscope #1491.

"New tool mars_reit_fundamentals: REIT FFO / AFFO / NOI / occupancy / WALT per fiscal period"

mars_reit_fundamentals

mars_reit_fundamentals returns one REIT's operating fundamentals as an ascending per-period series: FFO (+ per share), AFFO, NOI, same-store NOI, occupancy, same-store occupancy, debt maturing in 12 months and WALT. It's the operating companion to mars_reit_deal_flow. Data is reit's Track-2 extraction from earnings 8-K supplementals: 4,525 rows across 222 REITs, quarterly from 2020.

Family: fundamentals. Example: Realty Income (O) 2026-Q2 core FFO $996.6M ($1.07/share), AFFO $1.02B, NOI $1.38B, occupancy 98.8%, WALT 8.6 years.

"Insider tools: gift class with estimated values labeled, local-currency prices flagged, canonical SIC labels"

insider_history names_transactions insider_sector_pulse

Gifts (#1457). Form 4 code G rows are now class gift (they were other), with gift_kind = gift or intra_owner_transfer.

Local-currency prices (#1447). insider_history rows carry price_market_ratio (reported price ÷ market close). Where it falls outside 0.67–1.5, value_note says "reported price in local currency; value_usd converted". Example: BBD's 2026-10-15 buy is filed at R$17.98, about 5.2× the USD close, and value_usd is $219K.

SIC labels (#1467). insider_sector_pulse now uses insider's canonical upper-case label for each SIC code (e.g. "STATE COMMERCIAL BANKS"); our own SEC table only fills gaps. Null sectors stay null. Junk filed tickers (NONE, N/A) are replaced upstream with the issuer's real ticker.

"search_classifiers: the `source` parameter's human/auto counts are now read live"

search_classifiers

The source parameter's description used to hard-code 159 human + 370 auto, which totals 529. The registry now holds 531, so the split is read from the live registry: 161 human + 370 auto. It matches the classifier count in the tool description and in /meta. (Kscope #1473)

"insider_holdings / holders_combined: relabeled and stale vehicle lines labeled; former insiders use insider's owner_is_current"

insider_holdings holders_combined

Insider holdings lines now carry insider's line-status flags.

New row fields: superseded_by_nature, missed_later_filings, owner_last_filing_date, owner_is_current, line_status. (Kscope #1466)

"thirteenf_smart_money_consensus: signal needs flow and breadth to agree; new mixed_dollar_buying / mixed_dollar_selling"

thirteenf_smart_money_consensus

signal now weighs breadth as well as dollar flow. It used to be the sign of notable net flow alone. GOOGL 2026-Q2 read "net_buying" on +$62.5B even though 226 notable funds trimmed or exited against 182 that added. A few large buyers had outweighed most holders.

Every per-ticker response and leaderboard row now carries:

The all-holders block gets its own breadth and signal the same way. Per-ticker responses also add notable_value_change_usd, the price-inclusive change in the cohort's position value (GOOGL +$261B), labeled so it isn't read as buying.

2026-Q2 results: GOOGL, GOOG and AAPL are mixed_dollar_buying; NVDA, AMZN and MSFT are net_buying. (Kscope #1456)

"/meta: public_total leaves out prospect by default and names the unfamilied tool"

/meta now leaves the prospect family (9 donor/DAF tools) out of public_total by default. prospect isn't a Console module. On today's catalog that takes public_total from 150 to 141, and excluded_families shows ["prospect"]. ?exclude=a,b still overrides the default, and ?exclude= (empty) excludes nothing.

The response also adds unfamilied_tools, which lists the always-on tools that belong to no family. Today that's just resolve, the shared ID resolver. Every user sees it, so it stays in public_total. (Kscope #1476)

"holders_combined Form 4 rows: one labeled, priced line per vehicle; emptied vehicles and former insiders dropped"

holders_combined insider_holdings

holders_combined: Form 4 lines are labeled and priced. Each insider row is one holding vehicle, with ownership = direct or indirect: <vehicle>, for example Huang's indirect: By Trust 467.7M, direct 70.1M, indirect: By Limited Liability Company 3 30M and indirect: By Irrevocable Remainder Trust 109M. value_usd is shares × the last close, and value_basis names the close and its date. Two kinds of line are now dropped, and each drop is counted in warnings:

For NVDA that drops 5 emptied vehicles and Burgess, Ochoa, Perry and McCaffery. insider_holdings rows now carry nature_of_ownership. (Kscope #1466)

"thirteenf_holders / thirteenf_holders_full accept a CUSIP alone"

thirteenf_holders thirteenf_holders_full

thirteenf_holders accepts cusip without a ticker. thirteenf_holders(cusip="436440101", report_date="2026-06-30") returned "ticker is required"; it now returns the holders. This reaches lines that have no ticker mapping, such as new share classes and retired CUSIPs (HON 438516106: 153 holders at 2026-06-30). thirteenf_holders_full takes the same cusip parameter. CUSIPs are upper-cased on input, and the lower-case spelling is also matched until the 13F data is normalized (22160k105 → COST, 4,258 holders). (Kscope #1484)

"search_activist_positions ownership_summary counts current stakes only"

search_activist_positions

ownership_summary now sums only current >5% stakes. Previously it summed every filer's last 13D/G back to 2000. For NVDA that gave 48.54% "passive" across 13 filers, including 2002-era 13Gs, a zeroed-out Vanguard amendment and FMR at 4.07%. A stake now counts only when latest_percent >= 5, latest_shares > 0, and it was last filed within 3 years of as_of. The window is 3 years because since the 2024 SEC rule change a 13G is amended only on a material change, so a steady holder can go quiet for over a year. Everything else is counted in the new n_dormant_filers and is not summed; those positions are still listed. The criteria are echoed in current_criteria. subject_cik is now filled even when you query by ticker. NVDA: 14.61% across 2 current filers (Vanguard Capital Management 7.31%, BlackRock 7.30%), 11 dormant. (Kscope #1482)

"Aggregate insider and 13F tools now trace to filings: insider_flow include_participants, cluster buyers with accessions, 13F holder accessions, drilldown hints"

insider_flow insider_cluster_screen thirteenf_holders thirteenf_holders_full insider_sector_pulse thirteenf_smart_money_consensus thirteenf_screen insider_proven_buys

For consumers: "insider_flow takes include_participants (and participants_limit, default 10): the top insiders by value with insider_cik, name, title, side, value_usd, n_transactions and up to 5 recent accessions. AAPL 180d: Levinson $86.7M sell, Cook $63.8M planned sale. insider_cluster_screen clusters now carry buyers (buyers_per_cluster, default 10; 0 omits) with their accessions. thirteenf_holders / thirteenf_holders_full holders carry accessions: the 13F filing(s) for the quarter, newest first, mapped through sub-CIKs (Vanguard lists all 9 sub-CIK filings). insider_sector_pulse, thirteenf_smart_money_consensus, thirteenf_screen and insider_proven_buys have no per-trade accessions in their source, so they return a drilldown field naming the tool call that does."

"Insider flow counts each economic trade once: names_transactions distinct_events (default on), insider_flow and insider_cluster_screen dedupe joint filings"

names_transactions insider_flow insider_cluster_screen

For consumers: "A joint Form 4 (a fund and its GP, say) reports one sale under several owners. These tools now count it once. names_transactions takes distinct_events (default true): Raul's sells >$100M since 2026-07-01 drop from 45 rows / $11.05B to 35 rows / $8.75B; pass false to see every reporting owner. Rows also carry same_event_group_id, is_primary_in_group, value_estimated and gift_kind. insider_flow and insider_cluster_screen buyer/seller counts and dollar sums now use one row per event: LTH's 2026-08-10 sale reads 1 seller / $220.9M (was 3 / $662.8M). insider_sector_pulse was already deduplicated; insider_history stays row-level."

"BREAKING RENAME: thirteenf_cohort_flow n_funds_in_cohort_holding is now n_funds_in_cohort; thirteenf_screen exclude_corporate_actions covers CUSIP successions and holder collapses" BREAKING

thirteenf_cohort_flow thirteenf_screen

For consumers: "Breaking: thirteenf_cohort_flow rows renamed n_funds_in_cohort_holding to n_funds_in_cohort; update any code reading the old name. thirteenf_screen exclude_corporate_actions=true now also drops rows flagged is_new_security (a new CUSIP succeeding an old one) or is_holder_collapse, and each row reports both flags. Today it drops spin-offs such as HON, DD and CCL; once smart_money's 2026-10-20 rebuild fills the new flags it will also catch splits and take-private exits (KLAC, CTRA, HOLX)."

13F — all-holders share-based consensus, fund exclusion, notable exits on the screen, named cohorts and exits in cohort_flow

thirteenf_smart_money_consensus thirteenf_screen thirteenf_cohort_flow

For consumers: thirteenf_smart_money_consensus takes universe=notable (default) or all_holders; all_holders ranks on the share-based net_change_value across every 13F filer (GOOGL Q2 = +$42.6B) with n_increased/n_decreased breadth, and per-ticker calls now return an all_holders block beside the notable one. exclude_funds=true drops ETPs and open-end, closed-end and fund-of-funds securities (SPY, IVV and similar) on both the consensus tool and thirteenf_screen. thirteenf_screen rows add n_notable_exits, notable_net_value_usd, n_increased, n_decreased and security_type, with a n_notable_exits_min filter and those fields as sort keys. thirteenf_cohort_flow no longer requires cohort_ciks; build a named cohort with cohort_style (activist, value, quant...), cohort_tier (top_50, top_200, long_tail), cohort_manager_type and notable_only (capped at 1,000 managers, flagged when truncated). include_exits (default true) counts exiting managers' sold shares in net flow; each row carries n_new, n_increased, n_decreased, n_exits and n_unchanged. The field n_funds_in_cohort_holding is renamed n_funds_in_cohort.

insider_sector_pulse — sort order, pagination, totals, 2-digit SIC and division rollups

insider_sector_pulse

For consumers: insider_sector_pulse takes order (desc default, asc puts the heaviest net-selling sectors first), offset for paging, and sort=sell_usd. Every response carries total_sectors, has_more and a totals block (buy/sell/net USD and issuer counts summed across all sectors, not just the page). New rollup param groups at 4-digit SIC (default), 2-digit major_group (each row also names its division) or division letter A-J. Coarser rollups and custom windows run live in ~0.1-0.2s; pct_issuers_with_buying is now filled on live queries too (it was null).

names_transactions issuer filter; live tool, classifier and insider-corpus counts

names_transactions search_classifiers insider_flow insider_cluster_screen insider_sector_pulse

For consumers: names_transactions takes issuer_cik and issuer_ticker (comma-separated lists OK) to return every insider trade in a company's stock; cik still means the insider. Each row carries primary_owner_cik and joint_filing — count a joint filing once per primary_owner_cik. search_classifiers now states the live classifier count (531), and the three insider tools state the live corpus size (11.7M Form 4/5 transaction lines from ~4.6M filings) instead of the stale 529 and 10.2M. New GET /meta returns tool counts generated from the live catalog (total, public_total, by_family, classifiers); ?exclude=prospect drops whole families from public_total.

Sign-in flow rebranded from ARGOS to Kaleidoscope MCP

For consumers: The OAuth magic-link sign-in pages (/signin, "Check your email", confirm, error) and the sign-in email now carry the Kaleidoscope logo and Console-login styling, and say "Kaleidoscope MCP" instead of ARGOS — subject, sender name and body. No change to the flow itself, the 15-minute link expiry, or the allowlist.

company_deals — non-USD merger values no longer shown as US dollars

company_deals

For consumers: News-sourced deals now take their value from the FX-converted deal value. They fall back to the raw value only when the deal's currency is USD or not recorded. Before this fix, a non-USD deal with no converted value (20 rows in merger_deals_v2, e.g. a €320M deal) showed its native-currency amount as dollars; it now shows no value. This applies the same currency rule the other merger and funding tools already use.

company_fund_votes — proposals grouped by the linked 8-K proposal, with the meeting's own result

company_fund_votes

For consumers: N-PX items that MARS linked to the company's 8-K voting-results proposal are now grouped as that one proposal, however each fund worded it; unlinked items still group by side + wording. Linked rows take their side from the 8-K proposal type (funds sometimes mislabel management items as shareholder ones) and carry a new meeting_result block: the meeting's own support (for / (for + against + abstain)), vote counts, the outcome as reported (null when not yet recorded, never inferred) and the 8-K accession. Tesla 2025-11-06: 13 main proposals; the 2025 CEO award is one row with 1,494 fund votes, 87.9% N-PX support by shares, 149 managers against, and a meeting result of 76.6%, approved.

mars_reit_deal_flow — non-US trade-press deals now left out upstream

mars_reit_deal_flow

For consumers: reit has flagged trade-press (GDELT) deals by non-US or unknown-country companies as unverified (fx_unverified, 8,349 rows), because their amounts were never converted to USD. They are now left out entirely, not just their amounts, so deal counts for non-US REITs and operators can drop. Counts and $ for those issuers cover Form D, news-path and any US-company deals only; treat them as a lower bound. n_amount_withheld should normally be 0 now.

operating_metrics serves cruise lines from MARS's KPI panel again (no Qdrant)

operating_metrics

For consumers: Cruise lines (CCL, RCL, NCLH, Viking, Lindblad) are back in operating_metrics. They now come from MARS's company KPI panel like the other industries: occupancy, passenger cruise days, APCDs, fuel cost per ton, net cruise costs and yields. The old cruise route read the parked Qdrant collection and failed. Annual (FY) rows are flagged upstream and left out; Q4 values MARS derives as FY minus Q1-Q3 are kept and listed in each period's derived_kpis. A company with no KPI rows returns status='not_covered'.

New company_fund_votes tool (how funds voted at a company's shareholder meeting, N-PX)

company_fund_votes

For consumers: company_fund_votes (activism family) takes a ticker or CIK and an optional meeting_date (default: the latest past meeting). It returns the meeting (date, ballot items, fund votes, managers) and one row per proposal. Each row has the side, description, fund votes, support by shares (FOR / (FOR + AGAINST/WITHHOLD)), support by fund votes, and how many managers voted For / Against / Abstain / Split. Pass `proposal` to list each manager's vote with shares and fund votes. Console Fund Votes rules apply: item dates within 7 days form one meeting, a meeting needs 3+ items, and a proposal is side + description (letters and digits, first 80 characters) with 2+ fund votes. Small wording variants (under max(3, 5% of the largest proposal's votes)) are kept, flagged `minor` and listed last; n_proposals counts the rest, and meeting.lowest_support_proposal names the weakest non-minor proposal with at least 30% of the top proposal's votes. Managers are adviser families via the voter's filed or MARS-inferred adviser, or the fund's own name when it has neither. Predecessor and successor CIKs from atlas cik_canonical are included (lineage_ciks), so a reorg's old holding company's meetings show up (e.g. HollyFrontier under HF Sinclair). This covers registered funds only; for the official meeting result use company_proxy_votes.

New short_volume tool (FINRA Reg SHO daily short-sale volume)

short_volume

For consumers: short_volume (ownership family) returns a stock's daily FINRA Reg SHO short volume, short-exempt volume, FINRA total volume and short ratio. It adds volume-weighted 5-day, 20-day and 1-year (252 trading days) short ratios and the z-score of the last 5 days' average ratio against the year. This is off-exchange, FINRA-reported volume only; it is not total volume and not short interest (use short_interest for that). Class shares use a slash (BRK/B), and BRK.B / BRK-B are converted.

operating_metrics covers P&C insurers, homebuilders and airlines

operating_metrics

For consumers: operating_metrics now returns MARS's per-company KPI panel for P&C insurers (combined / loss / expense ratios, premiums), homebuilders (net orders, closings, backlog, average selling price, cancellation rate, active communities) and airlines (ASMs, RPMs, load factor, TRASM / PRASM / CASM, yield). It follows Console's rules: 'warn' rows are left out, each KPI keeps its most common unit, the latest filing wins per period, and a KPI needs 2+ periods. It returns the latest 20 periods by default (`max_periods`, up to 60), newest first.

exec_compensation returns a normalised role (executive / director)

exec_compensation

For consumers: exec_compensation's `role` is now always 'executive' or 'director'. It reads comp's normalised role column; the source data spelled executives both 'exec' and 'executive'. The source spelling is kept in a new `role_raw` field. Director pay rows are back in the data (comp's view now has ~535K rows), and reads come from comp's new materialized view, so a call now takes well under a second.

Credit est_yield_pct is now true yield to maturity (Console parity)

tranche_bond_price_dispersion_screen tranche_widest_held tranche_refi_screen

For consumers: est_yield_pct is now the yield to maturity, matching Console. It uses semi-annual compounding, is solved on the clean price over fractional periods, and counts years as (maturity - report date) / 365.25. It replaces the earlier approximation. Following Console: no yield outside -5% to 80% (Console's bisection bounds), bonds with no maturity date or one on or before the report date are left out of both screens, and distress is only checked below a price of 110 (or for a disputed bond). There is still no yield from a coupon above 20% or a price outside 1-2000% of par (in-the-money convertibles trade up to ~1,475), and none under 0.08 years to maturity (the cutoff was 0.05). The securitisation filter also drops Conseco Finance and Mount Logan names.

New manager_proxy_voting tool (how a manager's funds voted their proxies, N-PX)

manager_proxy_voting

For consumers: manager_proxy_voting takes a 13F manager cik (expanded to its adviser family), an adviser crd or an investor name, plus a proxy_year (2024-2026; July 1 - June 30). It returns fund votes cast, the share against management (of votes with a known direction), a by-category breakdown with shareholder proposals counted separately, proposals a majority of the family's funds opposed, say-on-pay votes the majority opposed, and shareholder proposals the majority backed out of all voted. N-PX notice reports (funds with nothing to vote) are listed as "notice report: no votes" and left out of the counts; a voting report with 0 extracted votes is listed under coverage_gaps. Proposals and counts follow Console's rules exactly: a proposal is issuer + meeting date + vote source + description (letters and digits, first 80 characters); counts are fund vote rows; opposed = more than half of the votes with a known direction against management; shareholder proposals are backed when more than half of all votes on them are FOR, and shareholder_proposal_support_pct gives FOR votes over all votes. by_category carries proposals and proposals opposed per category. It reads mars's precomputed npx_adviser_item_votes_mv (refreshed daily after the N-PX load) and rolls up proposals in one statement. A manager's own N-PX filed under its 13F CIK (e.g. BlackRock, Inc.'s institutional-manager report) counts toward its family. A first call takes about 3-4 seconds for the largest families (BlackRock, Vanguard); repeat calls for the same manager and year within 24 hours return instantly (cached_at says when it was computed).

Adviser screener and credit distressed / disputed screens aligned to Console's definitions

adv_screen_managers tranche_bond_price_dispersion_screen tranche_widest_held

For consumers: adv_screen_managers' universe is now every non-placeholder adviser CRD with current RAUM above 0 (~21.6K, Console's universe; previously SEC-registered only). large_wealth is now wealth focus with AUM of $10B+ (was $5B). The regulatory preset now needs AUM of $100M+, and events_5y counts every regulatory event initiated in the last 5 years, firm- and affiliate-level (previously firm-level only). Advisers whose last AUM is over 2 years old with no registration type stay hidden unless include_stale_unverified=true. tranche_bond_price_dispersion_screen's view=disputed and view=distressed now read MARS's company-linked bond panel, which is Console's source; bonds with no linked company are left out. Securitisations and agencies are matched on both the bond issuer name and the MARS company name, and the list now also covers REMIC, tobacco settlement, FHLB, Farm Credit, Alt-A, "Capital I Trust", "Series NNNN", vintage deal codes such as "2017-1", mortgage shelves (Residential Accredit Loans, Ace Securities), credit-linked notes, manufactured-housing contracts, BSL / loan-funding CLOs and CFOs. distressed has no fund minimum. disputed leaves out structured issuers and prices outside 1-2000% of par, and is_disputed follows the same rule wherever it appears. For both views, a fund mark outside 1-2000% of par (a 0 placeholder or a per-unit price; in-the-money convertibles at 300-1,475 stay in) is left out of the bond's price mean, std dev, min and max before the rules run; n_marks_dropped on each row counts them.

Credit tools follow Console's Credit Monitor rules + new bdc_summary tool

bdc_summary tranche_refi_screen tranche_bond_price_dispersion_screen tranche_widest_held tranche_debt_flow_movers credit_bdc_loan_book credit_cross_held_loans

For consumers: New bdc_summary lists every covered public BDC at its latest quarter with debt fair value, cost, mark %, share of fair value marked below 90% of cost, non-accrual % by fair value and by cost, NAV per share and borrower count. A 100+ borrower book with zero non-accruals comes back as a source gap (null, non_accrual_source_gap=true), not 0%; credit_bdc_loan_book now carries the same fields and flag. tranche_refi_screen defaults to the newest settled month and blanks coupons above 20% (structured notes). tranche_bond_price_dispersion_screen gains view=disputed (price std dev of 5+ points across 3+ funds) and view=distressed (estimated yield of 12%+ or price under 50, securitisations, agencies and mortgage / lease trusts excluded); every row, and every tranche_widest_held row, carries est_yield_pct, is_distressed, is_disputed and is_structured_issuer. tranche_debt_flow_movers reads Console's monthly issuer-flow table, defaults to the newest month with at least 85% of the prior month's issuer coverage and keeps issuers with $50M+ of fund-held bonds. credit_cross_held_loans leaves marks above 130% of cost out of mark_spread. Rows on the refi, dispersion, widest-held and movers tools carry listed_parent (ticker, CIK, name, and how it was found: bond CIK, MARS company CIK, Ex-21 parent or name match; a "<Parent>/<suffix>" issuer name such as "Jetblue Airways/loyalty" falls back to the part before the slash). The tranche and credit tools that take top_n also accept limit. Data fixes for BDC non-accruals and structured-note prices are in progress and flow through without further changes.

adv_screen_managers now matches Console's Adviser Screener (growth, client base, custodian, adviser moves, presets)

adv_screen_managers

For consumers: adv_screen_managers now reads mars and returns, per adviser CRD, aum (the newer of the profile figure and the latest ADV filing), 1y / 3y AUM growth (vs the last filing on or before that date; null when AUM at the start was under $50M), focus (wealth / institutional / private_funds / mixed by the 50%+ RAUM rule), clients, employees, IARs, headline management and performance fee, private funds, main custodian, office city / state, firm-level regulatory events in the last 5 years, the serious-disciplinary flag and net advisers gained over 1y (partial coverage). New filters aum_min/max, g1_min/max, g3_min, focus, state, custodian, fee_max, pf_min, net1_min/max, events_min, name, crd, plus nine presets (growing_wealth, large_wealth, emerging, private_funds, institutional, adding_advisers, losing_advisers, shrinking, regulatory). Existing filters and the current_raum / raum_as_of fields still work. The universe is now SEC-registered advisers with AUM above 0 (previously all 52K ADV filers). Preset thresholds are working defaults until Raul's Console diff. as_of still uses the point-in-time ADV history, and the new filters are rejected there.

New funds_managed tool (every fund an adviser family runs) + advisers on fund_equity_holders rows

funds_managed fund_equity_holders

For consumers: funds_managed(cik | crd) lists every registered fund series (mutual funds, ETFs, closed-end funds) a 13F manager's adviser family runs as adviser or sub-adviser, from SEC N-CEN adviser links (terminated links excluded) joined to the N-PORT fund registry. Per fund it returns series_id, fund name, registrant (the filing trust, not the adviser), advisers and sub-advisers (name, CRD, role), latest N-PORT net assets with its date, and the N-CEN as-of date. It also returns per-adviser fund counts and net assets, and family totals. Advisers are grouped by CRD under one canonical name. role=adviser|subadviser narrows the list. Matches Console's Funds Managed app for State Street (261 funds / $1.23T) and BlackRock (814 / $4.07T). fund_equity_holders rows now carry adviser and sub_advisers per fund series. In the ownership family.

mars deals — new 'reported' integrity tier (single secondary news report) excluded by default

mars_mergers_screen mars_mergers_for_company mars_trending_acquirers mars_funding_history mars_funding_search mars_investor_portfolio mars_sector_funding_flow

For consumers: mars added a tier between vouched and warn on funding_deals_v2 and merger_deals_v2 — verified > vouched > reported > warn. 'reported' means the deal rests on one secondary news report (no filing, Form D, 8-K or agreement exhibit, no wire press release, no major outlet, fewer than 2 outlets); 25,083 mergers and 65,211 funding deals moved there, e.g. Amazon/Decart $7B (one Korean article). Every mars deal tool already filters to verified / vouched, so these deals drop out of default results with no code change; descriptions now name the tier. include_warns=true on the merger tools returns reported and warn deals, with integrity_tier on each row. A reported deal goes back to vouched by itself when a second outlet, a release or a filing attaches.

Console parity — Raul's code check on red flags, proxy votes, insider liquidity, track record, dilution, ECM

company_red_flags company_proxy_votes insider_liquidity insider_track_record insider_proven_buys company_dilution ecm_monitor

For consumers: Raul compared these tools with Kaleidoscope Console's code and the rules now match. company_red_flags — a deficiency notice with no stated deadline stays open 365 days (was 360); "recent" is 455 days (was 456). company_proxy_votes — a meeting date after the 8-K or more than 120 days before it (was 365) falls back to the 8-K date; golden-parachute votes use Console's exact pattern (golden parachute, merger, business combination, tender offer, change in control, Item 402(t), plan of sale/liquidation/dissolution, "paid or become payable", ...); a meeting with a merger or golden-parachute vote and no auditor ratification is "special". insider_track_record / insider_proven_buys — SPY is priced on the first trading day on or after each date (was on or before), the way MARS dates the stock's returns, so excess-vs-SPY figures move slightly. company_dilution — on one filing and instrument, a conversion-price row under 2% of its twin's price (or blanked as a misread) is dropped: Snowflake 2024-09-24 keeps $112.50 and loses the $0.1125 copy. insider_liquidity already ended a 10b5-1 plan on "cancelled" as on "terminated"; its description now says so. ecm_monitor is unchanged (Console adopts its unknown-cap = under $2B rule). pending_console_diff is gone from all of them; company_deals (#1304) keeps it.

exec_compensation reads comp's authority view — no more identical-value LLM pay rows

exec_compensation

For consumers: exec_compensation now reads mars.person_compensation_best_v instead of the raw pay table, and drops rows flagged is_suspect_dup. The old "highest total wins" pick could choose an LLM row that copied one value onto several execs — HF Sinclair (DINO) 2025 showed $5.8M for each of its NEOs; it now reads the proxy table (Atanasov $4.55M, Nitcher $3.24M, Pompa $3.12M, Ledbetter $2.58M). Each row adds source_rank (1 = deterministic proxy-table lane; higher = less trusted), fiscal_year_raw (as extracted — fiscal_year is now normalized) and proxy_year (the year the proxy was filed). include_implausible is gone; the view always excludes implausible values. Reads currently take about 2s.

agreement amounts — consideration_tier ready (undisclosed / redacted, never $0)

company_deals chart_events

For consumers: mars is splitting agreement_exhibits_v2's quality flag into integrity_tier (the document and its parties) and consideration_tier (the amount). Ready ahead of it — company_deals carries consideration_tier on each agreement block (deal values still come from news, never from the agreement), and chart_events' contract lane labels a withheld amount "undisclosed" or "redacted" with value null instead of showing $0 or a blank. Until mars ships the column, "redacted" comes from the existing consideration_redacted flag. When the column lands, company_deals keeps exactly verified / vouched agreements (amount problems no longer mark a document warn), so the ~1,421 newly verified agreements are the same ones it already showed — no double counting.

tranche_issuer_debt_flow + tranche_debt_flow_movers — fund flows in corporate bonds (tranche TR-26)

tranche_issuer_debt_flow tranche_debt_flow_movers

For consumers: Two credit-family tools over tranche's issuer_flow — the change side of the fund-holder data. tranche_issuer_debt_flow(ticker_or_cik, n_quarters) is one issuer's quarter-by-quarter net fund flow in its bonds (are holders adding or trimming?); tranche_debt_flow_movers(direction, as_of, min_funds, top_n) ranks the issuers funds are fleeing (the credit-distress feed) or piling into. net_flow_usd is quarter-over-quarter and coverage-immune (only funds that held the bonds at both reports, each against its own prior report — real trading, not the filing-calendar sawtooth); n_continuing_funds is the basis and is always shown; prefer net_flow_pct to compare; the movers screen keeps a 20-fund floor and defaults to the latest SETTLED quarter (the newest is still filling in; rows say settled true/false). Example: JPM 2026-06-30 net −$54M across 601 continuing funds (−0.27%). Fund-held N-PORT sample, corporate bonds only.

company_deals — a company's M&A history, one row per deal, SEC first (Console M&A History parity, ticket #1304, pending Console diff)

company_deals

For consumers: New transactions-family tool. One row per real deal for a company — role (acquirer / target / seller), counterparty with aka names, kind (Merger / Stock purchase / Asset purchase / Acquisition), status, announced and closed dates, value + value_source, whether a definitive agreement is on file, confirmed, communications count (425 / DEFA14A), advisers (Financial / Legal, which side) and an event trail with accessions (agreement, merger proxy / tender / S-4, 8-K milestones, delisting). SEC merger / stock-purchase / asset-purchase agreements come first (merger vehicles never shown as counterparties; agreement amounts are not deal values — often termination fees); news deals are merged in and deduplicated (misspellings, same counterparty chained, value matches); investments reported as acquisitions are hidden (investments_hidden). Status comes from the filing trail — delisting or 8-K 2.01 Closed, 8-K 1.02 Terminated, proxy / tender / S-4 Pending — else the news stage; old news-only deals still open are flagged stale. News-only acquisitions of US public targets with no SEC filing are unconfirmed and hidden unless include_unconfirmed. Verified: Microsoft → Activision (one deal, $69B, closed 2023-10-13 by 25-NSE, 24 communications), J&J → Abiomed ($16.6B, tender trail, closed 2022-12-22), Pfizer → Metsera (five misspelled records merged, closed 2025-11-13, advisers both sides).

names_affiliations — status is now Active / Former (names-api bd09b49)

names_affiliations

For consumers: names-api now sets each affiliation's status from the person's latest Form 4 or holding filing at that issuer — Active when it's within ~13 months, Former when older (it used to say Active for every issuer). Same field name. Based on filings, not a stated end date, so a quiet but still-serving director can show Former.

donor_person_profile — career "current" now computed at read time (#1146)

donor_person_profile

For consumers: Each career row now carries current (true / false / null), current_status (current / ended / unknown) and evidence_date, read from mars.person_work_history_current_v — TRUE when the evidence or start date is within 13 months, FALSE once the role ended, NULL when there's no recent evidence. The writer's own flag is kept as is_current_as_asserted (it may be stale). Treat unknown as unknown, not current; expect fewer current roles until bios and signals backfill evidence dates.

catalyst_calendar — Console Catalyst Calendar parity (ticket #1303) + past-as_of history reads fixed BREAKING

catalyst_calendar

For consumers: BEHAVIOUR CHANGE — two Console defaults now apply: (1) the confidence floor default is 0.4 (was 0.3); pass min_confidence=0.3 for the old behaviour; (2) debt_maturity rows are OFF when event_types is omitted — name debt_maturity in event_types/types or pass include_debt=true. Added inputs window (7/30/90/180), types (alias of event_types), sector (SIC description or 2-digit SIC), min_market_cap (approximate), q (name/ticker). Two new event types — listing_deadline (open exchange-compliance deadlines, open until the deadline, regained compliance, or for late filers a later 10-K/10-Q) and earnings (PROJECTED earnings dates, $10B+ market cap by default; earnings_min_market_cap to change; any cap when tickers given) — both current-state only. Every row now carries group (Healthcare / Capital markets / Policy / Credit / Earnings), market_cap_approx and sector_sic. The response adds n_total, truncated and by_group counted over the WHOLE window (by_group_returned for the rows sent); with no event types given the default limit is 200 and the cut is shared across groups so projected earnings can't crowd out the rest. cik is an integer on every row. FIX: past-as_of reads on the history table (pdufa / adcom / ipo_lockup / device / tariff) had been failing since catalyst changed its detail column from jsonb to text — they work again, and query failures no longer return raw SQL.

ecm_monitor — market-wide equity capital markets and dilution (Console ECM Monitor parity, ticket #1302, pending Console diff)

ecm_monitor

For consumers: New fundamentals-family tool for listed US companies over 30 / 90 / 180 / 365 days. KPIs (primary offerings, $ raised ex-ATM, ATM programmes, resales, reverse splits, authorized-share increases, variable-price converts, lock-ups ending in 120 days), a list for each, an underwriter league table (deals, leads, $ credited, ATM mandates) and a by-sector rollup. Filters: sector (SIC description or 2-digit SIC), under_2b (approximate cap), q (name / ticker). Same rules as company_dilution (misread conversion prices not counted as variable, amount cap, generic underwriter names removed); $10 blank-check IPOs and warrant / unit tickers left out of lock-ups. League $ credit comes from primary offerings only — an ATM amount is programme capacity, not money raised. Jefferies LLC carries a source_issue note: the source table names it on ~4,600 offerings, likely an extraction default under investigation.

company_dilution — offerings, converts, reverse splits, PIPEs and IPO per company (Console ECM parity, ticket #1301, pending Console diff)

company_dilution

For consumers: New fundamentals-family tool. Registered equity offerings (amended filings folded into the original) with debt offerings kept separate, private placements and converts (conversion price, discount, variable/reset flag, investors, maturity), dilution-relevant charter changes (reverse-split ratio, authorized shares before/after), Form D private raises, the IPO with lock-up tiers, cards (raised in 12 months ex-ATM, ATM programmes, reverse splits, authorized increases, variable converts) and coverage. Console rules — a conversion price under $0.10 and under 2% of the last price is a misread conversion rate / par value, dropped with its variable flag (GameStop, Snowflake); implausibly large amounts dropped (the >$100B rows are non-USD notes read as dollars); generic underwriter names removed. Registered-offering coverage is ~900 filers a year, mostly small caps — none on record is not none filed. Amount cap, generic-underwriter list and fold key are named constants pending Raul's Console diff.

company_red_flags — accounting, listing, risk and related-party trouble in one call (Console RF parity, ticket #1297, pending Console diff)

company_red_flags

For consumers: New filings-family tool. Cards for going concern, material weakness, restatement, impairment, listing status, new risk factors, related party and bankruptcy 8-K (state clear / info / warn / hot, latest date, accession, count), plus the financial events, exchange-listing events with the open notice, the latest 10-K/10-Q risk-factor comparison (new paragraphs high → medium → low) with history, 3 years of related-party transactions, and newest date per source. Console rules — MARS 'warn' rows dropped; going concern counts only when the text says it, it's about the company (not a subsidiary / segment / country) and market cap < $2B (cap approximate); a listing notice stays open until its deadline, regained compliance, or for a late-filing notice a later 10-K / 10-Q; implausible event dates dropped (date_dropped). Thresholds the spec leaves open are named constants pending Raul's Console diff.

insider_track_record + insider_proven_buys — graded insider buy records vs the S&P 500 (Console parity, ticket #1300, pending Console diff)

insider_track_record insider_proven_buys

For consumers: Two new ownership-family tools (insider_track_record also in people). insider_track_record — for a person (person_id / owner_cik) their buy record and every buy episode, or for a company everyone who bought there scored on all their buys. insider_proven_buys — market-wide recent buys (30/90/180 days) by insiders graded proven / strong / weak / all, with role, value, approximate market cap and SIC-sector filters, companies with 2+ such buyers, KPIs and leaders. Scoring is Console's, not MARS's insider_track_record_v2 — buys within 30 days count as one episode priced at the first; excess = stock − SPY over the same calendar window; returns above +200/300/400% and employee plan-buying days (8+ insiders on one day at issuers with 3+ such days) are left out; grade needs 3 scored 90-day buys (strong = 4+, ≥65% beat, median ≥ +5 pts; good = ≥55%, median > 0; weak = median ≤ −5 pts and <45%). Market cap here is the earnings panel's approximation, not Console's guarded cap. Marked pending_console_diff until checked against Console's code.

insider_liquidity — Form 144 + vesting/option calendar + 10b5-1 plans per company (Console LIQ parity, ticket #1299)

insider_liquidity

For consumers: New ownership-family tool: when a company's insiders can sell, and have said they will. One call per company returns Form 144 notices (12 months, with broker and value at last close), the vesting / option-expiry calendar for the next 365 days with a monthly rollup, Rule 10b5-1 plan adoptions and terminations since 2023 with which plans are active, and cards (Form 144 $ last 90 days, vesting next 90 days, options expiring 12 months, active plans). Console rules: an implausible Form 144 value is withheld and shares × last close shown as estimated; a stated value >3× off shares × last close is kept but flagged mismatch and totals use the smaller; one active 10b5-1 plan per person (a newer adoption or termination closes the earlier; no end date = 15 months after adoption); event dates >11 years past their filing dropped as typos. insider_liquidity_calendar stays for person-axis and market-wide scans.

company_proxy_votes — shareholder-meeting results + next ballot (Console PV parity, ticket #1298)

company_proxy_votes

For consumers: New activism-family tool. One call returns a company's 8-K Item 5.07 vote results (per proposal for/against/abstain/broker, support %, outcome, against-management flag), the next ballot (DEF 14A filed after the latest result — null when no meeting is pending), say-on-pay history, 5 years of shareholder proposals, and cards for last meeting / say-on-pay / director support / auditor ratification. Same math as Kaleidoscope Console — support = for ÷ (for+against+abstain), broker non-votes excluded; say-on-pay < 70% and director support < 80% flagged warn; directors never failed by the vote count (plurality voting); for say-on-pay and auditor the count overrides a contradictory extracted outcome (see outcome_note); missing or impossible meeting dates fall back to the 8-K filing date (meeting_date_estimated). Merger-related compensation votes (typed say_on_pay by the extractor) are recognised as golden-parachute votes and kept out of the say-on-pay card/history, so a deal-only ballot reads as a special meeting. Where the source copied broker non-votes into abstain (~7% of rows), abstain is left out of support % with a note. Follows merger-legacy CIKs — XOM's 2026-07 holdco reorg (2115436 ← 34088) no longer hides its history. Filings returned by accession.

dedicated graph-only bearer minted for mars-ui (#757 Find Path / Graph view)

graph_path graph_neighbors graph_entity graph_resolve graph_schema graph_aggregate graph_ledger

For consumers: mars-ui's backend (ticket #757, built with nexus) now calls graph_path / graph_neighbors / graph_entity on the MCP instead of reimplementing path-finding. It gets its own credential — admin.bearers row mars-ui@kscope.io, verticals={graph} — separate from nexus-web's, so each can be revoked independently. Same isolation pattern as the earlier inferno/crown graph-only bearer. Verified live on mcp.kscope.io before handoff: exactly the 7 graph tools + discover_tools + resolve, nothing else; unauthenticated → 401.

mars_reit_deal_flow — non-USD trade-press amounts withheld instead of summed as USD (ticket #825, 4th table)

mars_reit_deal_flow

For consumers: names confirmed live that mars.real_estate_deals_v2 has the #825 currency bug in its biggest source. The trade-press (GDELT) extraction path (17,141 of 19,491 rows) stored the article's number as amount_usd with no currency handling — e.g. a Philippine REIT's PHP 77B deal (~USD 1.4B) sat in the table as $77B; PH/JP/CN medians run 3-4× the US median. The table has no currency column to gate on, so until the reit team routes GDELT through FX conversion, mars_reit_deal_flow now sums $ only from Form D filings (genuine USD), the news path (converts properly), and GDELT deals by US-domiciled companies. Every other deal still COUNTS toward n_deals / n_acquisitions / etc. — only its amount is withheld. Each bucket now carries `n_amount_withheld` (how many deals' $ were excluded) and `n_amounts_gdelt_approx` (how many summed amounts came from the approximate GDELT path). US REITs are effectively unchanged (verified O, SPG, PLD); for non-US issuers (verified PHI, MFG) $ sums are now a lower bound — lead with counts and net direction, which were always the reliable signal here. Upstream fix (FX conversion + a currency column) is with the reit team.

mars_* tools — fixed a live currency-mislabeling bug on amount_usd/deal_value_usd fallbacks (ticket #825)

mars_funding_search mars_investor_portfolio mars_mergers_for_company mars_mergers_screen mars_trending_acquirers mars_sector_funding_flow mars_find_similar_deals chart_events

For consumers: names + mars flagged (ticket #825) that `amount_usd`/`deal_value_usd` on funding_deals_v2/merger_deals_v2 hold the RAW figure in whatever currency the source printed, not USD -- the correct value lives in the `*_converted` columns. Six tools already preferred `*_converted` via `COALESCE(converted, raw)`, but that fallback wasn't fully safe: when `converted` is NULL AND the row's own currency is non-USD, it silently used the raw foreign-currency number AS IF it were USD. Quantified before fixing -- 2,075 funding_deals_v2 rows + 330 merger_deals_v2 rows hit exactly that gap (e.g. a single row mis-tagged currency='UF', a Chilean inflation unit, was inflating one day's funding-flow sum by $3.4B when counted at face value). Fixed by guarding every fallback on the row's own currency: `COALESCE(converted, CASE WHEN currency IS NULL OR currency='USD' THEN raw END)`, across ~16 filter/sort/aggregate call sites. names then confirmed two more tables share the bug: `mars_find_similar_deals` (reads the older article-grain funding_rounds_v2/mergers_resolved_v2, which have NO `*_converted` column at all -- non-USD rows now null out instead of mislabeling; this tool is currently OFFLINE per the Qdrant decommission so the fix is hygiene for its eventual revival, not a live-traffic fix) and `chart_events`'s contract lane (agreement_exhibits_v2 has real currency diversity + a real `amount_usd_converted` column, but the code read `deal_value_usd or amount_usd_converted` -- backwards priority -- now flipped converted-first with the same currency guard on the raw fallback). Verified live on dev and prod post-deploy, including against a real non-USD contract filer. CORRECTION (2026-10-11): this entry originally said `mars_reit_deal_flow` / real_estate_deals_v2 was low-risk and parked. That was wrong — names checked live and it IS contaminated (17,141 of 19,491 rows come from the GDELT path, which never converted currency). Fixed separately; see the 2026-10-11 mars_reit_deal_flow entry above.

graph_ledger — live-compute fallback on entity_ledger_v2 cache miss, instead of flat found:false

graph_ledger

For consumers: nexus (a consuming agent) and mars independently found that entity_ledger_v2 has only ~51 rows from the original 2026-09-30 pilot with no refresh since, so graph_ledger returned found:false for essentially every real company/person outside that set -- a tool that reads as general-purpose but was ~99% miss in practice. mars fixed a catalog bug blocking `mars.compute_entity_ledger(node_id, node_kind)` and confirmed it works live; this wires that in as a fallback on any cache miss instead of giving up. Every response now carries `cache_status` (`hit` | `miss_live_computed`) so it's never ambiguous which path served a call -- a live-computed one costs ~0.5-0.7s instead of an instant cache hit, no write-back to entity_ledger_v2 from our side (we're read-only on proc2; a standing nightly materialization job is a separate scope/cost decision mars is bringing to kee). Verified live through the dev MCP endpoint: Eloxx Pharmaceuticals and a person tied to it (David Rector, via officer_of) both now resolve with full sections; an existing pilot node still hits cache; a genuine non-existent identifier still returns a clean found:false via normal resolve failure, not a live-compute attempt.

7 new tools shipped from Console's P1 gap list — short_interest, holders_combined, company_filings, price_history, chart_events, adv_private_funds + a bond_yields extension on tranche_bond_card

short_interest holders_combined company_filings price_history chart_events adv_private_funds tranche_bond_card

For consumers: Follow-on to yesterday's Console-family rework -- Raul's kbrain spec named ~23 P1 tools (a live Console report with no MCP backing); these 7 were the ones confirmed fast/unblocked after a status-check round with the owning data teams (adv, sections-fc, xbrl all replied same-day; material_contracts and geographic_revenue stay deferred pending those teams' own work). `short_interest` surfaces mars's FINRA short-interest panel with its own precomputed squeeze/spike flags verbatim. `holders_combined` merges thirteenf_holders_full + search_activist_positions + insider_holdings into one register (13F/13D/13G/Form 4), each row carrying its own as_of so a 13D/G filing date is never read as a 13F quarter-end. `company_filings` tags a company's filings with fired classifiers + evidence sentences -- required working around a live query-planner pathology against JPM (a mega structured-note issuer) where combining cik+role+form filters with ORDER BY+LIMIT made Postgres scan every company's filings globally instead of just JPM's; fixed by sorting/windowing in Python instead of SQL. `price_history` is the first MCP tool to return a price series at all (atlas market.prices_v), implementing Console's split-adjustment convention and verified against NVDA's 2024 10-for-1 split and AAPL's full 4-split history. `chart_events` composes 9 separate sources into Console's 10 event lanes (earnings/dividend/split/insider/stake/campaign/formd/debt/contract/auditor), each lane isolated so one source failing doesn't break the rest. `adv_private_funds` wraps adv's newly-confirmed-live Schedule D private-funds tables, with the spec's `include_prior` parameter dropped entirely once adv clarified the source has no time-series dimension to build it against. `tranche_bond_card` gained a yield_analysis block (YTM, 1M/3M price deltas, a distressed flag) rather than shipping bond_yields as a separate near-identical tool, per that spec's own suggestion. All 7 verified against real live data including deliberately-chosen worst cases (JPM, Ford, GME, Politan Capital) before shipping; family notes in kbrain (Ownership / Filings / Charts / Advisers) updated to move these from "to build" to "live," each flagged with its honest gaps vs. the original idealized spec.

MCP tool families re-cut to match Console's modules — 17 new keys, shared resolve tool, every live grant migrated BREAKING

resolve resolve_company_name mars_company_resolve graph_resolve bio_entity_resolve names_lookup donor_person_resolve

For consumers: Raul decided (kbrain Reference/MCP families.md, 2026-10-02) that the MCP's tool families should match Console's own module keys one-for-one, so a family name means the same thing in both products -- replacing the old data-source-shaped families (sec, names, tranche, bio, funding, mergers, power, sedar, graph, catalyst, news, donor, insider, agreements) with 17 Console-aligned keys: overview, filings, ownership, activism, credit, private, advisers, graph, fundamentals, intelligence (plus three independent add-on packs intelligence.bio / intelligence.power / intelligence.mining), charts, people, transactions, prospect. 127 live tools redistributed across them exactly per Raul's tool-by-tool mapping (verified to match his count precisely) -- most tools moved to a different family than before; e.g. the old `sec` bundle (40 tools) now spans 8 different families. New shared `resolve` tool added (company/person/investor/institution in, canonical id out; pass kind='gene' for biotech gene/target resolution) -- callable regardless of which families you hold, same as discover_tools. The 6 older single-purpose resolvers (resolve_company_name, mars_company_resolve, graph_resolve, bio_entity_resolve, names_lookup, donor_person_resolve) are UNCHANGED in behavior and response shape, just marked superseded in their descriptions -- no prompts break. The tranche__*/sections__* double-underscore tool names are gone entirely (single-underscore was already the preferred spelling via alias; now it's the only one). Every live customer's and test account's entitlement grant was migrated to the new keys in the same deploy window as the code (old-key grants would otherwise match nothing) -- if a tool you used yesterday is suddenly ungranted today, your family assignment likely needs a follow-up; reach out and we'll fix the mapping. Three families are placeholders for now: `charts` has zero live tools (Console's own Charts module went live 2026-10-02 with no MCP backing yet -- price_history/chart_events are the first P1 builds); the 6 non-mining sedar_* tools, search_news/search_news_articles, and mars_find_similar_deals stay offline exactly as before (unrelated to this change -- parked pending other decisions) and aren't in any new family yet.

id: 2026-10-02-graph-neighbors-reverse-edge-fix date: 2026-10-02 title: graph_neighbors — fixed 10 catalog edges (20 label directions) that were never walkable in reverse, including successor_of kind: fix tools: [graph_neighbors, graph_entity] breaking: false env: prod (main, graph vertical) + dev for_consumers: nexus reported (team-mail) that mars.graph_edge_catalog declares a reverse_label "has_successor" for successor_of, and graph_schema advertises it as live, but graph_neighbors never actually walked it -- querying from the dormant/predecessor CIK worked, querying from the current/active company (the direction that matters 100% of the time in practice) returned zero edges. Confirmed 3-for-3 against Ocean Biomedical, IonQ, and Rocket Lab. Root cause: the internal reverse-direction guard only auto-reversed an edge when the two node types differed or the edge was marked "undirected" -- a same-type DIRECTED edge (company->company) could never qualify no matter what the catalog declared, regardless of whether a reverse_label was set. Auditing the full catalog found this wasn't specific to successor_of -- it silently affected 10 edges / 20 label directions total: subsidiary_of, owns_subsidiary, divested_to, signed_agreement, managed_fund_of, uses_prime_broker, markets_fund_of, adr_of, and branded_vehicle_of, alongside successor_of. Fixed by treating a declared reverse_label as an unconditional auto-reverse trigger (the catalog's own contract), with one collision guard: skipped when another catalog row already provides that exact label as a forward edge for the same node_type -- caught and confirmed this actually happens for owns_subsidiary (its reverse_label is literally the string "subsidiary_of", colliding with the standalone subsidiary_of row across two different substrate tables; without the guard a child company showed the same parent twice, verified present before the fix and suppressed after). Verified the guard doesn't over-fire on a same-string-different-meaning case either (managed_fund_of's reverse "manages_fund" shares a name with the real investor->fund manages_fund edge but never collides in practice, since the two apply to different node types). graph_entity's edge-count histogram shares the same underlying function and picks up the fix too. Verified end-to-end against the live dev endpoint before and after deploying to prod.

scoped names-only bearer minted on mcp.kscope.io for nexus's tool-selection testing

names_990_boards names_affiliations names_holdings names_lookup names_profile names_search names_transactions names_vesting_calendar

For consumers: kee flagged a report that Robin (a real Insider Focus/names customer) hit a case where the model reached for web search instead of an MCP tool she was actually entitled to -- the same class of problem nexus had just fixed for graph_ledger (see the discoverability-fix entry above). Audited Robin's account first: verticals={names}, tool_access=["@names"], 44 calls total since 2026-09-29, zero failures, nothing ever touched a SEC-vertical tool -- so whatever went wrong happened upstream of us, in the model's own web-vs-MCP-tool choice, not at our layer. To let nexus reproduce it with the same real-tool-call-trace methodology that solved graph_ledger, minted a new admin.bearers row (email nexus@kscope.io, verticals={names} only, same isolation pattern as the earlier graph-only bearer for inferno/crown -- narrowly scoped, no off-ramp to other families, soft-revocable). Verified live before handing it over: sees exactly 12 tools (the 8 names_* tools + the 4 always-on core resolvers discover_tools/resolve_company_name/resolve_company_profile/semantic_search), nothing else. Sent nexus Robin's exact query sequence/args from both her first session and her most recent one as realistic repro material. Awaiting nexus's findings.

id: 2026-09-30-graph-ledger-discoverability-fix date: 2026-09-30 title: graph_ledger fixed — LLM tool-selection was never picking it, even when told to; now resolves a bare company name in one call kind: fix tools: [graph_ledger] breaking: false env: prod (main, graph vertical) + dev for_consumers: nexus ran a real grok-4.7 session live against argos-dev (exact tool-call trace via grok export, not inferred) and asked it to look up AllSpice.io as a company. It never called graph_ledger -- its own tool-selection reached for mars_company_resolve + mars_funding_history + mars_search_articles across 3 separate calls plus an external WebFetch instead, reasoning "I'll look for a company-profile or resolve tool rather than a public-markets one." Root cause was two-part, both fixed: (1) graph_ledger required a pre-resolved node_id, so even considering it meant a separate graph_resolve call first -- now it accepts an optional identifier (name/ticker/domain/CIK/crd/linkedin_url) and resolves internally via the same cascade graph_resolve uses, so picking the tool at all is now sufficient; (2) the description read as a narrow graph-jargon primitive ("PILOT -- a precomputed neighborhood digest for one graph node") rather than the general-purpose company/person/investor overview it actually is -- rewrote it and measured the change with the server's own E5 discover_tools embedding against the exact 3 competing tools nexus's trace showed it lost to, across 4 natural phrasings of the same question: moved graph_ledger from dead-last on every phrasing to winning 2 of 4 outright and closing most of the gap on the other 2. Also closed a silent gap found while in there -- graph_ledger had no TITLES entry (every sibling graph_* tool does), which was weakening its own discover_tools embedding independent of the description content. This is a measured improvement on the search mechanism we control (our own discover_tools), not a guaranteed fix for every client's own opaque tool-selection algorithm (grok's internal one included) -- but AllSpice.io itself is a good honesty check: it resolves cleanly as a real company now, but isn't one of the 50 companies/people in the ledger's current precomputed pilot set, so it correctly degrades to found:false with resolved_via surfaced (proving resolution worked) rather than requiring a separate graph_resolve call just to learn the same thing. Verified end-to-end against the live dev endpoint: an in-pilot company (Prolific Machines) resolved by name returns the full ledger in one call; AllSpice.io resolves and honestly reports no pilot coverage yet; a non-existent name and a call with no input at all both fail cleanly; the original node_id-only path is unchanged.

fixed a real sign-in bug — corporate email security scanners were burning magic links before the user clicked them

For consumers: A customer (berenice.munoz@arcosa.com) reported a sign-in error; prod logs plus IP-org lookups traced the root cause to Microsoft Defender for Office 365 Safe Links -- it GET-prefetches every link in an inbound email before the recipient opens it, and our /signin/verify was a single GET that both validated AND consumed the one-time magic-link token in one step, so the scanner's prefetch burned the token before she ever clicked it. She then got "This sign-in link is invalid, expired, or already used" on her own real click, seconds after a Microsoft-owned IP had already (successfully, from the server's point of view) consumed it. Not an entitlement/allowlist issue -- her account was correctly provisioned throughout. Fixed by splitting the flow: GET /signin/verify now only peeks (db.peek_magic_link, no side effect) and renders a "click to confirm" page; the actual single-use consumption (db.consume_magic_link, unchanged) now happens on the POST that page's button triggers. Email/link scanners GET-prefetch but essentially never submit HTML forms, so this closes the race without weakening the one-time-use guarantee -- verified with a 5-step smoke test (repeated GETs never consume; the first POST consumes and redirects with a session cookie; a second POST on the same token correctly 400s; a GET after real consumption correctly shows the expired/used error) before and after deploying to prod. This will fix the same failure for any future customer behind a similar corporate link-scanning gateway (Microsoft Safe Links, Proofpoint, Mimecast, etc.), not just this one case.

names_search — issuer= filter found to be a complete no-op upstream, stripped + warned

names_search

For consumers: Found while investigating a 13s names_search call from a grok CLI test run (see the graph/mars-mcp retirement entries for that testing thread). Root cause turned out to be a correctness bug on names-api's backend, not latency: the issuer= filter on /search has zero effect -- issuer=<a nonsense string that should match nothing> returns the exact same total=305293 as no filter at all, meaning it silently returns the entire unfiltered dataset regardless of what's passed. Sanity-checked name= and ticker= both work correctly (real value -> real selective count, nonsense -> 0), so this looks isolated to issuer specifically. This is a correctness issue, not just performance -- any caller using issuer= got back a real-looking total with no indication the filter did nothing. Fixed on our side: names_search no longer sends issuer= upstream and returns an explicit warning when a caller passes it, rather than silently passing through the misleading total; other filters in the same call (ticker, title, etc.) are unaffected. Reported to the names team directly, who own the backend fix.

graph-mcp.staticpipe.com and mars-mcp.staticpipe.com retired — mcp.kscope.io covers both

For consumers: Retired the last two standalone MCP services after the customer-vertical teardown earlier today. graph-mcp.staticpipe.com: its 12-tool graph family is available on mcp.kscope.io via a scoped verticals=['graph'] grant, but inferno (the actual consumer) turned out to need full access rather than isolation, so this was superseded by giving inferno's existing bearer verticals=['full'] instead (see the earlier entry) -- no distinct need for the isolated surface remained. mars-mcp.staticpipe.com: its 20-tool catalog is exactly VERTICAL_TOOL_SETS['funding'] | VERTICAL_TOOL_SETS['mergers'] (verified: zero set difference either way -- 7 funding-only, 7 mergers-only, 6 in both), fully covered by mcp.kscope.io's funding+mergers verticals, and inferno's now-full access already includes both. Both services stopped + disabled (argos-mcp-graph.service, argos-mcp-mars.service), their single whole-host ALB rules removed from the argos-mcp-dev listener (priorities 6 and 7), and their target groups (argos-mcp-graph-tg, argos-mcp-mars-tg) deleted. Neither hostname had a dedicated catch-all the way the kscope.io customer verticals did, so both now fall through to the ALB's default action (mcp.staticpipe.com's own backend) rather than a clean 4xx -- functionally safe (still requires the dev bearer token, no new exposure) but not the cleanest failure mode; noted as a known characteristic, not fixed further. mcp.staticpipe.com and the unrelated mars.staticpipe.com (Metabase, sharing the same ALB) both confirmed unaffected. Active ARGOS MCP surface is now down to 2 services: mcp.kscope.io (primary) and mcp.staticpipe.com (dev) -- the whole per-vertical-subdomain model from earlier this year is gone.

fixed a real access-control gap in the unified catalog — grok-compat alias tools bypassed per-user gating

sections__ipo_screen and 4 siblings tranche__bond_card and 6 siblings -- both canonical and single-underscore alias spellings

For consumers: Found while setting up a scoped graph-only bearer (see the entry below) to consolidate graph-mcp.staticpipe.com into mcp.kscope.io. The 13 single-underscore grok-CLI-compat aliases for sections__/tranche__ tools (sections_ipo_screen, tranche_bond_card, etc.) were completely ungated on the unified catalog -- can_see() returned True unconditionally for the alias name regardless of caller entitlement, while the canonical double-underscore name was correctly gated. This affected every narrowly-scoped grant already live, not just the new bearer -- e.g. peter@rockpointadvisory.com (sedar-only) could see tranche_*/sections_ipo_* tools he has no grant to, via their alias names. Root cause: UNIFIED_BETA_TOOLS's gate-mirroring pass for grok aliases ran on BETA_TOOLS before UNIFIED_BETA_TOOLS added its own batch of fresh gates from VERTICAL_TOOL_SETS -- those fresh entries never got a second mirror pass. Fixed with a second alias-mirror pass over the fully-built UNIFIED_BETA_TOOLS. A companion bug surfaced by the same fix (over-restrictive this time): VERTICAL_TOOL_SETS only stores canonical names, so a legitimately entitled caller could see the canonical name but not the alias for the same tool -- fixed with a new ALIAS_TO_CANONICAL reverse map consulted by _vertical_grants() before membership checks, and a second, independently-drifted copy of the same "@vertical -> VERTICAL_TOOL_SETS" logic duplicated inline in can_see() was routed through the shared helper instead of patched separately. Verified across 4 real identities post-deploy (an @sec external, a graph-only bearer, an unentitled account, and full-admin) that both canonical and alias names now return identical, correctly-scoped results in every direction.

scoped graph-only bearer minted on mcp.kscope.io — path to retiring graph-mcp.staticpipe.com too

graph_resolve graph_entity graph_neighbors graph_schema graph_aggregate graph_path graph_ledger mars_funding_history mars_investor_portfolio mars_mergers_for_company donor_person_profile donor_person_track_record

For consumers: Minted a new admin.bearers row (email inferno@kscope.io, a SEPARATE credential from their existing broad bearer -- deliberately not added to it, since combining would give an off-ramp to relational tools and defeat the isolation property below) scoped to verticals=['graph'] only, for inferno/crown to use on mcp.kscope.io in place of graph-mcp.staticpipe.com's shared static bearer. Verified live: sees exactly the 12 graph-family tools plus the 4 always-on core resolvers (discover_tools, resolve_company_name, resolve_company_profile, semantic_search) -- no relational mars/sec/tranche tools, same "no off-ramp" isolation graph-mcp.staticpipe.com's dedicated ARGOS_VERTICAL=graph server provides today. Once inferno/crown switch to this credential, graph-mcp.staticpipe.com can be retired the same way the 10 customer verticals were earlier today. Not yet decommissioned -- awaiting the actual handoff/switch.

full teardown of the 10 vertical services complete — disabled, ALB rules + target groups removed

For consumers: Completes the retirement started earlier today (stopping the 10 services). Disabled all 10 systemd units (no longer restart on a proc3 reboot). On the argos-mcp ALB: removed the 30 vertical-specific listener rules (the 3 /mcp*+OAuth-path rules per vertical pointing at each dead target group) while deliberately KEEPING each host's existing catch-all rule to MCP-LANDING -- so mcp-sedar.kscope.io etc. still render Raul's landing page instead of a raw ALB 503. Deleted the 10 now-orphaned target groups (argos-mcp-{sedar,funding,mergers,sec,news,catalyst,tranche,agreements,bio,insiderfocus}-tg). DNS itself is untouched -- those CNAMEs are Raul's Namecheap registrar, outside our control, and harmless to leave pointed at the ALB now that routing falls through cleanly. Verified externally against the real public hostnames post-teardown: mcp.kscope.io/health=200, mcp.kscope.io/mcp/ (unauthed)=401 (routing intact), mcp-sedar.kscope.io/mcp/=404 (graceful fallthrough, not 503), mcp-sedar.kscope.io/ and mcp-ifocus.kscope.io/ =200 (landing pages render normally). graph-mcp.staticpipe.com and mars-mcp.staticpipe.com (internal research surfaces, different ALB) untouched throughout.

10 customer-facing vertical MCP services stopped — consolidated onto mcp.kscope.io

For consumers: Following the unified-catalog cutover (same day, see the entry below), stopped the 10 dedicated per-product vertical systemd services -- argos-mcp-{agreements,bio,catalyst,funding,insiderfocus,mergers,news,sec,sedar,tranche} -- freeing ports 8002/8004-8010/8013. Their customer-facing hostnames (mcp-sedar.kscope.io etc.) will now 503 at the ALB (target groups have no healthy backends) rather than serve traffic; DNS/ALB rules themselves are untouched for now. Justification: audit-log analysis showed zero real external customer traffic on any of these 10 hostnames across their entire history -- every real call was from internal team accounts. The two customers who'd been specifically handed a vertical URL (Peter/Rockpoint on mcp-sedar, Robin on mcp-ifocus) were confirmed migrated to mcp.kscope.io before this step. Services were stopped, not disabled -- still enabled for boot, so a proc3 reboot would restart them; a full teardown (disable + remove ALB rules + DNS) is a separate, later cleanup step. mcp.kscope.io confirmed healthy post-stop; zero failed systemd units. graph-mcp.staticpipe.com and mars-mcp.staticpipe.com (internal research/test surfaces, not customer product URLs) were left running, untouched.

mcp.kscope.io now serves the unified, per-user-gated catalog (the actual cutover)

every tool

For consumers: ARGOS_UNIFIED_CATALOG=true is now live on the main mcp.kscope.io service (was built + deployed dormant on 2026-09-27, flipped live today). mcp.kscope.io now registers the full ~140-tool catalog and gates per-user via admin.users.tool_access / admin.users.verticals, replacing the old ARGOS_DEFAULT_VERTICAL=sec grab-bag. Every current user's grant was pre-migrated before this flip so nothing changed for existing users: verified live post-cutover against the real deployed code + real DB for 4 representative accounts (an @sec-only external contact, peter@rockpointadvisory.com on @sedar, robin@kscope.io on @names, and an admin) -- each sees exactly their entitled family's tools and nothing else, core resolvers (resolve_company_name, semantic_search) visible to everyone. This is the first real step toward retiring the 9 dedicated per-product vertical subdomains (mcp-sedar/funding/bio/mergers/sec/news/catalyst/tranche/agreements.kscope.io) -- audit-log analysis shows zero real external customer traffic on any of them to date, so the actual decommission is lower-risk than a typical migration, but the two customers who were handed a specific vertical URL (Peter -> mcp-sedar, Robin -> mcp-ifocus) need to be pointed at mcp.kscope.io before their old URLs go away, not after.

mars_search_articles — cut over from frozen mars.article_meta_v2 to sections' OpenSearch news_* index BREAKING

mars_search_articles

For consumers: Full backend replacement, same tool name/identity. mars.article_meta_v2 (the Postgres table this tool used to read) has been frozen since 2026-09-24 with no path back to a live feed -- see the earlier same-day "frozen notice" entry for how that was discovered and verified. Per house standard ("no tool ships that doesn't work -- fix it or remove it"), this tool is now backed by sections' actively-maintained OpenSearch news_* cluster instead: a genuinely broader, live corpus (hundreds of domains, global, multi-language, three source buckets rss/benzinga/gdelt -- benzinga synced directly off the same source DynamoDB stream mars used to read, confirmed the exact set) rather than mars's narrower Benzinga-plus-one-GDELT-source table. BREAKING for existing callers: `company_id`/`sector`/`min_deal_link`/`linked_deal_ids` are gone (those joined to mars's own deal tables, which this corpus has no relationship to); new `source_bucket`/`source_domain`/`tickers`/`language`/`include_incomplete` filters replace them. Query syntax moved from Postgres websearch_to_tsquery to OpenSearch query_string (same bare-terms-AND/quoted-phrase/OR/NOT shape, not byte-identical -- malformed syntax now falls back to a lenient search instead of erroring). Two honest caveats baked into every response's `warnings[]`: (1) freshness is NOT real-time -- publish-to-queryable is a MEDIAN of ~2.5hr for the rss-sourced bucket (RSS polling lag), benzinga is much fresher (CDC stream); (2) known-incomplete (teaser-only) docs are excluded by default, opt in with `include_incomplete=true`. One data-quality bug caught and fixed before shipping: `body_complete` is only ever populated on rss-bucket docs -- an early version of the default filter used `body_complete=true`, which silently excluded the ENTIRE benzinga+gdelt corpus (null field never equals true) -- fixed to exclude only docs affirmatively known incomplete (`must_not body_complete=false`), leaving unset fields untouched.

mars_search_articles — honest staleness notice (underlying corpus frozen since 2026-09-24)

mars_search_articles

For consumers: A cross-team research thread (sections, rss, mars-spider-optimization) uncovered that mars.article_meta_v2 — the table this tool reads — has been completely frozen since 2026-09-24 21:31 UTC, the exact moment mars retired both of its writers (mars-feed-gdelt.timer, mars-feed-sqs.service) as part of exiting the ETL business (Benzinga ingestion is moving to sections' OpenSearch infrastructure). The table was never rewired to a live source afterward, so it will not grow further pending a mars decision (rebuild the feed vs. formally supersede it). We verified this isn't a data-loss event: 479/500 sampled doc IDs from sections' OpenSearch news_2026 index match mars's docids exactly (same upstream DynamoDB table + CDC stream), and sections' historical coverage goes back to 2003-07-15 vs. this table's own 2009-08-10 start — a superset, not a smaller replacement. Added an explicit frozen-as-of-2026-09-24 warning to both the tool description and every response payload (`warnings[]`) so callers don't mistake results for live news. No data or behavior change otherwise. mars is deciding the long-term architecture and will follow up; sections' news_* index (90x this corpus's yearly volume, hundreds of domains, global/non-English) is a real, currently-unwrapped capability gap worth a future tool if that data can be relied upon going forward.

can_see() now gates DB-backed customer bearer tokens too (unified-catalog Phase 1 follow-up)

every gated tool via auth/identity.py:can_see

For consumers: Internal-only fix inside the still-unreleased unified-catalog mechanism (ARGOS_UNIFIED_CATALOG=true, see the 2026-09-27 "Unified catalog mode" entry below) -- caught before any deploy. can_see() previously bypassed ALL per-tool gating for kind=="static" identities, which was safe under the old per-vertical-hostname model (the catalog was already narrowed to one product's tools before can_see ever ran) but would have let every existing per-customer bearer token (e.g. a sedar-only bearer) see the ENTIRE catalog once registered against the new single-URL unified interface. Fixed by distinguishing the legacy break-glass/dev bearer (unchanged: still sees everything) from a DB-backed per-customer bearer, which is now gated by its admin.bearers.verticals grant exactly like the dedicated-URL model already does -- so a customer's existing vertical grant governs both surfaces identically and requires no separate migration step once the unified catalog goes live. Also wired admin.users.verticals into the same check for OAuth identities (the coarse "which products" grant, alongside the existing fine-grained tool_access patterns) -- toggling a family on for a customer in the admin UI now governs the unified catalog the same way it already governs the dedicated vertical URL. Verified with 7 direct can_see() cases (sedar-only bearer sees sedar not bio; break-glass and open-kind unchanged; 'full' grant sees everything; graph-only bearer sees graph not sedar).

resolve_ticker_to_cik gains as_of — point-in-time ticker resolution (S=Sprint->SentinelOne class reuse)

get_company_profile get_fundamentals and every tool that resolves a ticker via db/qdrant.py

For consumers: The shared ticker->CIK resolver used across the catalog now accepts an as_of date for point-in-time resolution over ticker-reuse eras (a ticker symbol changing hands between unrelated companies over time — verified live: "S" resolves to Sprint Corp for dates in 2005-2023, flips to SentinelOne for 2024 onward, the same real symbol). Implements atlas's exact era-read contract (inclusive bounds + an explicit tie-break on the shared boundary date) rather than a simplified version, since the underlying data mixes two incompatible conventions for what a range's end date means and a naive "cleaner-looking" query silently drops results on the last day of many eras. Purely additive — omitting as_of preserves the existing current-occupant behavior exactly. Historical resolution returns cik/company_name/valid_from/valid_to only; other metadata (exchange, entity type, SIC) has no point-in-time source yet and comes back empty rather than incorrectly showing today's value as if it held historically.

graph_neighbors — consumes catalog-declared dedup_key_cols (fixes works_at duplicate edges)

graph_neighbors

For consumers: Fixed a duplicate-edge bug on the works_at graph edge — the same person could appear twice for one real employment relationship when the underlying extraction pipeline had multiple raw rows for the same fact (e.g. a bio-text extraction and a web-crawl extraction of the same job). graph_neighbors now consumes a new catalog mechanism (graph_edge_catalog.dedup_key_cols + dedup_order_by, mars-declared) that collapses genuine source-row duplicates to one edge per real fact, picking a deterministic winner (curated source over mechanical, titled over blank, higher confidence, more recent). Purely additive — every other edge type has no dedup key declared and is unaffected; new edges can opt in with zero argos-side code changes, same zero-deploy property as the rest of the catalog-driven graph.

thirteenf_holders / thirteenf_who_holds — fixed a bug that silently dropped Vanguard (and every other multi-CIK family) from holder lists

thirteenf_holders thirteenf_holders_full thirteenf_who_holds

For consumers: Fixed a significant correctness bug — Vanguard (and any other institution whose 13F filings are split across multiple sub-CIKs rolled up via cik_canonical) was being silently EXCLUDED from holder lists entirely, not just under-counted. Root cause: Vanguard's canonical CIK has no row of its own in the underlying table (its real filings are all under 8+ sub-entity CIKs); the prior query excluded every non-canonical row outright rather than rolling them up, a workaround originally built for a different problem (a filer-identity-transition duplicate that needed deduping, not excluding). The fix distinguishes both cases correctly — sums genuinely distinct sub-entity positions under their canonical holder, while still collapsing true duplicate rows to prevent double-counting. AMZN's total 13F value moved from ~$1.455T (undercounted) to the correct ~$1.655T as a direct result; Vanguard now correctly appears as a top holder (often #1) wherever it should. This likely affected most large-cap tickers, since Vanguard holds a position in nearly all of them.

graph-mcp.staticpipe.com grows 7 → 13 tools (graph_ledger + its reference set)

graph_ledger mars_funding_history mars_mergers_for_company mars_investor_portfolio donor_person_track_record donor_person_profile

For consumers: After the graph_ledger 6-node pilot passed its acceptance test, kee added graph_ledger plus the 5 vertical tools its response's `tools`/`dig` pointers reference to graph-mcp — a served ledger's pointers must resolve on the surface that serves it. graph-mcp goes from 7 tools (the bare graph primitives) to 13. graph_ledger itself stays precomputed-rows-only for now; the on-demand pre-walk path and further entity-count expansion remain separate, later decisions.

graph_ledger — PILOT precomputed neighborhood-digest tool (dev only, 6-node scope)

graph_ledger

For consumers: New graph_ledger(node_id, node_type?) tool — a precomputed, pre-joined neighborhood digest for one graph node (profile + ranked/fused neighbor sections + next_hops), so an LLM starts several steps into a relationship investigation instead of walking graph_neighbors cold. Pilot scope only: serves exactly 6 precomputed nodes from mars.entity_ledger_v2 (2 people, 4 companies); every other node_id returns an honest found:false, no live fallback yet. Each section's `dig` pointer (a ready graph_neighbors(...) call) is checked against this server's own live catalog before being returned, so a pointer you get back is always actually callable. Not on graph-mcp yet — that and the on-demand path both wait until the pilot's acceptance test passes.

mcp-insiderfocus.kscope.io — new dedicated vertical serving the 8 names_* tools (People & Insider Focus)

names_profile names_search names_affiliations names_transactions names_holdings names_vesting_calendar names_990_boards names_lookup

For consumers: A new dedicated MCP server for a sellable named-client deployment — serves exactly the 8 names_* tools (People & Insider Focus data from names-api) plus resolve_company_name/resolve_company_profile and discover_tools, nothing else. Standard 5-piece stand-up (env, systemd service on port 8013, target group, 4 ALB rules, OAuth+static-bearer auth) mirroring every other kscope.io vertical. Raul's existing OAuth login already has access; DNS (the registrar CNAME) is the one remaining step, pending Raul.

resolve_investor (mars_investor_portfolio / mars_funding_search) — exact distinctive-token-set match beats a superset match

mars_investor_portfolio mars_funding_search

For consumers: Fixes a resolver bug found via a batch of nexus regression reports (2026-09-21/22). A query like "AIG Asset Management" was resolving to an unrelated entity ("AVC AIG GP, LP") instead of the real "AIG" — word_similarity's length-neutral matching (needed for partial-name queries like "Arex Capital") scored the single-word anchor "aig" at 1.0 against both the true single-word target AND any longer unrelated name merely containing "aig" as one word among several, and a downstream CRD-presence tiebreak then promoted the wrong one. Now a candidate whose own distinctive-token set exactly equals the query's wins outright over any superset match. Zero behavior change for queries without an exact-token candidate (Blackstone/Amundi/Sequoia Capital/State Street/Berkshire Hathaway all verified unchanged); one incidental improvement (Stanford Management Company's ambiguous list tightened from 15 noisy candidates to 2 plausible ones). "U.S. Bank Trust Company" still correctly returns no match — no entity by that literal name exists in the underlying table (closest: "U.S. Bank National Association"), so refusing to guess remains correct.

donor_person_track_record — Universal Track Record, joins bankrupt/delisted/acquired outcomes per company

donor_person_track_record

For consumers: donor_person_track_record now joins outcome classification onto every returned company (include_outcomes, default true) — generalizing bio_person_track_record (SEC Form-4/biotech only) to every person in the mars graph. Three independent flags per company, not a collapsed verdict — a company can be both bankrupt AND later acquired (kept separate, not overwritten): bankrupt (bankruptcy_events_union_v — unambiguous), delisted_distress (delisting_events_v2 filtered to is_distress=true — the table also carries benign delistings like merger-driven voluntary delisting and exchange upgrades, which are excluded), acquired (merger_deals_v2, definitive + verified/vouched, the target side). Each flag carries a date. A response-level outcomes_summary rolls up counts. Honest caveat baked into the tool description: absence of all 3 flags means "no adverse outcome found in these sources," not "confirmed still healthy" — there's no active/defunct status column to assert that. Acceptance case (verified live): Ned L. Siegel's richest identity — 28 relationships / 19 distinct confirmed companies / 1 bankrupt / 4 distress-delisted / 4 acquired, including a company that was both bankrupt (2015) and later acquired (2019). Set include_outcomes=false to skip the 3 extra batch lookups if you only need raw relationships.

donor_property_screen — new min_network_links filter, a real connectivity metric (not just connected_only's boolean proxy)

donor_property_screen

For consumers: donor_property_screen's parcels mode gains `min_network_links` — filters owners by real connectivity (mars.entity_network_degree_v2.total_count, entity-general, built off the full graph_edge_catalog) rather than just connected_only's boolean "resolves to a named person." Setting min_network_links implies connected_only. Two honest caveats baked into the tool description: (1) entity-resolution fragmentation — a real prominent person can be split across multiple person_ids with very different degree counts, so a real prospect can be missed on a specific property tied to an under-resolved fragment (verified live: "Paul Tudor Jones II" has 3 distinct person_ids at total_count 168/6/no-row); if results look thin, also try connected_only=true alone. (2) The response may return fewer rows than requested if the internal search batch is exhausted before enough owners clear the threshold — flagged via the new `batch_exhausted` field, not silently treated as "no more matches."

New tool donor_property_screen — screen the deeds real-estate corpus by value, price, date, and geography

donor_property_screen

For consumers: New tool, sibling to donor_property_holdings (which is scoped to one person). donor_property_screen screens the WHOLE property-records and sale-events corpus — "who owns real estate matching these criteria" rather than "what does this one person own." Two modes: parcels (default, ranked by assessor value; `connected_only=true` restricts to owners resolved to a named person) and sale_events (buyer/seller transaction records, by price or date). Requires a narrowing filter (min_just_value for parcels; min_sale_price or sale_date_from for sale_events) — only those columns are indexed against the 20M+/19M+ row corpus. `connected_only` is a proxy for "resolves to a named owner," not a literal count of network connections — that data doesn't exist yet. Coverage is partial (~10 counties for records; King County WA only for sale events).

mars_mergers_for_company now resolves bare/short company names (was requiring the full legal suffix)

mars_mergers_for_company

For consumers: mars_mergers_for_company migrated onto the same universal resolver every other mars_* tool uses (previously it was the last one on an older, narrower resolver). A bare or shortened company name — e.g. "Imperative Care" instead of "Imperative Care, Inc." — now resolves the same as the full legal name, the company_id UUID, or the domain. It also now returns an explicit `error: "ambiguous"` with a candidate list when a name genuinely matches more than one company, instead of silently guessing. `resolved_via` values changed from {uuid, domain_match, name_match} to {uuid, domain, cik, ticker, name} to match the shared resolver's vocabulary — update any code that branched on the old string values.

Investor resolver hardening v2 — acronym handling, all-generic queries, and partial-name retrieval

mars_funding_search mars_investor_portfolio

For consumers: Three follow-on improvements to the shared investor resolver (same `investor` filter as the earlier over-match fix). (1) A period-separated acronym like "U.S." now tokenizes correctly as an identifier instead of losing its anchor entirely — "U.S. Bank Trust Company" now returns no match instead of a misleading unrelated firm. (2) A fully-generic query with no identifying content ("Retirement System", "Bank Trust Company") now returns no match instead of a list of unrelated organizations sharing only the generic phrase. (3) A partial company name ("Arex Capital" for "Arex Capital Management, LP") now resolves directly where it previously returned nothing, for names rare enough to trust unambiguously; a name too common to disambiguate (e.g. a bare "First Capital") still correctly returns either no match or an explicit list of candidates rather than guessing.

Investor resolver — fixed generic-phrase over-matches (Tiger Global / Veridian Private Wealth / Arex Capital / AIG Asset Management class)

mars_funding_search mars_investor_portfolio

For consumers: The shared investor resolver (used by the `investor` filter on mars_funding_search and mars_investor_portfolio) could silently return unrelated firms for a query sharing only a common descriptor phrase with the real target — e.g. "Tiger Global Management" returning Voyager/Toro/Meyer "Global Management", or "Arex Capital" returning CP/ACME/B5 "Capital" while dropping the real Arex Capital Management LP entirely. Fixed: added "global"/"private" to the generic-token stoplist, and closed a leak where the fallback fuzzy-match path wasn't applying the same generic-overmatch filter as the primary path. A query with no confident anchor now returns no match rather than misleading unrelated results. If you were working around noisy investor results by post-filtering, that's no longer necessary for this class of query.

mars_mergers_for_company now accepts a company_id UUID (was returning "no mergers")

mars_mergers_for_company

For consumers: mars_mergers_for_company was the only mars_* tool still on the legacy name/domain resolver, so a bare company_id UUID — the chain-in identifier every other mars_* tool emits (mars_company_resolve, mars_funding_history) — returned resolved_via="not_found" with an empty mergers list, even when real merger data existed. Fixed — it now resolves a UUID directly (resolved_via="uuid"), matching the other tools. Domain and name inputs are unchanged. If you chained company_id → mars_mergers_for_company and got empty results, re-run.

13F tools (thirteenf_*) moved to Postgres — full holder base + classification; response-shape change on the flow sub-structures BREAKING

thirteenf_holders thirteenf_portfolio thirteenf_who_holds thirteenf_overlap thirteenf_screen thirteenf_search_managers

For consumers: All six thirteenf_* tools now read proc2 holdings Postgres instead of the retired Qdrant 13F snapshot collections (Qdrant is being wound down for capacity). NET BETTER on the core data — thirteenf_holders / thirteenf_portfolio now return the HONEST FULL holder base / position book (NVDA = 5,927 holders, not a top-20 cap) with per-holder manager classification (manager_type / tier / is_notable) + qoq_action + top-level v3 signals. WHAT CHANGED IN THE SHAPE — the Qdrant-only precomputed sub-structures are no longer returned: the separate new_positions/exits/increases/decreases flow-event buckets (with bucket_rank), the notable-only holder subset as its own list, and the inline quarterly_history. The same information is available differently: per-holder is_notable + qoq_action (new/add/trim/exit/unchanged), the top-level v3 signals (is_cluster_entry / consecutive_accum_qtrs / accel_4q / …), and thirteenf_position_history for a manager+security multi-quarter trail. Also: thirteenf_search_managers name lookup is now case-insensitive; thirteenf_overlap uses full holdings (no top-50 cap). If you parsed the new_positions[]/exits[] arrays or the notable-subset list off these tools, switch to the per-holder fields.

mars_company_resolve + get_company_profile now return entity_kind (fund / SPAC / shell vs operating company)

mars_company_resolve get_company_profile

For consumers: Both tools now surface entity_kind, so a consumer can tell a fund / SPAC / shell (which correctly has no sector / funding / M&A profile) from a real operating company that simply has no data yet. Values: operating_company / operating_company_thin / op_co_no_domain / investment_vehicle / spac_shell / foundation_nonprofit / regulatory_body / individual / sec_shell_unknown / unknown. Requested by nexus, who hit The Taiwan Fund (a 1940-Act closed-end fund) rendering as an empty operating-company card when it's correctly-shaped data for what it is — entity_kind="investment_vehicle" now lets them build an honest fund/shell display path. On mars_company_resolve it's a native column (also on the ambiguity candidates); on get_company_profile it's a best-effort mars lookup by CIK (non-fatal — a cascade/Qdrant tool, so a mars miss just yields null). Verified: TWN -> investment_vehicle, AAPL/MSFT -> operating_company.

investor resolver — reject generic-phrase over-matches (Amundi Asset Management)

mars_investor_portfolio mars_funding_search

For consumers: Follow-up to yesterday's State Street fix, handling the broader case: "Amundi Asset Management" returned unrelated firms sharing only the generic phrase "Asset Management" (BQ / GC / MWA / APG … Asset Management) — the normalized-exact re-rank didn't catch these because they don't normalize to the same key. General fix: a fuzzy candidate is now kept only if it shares a DISTINCTIVE (non-generic) token with the query. Corporate-form / ultra-common finance words (asset, management, capital, partners, corporation, group, holdings, advisors, ventures, fund, associates, equity, llc/inc/lp …) no longer count as a match by themselves; the identity token (amundi, bridgewater) has to overlap. Sector words (real estate, energy, health) are deliberately NOT treated as generic since they can be distinctive. Verified: "Amundi Asset Management" → only the 3 genuine Amundi entities; "Bridgewater Associates" → the real one; State Street / Blackstone / KKR unchanged. Both mars_investor_portfolio and mars_funding_search (shared resolver). Thanks nexus.

mars investor name resolver no longer over-matches on generic tokens (State Street Corporation)

mars_investor_portfolio mars_funding_search

For consumers: BUG FIX — resolving an investor by a correctly-spelled multi-token name returned junk candidates that share only generic tokens (nexus: "State Street Corporation" surfaced "Main Street Capital Corporation" / "Centre Street Cambridge Corporation" / "SRB Corporation" as ambiguous matches). Raw trigram similarity() over-scores names sharing common words ("Street"/"Corporation") — it ranked Main Street Capital (0.543) ABOVE the real State Street (0.478), and the confidence floor then dropped the real one and kept the junk. Fix: re-rank the fuzzy candidate set by NORMALIZED-EXACT first (strip the legal suffix — "state street corporation" and "state street" both normalize to the same key), and treat a normalized-exact match as a confident resolution not subject to the similarity floor. "State Street Corporation" now resolves cleanly to State Street; "Main Street Capital Corporation" resolves to itself; typos/exact/miss unchanged. Both mars_investor_portfolio and mars_funding_search use the shared resolver. Thanks nexus / mars-spider-optimization.

adv_manager_people — who runs / owns a management company (Form ADV Schedule A/B)

adv_manager_people

For consumers: The people + ownership complement to the adv_manager_* tools (which give RAUM / fees / disciplinary). Given a firm CRD, returns the executives / control persons (Schedule A) and the corporate ownership chain up to the ultimate parent (Schedule B) from Form ADV — e.g. CRD 169355 (HealthCare Royalty Management) → Clarke Futch (Chairman & CEO, control person, 50-75%), Christopher White (President/CFO), + the full KKR ownership chain (KKR Harbor Holdings GP 75%+ → KKR & Co. Inc. …). People are deduped across Schedule A/B with their most-specific title; ADV surname-first names ("FUTCH, CLARKE, BRANNEN") are reordered to natural order for display; person_id chains to the person graph when resolved (~12%). Thin reader over mars.firm_owners_v2 (~139.9K owner rows / 25.3K firms). Substrate + canonical spec by bios/robin; requested by nexus for an Investors-page "who runs this firm" section.

mars_investor_portfolio / mars_funding_search — numeric CRD no longer 500s

mars_investor_portfolio mars_funding_search

For consumers: BUG FIX — passing a real CRD as a string, e.g. mars_investor_portfolio({"investor": "169355"}), threw a raw SQL error ("operator does not exist: bigint = text") instead of resolving. The shared investor resolver cast the param the wrong way (crd = %s::text) against the bigint crd column. Fixed to cast to ::bigint; "169355" now resolves (HealthCare Royalty Management). Also hardened: a CRD sent as a JSON number (not string) no longer AttributeErrors, and the resolver call is wrapped so any failure returns a clean error shape instead of a raw SQL string at the MCP boundary. Both fixes flow to mars_funding_search too (same resolver). Thanks nexus.

mars_funding_history surfaces raw investors when the canonical bridge lags (was empty + "no backers named")

mars_funding_history

For consumers: BUG FIX — mars_funding_history returned canonical_investors: [] plus a "source didn't name backers" note for rounds whose backers ARE named in the source but haven't been canonicalized yet (nexus: Joby's $200M round shows Morgan Stanley + Allen & Co. in the raw data, the tool showed nothing). The canonical bridge (deal_investors_v2) lags funding_deals_v2 — ~62K deals carry named investors in the raw `investors` field with no bridge rows yet. Now, when there are no canonical rows, the tool falls back to the raw names and surfaces them tagged position="uncanonicalized" with has_named_backers=true and a note that investor_id is null (so you know you can't chain to mars_investor_portfolio until the bridge resolves them). Genuinely-empty rounds still get the round-type "not itemized" note. So an empty list is never ambiguous: canonicalized, named-but-not-yet-canonicalized, or truly no backers. Thanks nexus / signals_v2.

graph_resolve ~850ms flat -> ~1-20ms (type-gated indexed cascade)

graph_resolve

For consumers: PERF FIX — graph_resolve cost a flat ~850-920ms on EVERY call, including a bare ticker or a total non-match (nexus). The cascade ran all resolver tiers with OR'd non-indexable predicates (cik::text=, upper(ticker)=, =ANY(name_variants)) that seq-scanned companies_v2 + persons_v2 every time. Now each tier is TYPE-GATED (numeric -> cik/crd btree; token -> ticker/domain/variants indexes; name -> institution name index + the companies normalized-suffix btree) and every branch is index-backed, so a strong ID resolves in ~1-20ms and never seq-scans. Warm: NVDA/AAPL 7ms, Stanford University 1ms, cik 22ms (were 847-921ms). Also fixed a resolution bug surfaced along the way: a domain shared as primary_domain by two rows (a real NVIDIA + a mis-tagged dupe) resolved to the wrong shorter-named one; strong-ID lookups now prefer the tickered/public company. Names resolve institution-first (Duke), strong IDs never seq-scan. Thanks nexus.

mars_company_resolve ~1s -> ~1ms on legal names (Apple Inc. / Stripe, Inc. / nvidia)

mars_company_resolve

For consumers: PERF FIX — several common name queries cost ~1-1.3s with a surprising inverted pattern (nexus): a verbatim legal name like 'Apple Inc.' / 'Stripe, Inc.' was SLOW while a bare 'stripe' was 17ms. Root cause: a legal name misses the exact-match phase (stored as 'APPLE INC' / 'Stripe Inc'), then its suffix trigrams (', in', 'inc.') blow up the trigram GIN scan to ~245K row rechecks (~1s); a short name like 'nvidia' clears that scan but scores below the 0.75 confidence floor on length-penalized similarity() and falls through to an unindexed name_variants scan (~1.2s). Fix: a new normalized-exact phase strips the legal suffix + punctuation and matches the indexed normalized-name btree (companies_v2_normstrip_idx) — 'apple inc.'→'apple', 'nvidia'→'nvidia' resolve in <5ms and skip both expensive phases. Verified: all six of nexus's cases now 0-1ms (nvidia→NVIDIA Corporation/NVDA, Apple Inc.→AAPL); distinct public companies colliding on a normalized name are flagged ambiguous. Typo/nonsense misses still fall to the fuzzy phases as before. Thanks nexus.

donor_person_resolve ~4.7s flat -> ~20ms (trigram index) + fixed broken exact-ID lookup

donor_person_resolve

For consumers: PERF FIX — donor_person_resolve cost a flat ~4.7s per name regardless of input, because the name match ran a leading-wildcard LIKE on the wrong column OR'd with a per-row unnest(name_variants) scan across all 3.7M person rows (neither indexable). Now the primary-name match uses the trigram GIN index (persons_v2_name_trgm on lower(primary_name)), and the name_variants substring scan runs ONLY as a fallback when the canonical name finds nothing (reordered / alternate spellings) — so a real, findable name never pays it. Warm timings: Ellen Ochoa 4836->19ms, Warren Buffett ->25ms, single-char 'x' ->873ms. Recall preserved; reordered names still resolve via the fallback. Also fixed a separate latent bug: exact-ID lookups (by person_id / crd / cik with no name) were fully broken (emitted `ORDER BY 0`, which Postgres reads as an invalid ordinal) — now 1ms. Thanks nexus (dogfooding argos-dev).

graph_neighbors no longer 500s on fund/foundation edges (default unfiltered walk)

graph_neighbors

For consumers: BUG FIX — graph_neighbors(node_id) with the default (unfiltered) edge set hard-crashed ("Error executing graph_neighbors: 'fund'" / "'foundation'") whenever the walk touched an edge whose far end is a fund or foundation node — reproducible on both a person node (via board_member_of_foundation) and a company node (via held_by_fund). The node-name registry only had the 4 uuid-keyed types (company/person/investor/institution); the newer fund (nport_fund_registry_v2) and foundation (foundation_orgs_v2, keyed by EIN) endpoint types KeyError-ed at query construction, 500-ing the whole walk. Fixed: the registry now covers fund + foundation, and an unknown endpoint type surfaces the edge with a null name rather than crashing. Verified: company → held_by_fund → "LVIP State Street Bond Index Fund" (with value/par/pct_of_nav signals); investor → manages_fund → Nuveen funds; person → board_member_of_foundation → "BABSON COLLEGE". No more need to pre-restrict edge_types as a workaround. Thanks nexus (dogfooding argos-dev).

Merger tools now show resolved company names (was blank for ~11K resolvable deals)

mars_mergers_screen mars_mergers_for_company mars_trending_acquirers

For consumers: The three merger tools rendered the counterparty from the RAW article name field (target_name / acquirer_name), which is NULL on ~11,038 target / ~10,607 acquirer deals that DO resolve to a real company node — so a known company (Tableau, Slack, MuleSoft on a Salesforce M&A view) rendered blank. Now they read the RESOLVED canonical name (companies_v2.primary_name joined on target_company_id / acquirer_company_id) and fall back to the raw field only when the id is unresolved. More correct anyway — the canonical name, not one article's spelling. Verified: Salesforce acquirer view → 0 blank counterparties (Spindle AI / Regrello / Convergence.ai / Slack / CLICKSOFTWARE / Tableau / MuleSoft all render). Thanks bios (ticket #774).

graph_neighbors — fixed empty-signal edges (silent 0-edge crash) + optional dst_name_col name fallback

graph_neighbors graph_entity graph_schema

For consumers: Two graph builder fixes surfaced by a bios audit of Raul's Goldman Company-360. (1) BUG FIX — any catalog edge with empty signal_exprs (e.g. owns_subsidiary) emitted a malformed SELECT (double comma) that was caught+swallowed, so the edge silently returned 0 edges. The SELECT list is now assembled robustly; no-signal edges walk correctly. (2) NEW — an optional catalog field `dst_name_col`: when the owning vertical sets it to a name column on the bridge table (e.g. company_subsidiaries_v2.subsidiary_name), graph_neighbors LEFT JOINs the canonical and falls back to that name, so edges whose destination company_id isn't resolved yet still surface BY NAME (target_node_id=null until entity-resolution backfills it). Verified: with dst_name_col set, Goldman's 44 subsidiaries surface by name. Both are catalog-driven — mars/bios set dst_name_col and it activates with zero argos deploy.

SEC agreements tools retired (data migrated off Qdrant to mars)

agreements_search agreements_get agreements_browse agreements_find_similar

For consumers: The four `agreements_*` tools (SEC contract semantic search over EX-10.x + 8-K Item 1.01) are retired. The agreements team migrated off Qdrant to mars postgres under the 2026-08-04 mandate, and the underlying agreements/clauses/meta Qdrant collections were dropped — so the tools had been returning "collection not available." They were dev-only with ~zero usage. The agreements corpus now lives in mars (`agreement_deals_v` + the `signed_agreement` graph edge); reach it via the mars/graph surfaces. Also removed from the mergers vertical's catalog. The agreements vertical URL stays up as a dormant shell for now.

sections_search now searches the full 2016-2026 corpus (was capped at 2024-2026)

sections__search

For consumers: `sections__search`'s omit-year default was hardcoded to 2024/2025/2026 — but the 10-K/20-F/S-1 corpus expanded to full 2016-2026 depth back in July. Cross-company semantic sweeps that omit `year` now fan out across every live year (verified: a "material weakness" search now returns hits spanning 2018-2026, previously 2024+ only). The year list is discovered dynamically from Qdrant (cached) so it tracks the corpus as it grows and won't drift again — each form family reports its true depth (10-K back to 2016; the separate 10-Q collection remains 2024-2026 until sections extends it). `sections__get_filing` was already full-depth; the two tools now agree. Thanks to flowlab (via sections) for flagging.

graph_path — bounded relationship-path finder (warm-intro / degrees-of-separation)

graph_path

For consumers: The "Paths In" / Find-Path / warm-intro-meter primitive. `graph_path(from_id, to_id)` finds the shortest relationship path(s) between two people — or a person and a company — over the person↔person `connected_to` graph (16.6M professional/board/deal co-occurrence ties), depth ≤ max_hops (default 3). Returns each path as a node sequence with per-hop `tie_strength` + an overall `strength_product` (short strong paths rank first) + `strength_min` (weakest link); top-level `connected` + `min_hops` are the warm-path meter ("reachable via N hops"). If `to_id` is a company, the final hop is a board-member / officer / employment edge into it (person →connected_to→ person →board_member→ company). Bidirectional beam-BFS — bounded, surfaces strong short paths, sub-second even from a 1,200-degree hub. Resolve endpoints with graph_resolve. Verified: 2-hop person→hub→person and person→connected→person→worked_at→Oruka Therapeutics both in <0.15s.

donor_person_track_record — a person's board/officer/founder/employment history, trust-tiered

donor_person_track_record

For consumers: The due-diligence surface for a person — every company they've been a board member / officer / founder / employee of, each tagged with its **trust tier** (`edge_status` = confirmed | candidate) and verbatim provenance (source_quote / source_url). Chain the returned `to_company_id`s into outcome tools (bankruptcy / litigation / M&A) to answer "has this person been near bad outcomes." **`min_edge_status` (default `confirmed`) is the key knob** — pass `candidate` to ALSO surface leads (an uncorroborated "director → X" is a DD lead to chase, not noise; the verifier honestly demotes obscure micro-cap directorships to candidate). Live over the base person-relationship table (always fresh — NOT the confirmed-only nightly person_network MV, which is why candidate edges are visible here). 4 relation types (board_member / officer_of / founder_of / worked_at); affiliated + studied_at excluded. Verified: Eugene I Davis (professional restructuring director) → 72 confirmed + 10 candidate leadership edges. Resolve a name → person_id with donor_person_resolve first.

graph_aggregate — the screen layer over the entity graph (rank entities by an aggregate over their edges)

graph_aggregate

For consumers: The cross-node complement to graph_neighbors' single-node walk. `graph_aggregate(edge_type, group_side, metric)` ranks entities BY an aggregate over their edges — "most widely-held stocks" (holds/dst/n_counterparties → Alphabet 6,205 holders), "biggest $ held" (holds/dst/sum:value_usd → iShares $3.1T), "most-targeted by 13D filers" (filed_13d_on/dst/n_counterparties), "busiest placement agents" (placed_offering_for/src/n_counterparties → Emerson Equity 714 issuers), "most-interlocked directors" (board_member_of/src/n_edges). Metric ∈ {n_edges | n_counterparties | sum:<signal>}; filters min_value + node_type. Thin reader over a mars-materialized nightly rollup (top-N is a <15ms index scan vs a multi-second/timeout live GROUP-BY) — catalog-driven, so every graph edge is screenable and new edges auto-join. Discover edges + their signals via graph_schema.

Two new donor-family tools — 990 foundation donor graph + RIA advisor mobility

donor_foundation_affiliations donor_advisor_mobility

For consumers: Two prospect-research reads for the donor family. **`donor_foundation_affiliations`** — the IRS Form 990 donor graph (276K people × 15.6K foundations): who sits on / officers which foundations, joined to that foundation's 990 financials (assets/revenue/contributions/grants-paid). Three modes — by person, by foundation, or by criteria ("people on boards of $100M+ grantmaking foundations"). Pass `ntee_prefix='T'` to filter to actual grantmaking foundations (drops realtor/credit-union boards that otherwise top the headcount lists). **`donor_advisor_mobility`** — RIA advisor movement between firms (recruiting/breakaway signal): per-firm gains/losses/net advisor flow (1y/3y/all) + current RAUM, plus firm→firm flow pairs ("where did Avantax's advisors go?"). Both are thin readers over bios-built serving views; validated (Osaic Wealth net +50 advisors/3y, Avantax→Cetera 14).

Two new tranche tools — corporate-bond price-dispersion screen + trend

tranche__bond_price_dispersion_screen tranche__dispersion_trend

For consumers: Two new reads over the N-PORT-P fund-mark panel that turn cross-fund mark *disagreement* into a public pricing-uncertainty / distress signal — useful for illiquid or distressed corporate bonds with no recent TRACE prints. **`tranche__bond_price_dispersion_screen(as_of?, min_funds=10, top_n=25)`** ranks CORP bonds by how much funds' marks diverge at a report date (absolute `price_stddev` + `dispersion_pct` CV), returning `price_min`/`price_max`/`n_funds` so you can see the spread — on settled quarters the top rows are defaulted/distressed bonds where funds' recovery marks legitimately split (some near par, some near 0: e.g. Saluda Grade min 0.43 / max 99.7). **`tranche__dispersion_trend(cusip, n_months=24)`** is the stddev trajectory over time — rising dispersion is a mark-uncertainty regime change / credit-event leading indicator. Both are N-PORT-P fund-held sample (~$3T, not total marks outstanding); read `n_funds` alongside stddev since dispersion off 2-3 funds is noisy, and the very latest report date is often thinly filed.

get_fundamentals + company profile now read live proc2 XBRL (off frozen Qdrant)

get_fundamentals get_company_profile

For consumers: The XBRL fundamentals dataset moved to a new home and is now refreshed hourly again. `get_fundamentals` and the fundamentals block in `get_company_profile` read the live source directly — same fields, same schema, same output, just no longer drifting stale. No consumer action needed.

Graph is now CATALOG-DRIVEN + new graph_schema tool (self-describing graph)

graph_neighbors graph_entity graph_schema

For consumers: The graph primitives now read their edge definitions from `mars.graph_edge_catalog` (a substrate table mars owns) at runtime instead of hardcoded specs — so **a new edge is usable the moment mars adds a catalog row, with zero argos deploy**. Plus a new **`graph_schema(query?)`** tool: discover what's traversable via E5 semantic search ('board seats', 'M&A', 'edges from a person to a company') — searchable not a dump, returning each edge's correct-use metadata inline (signals to rank by, the mandatory integrity/edge filters graph_neighbors applies for you, pit_column, coverage). This is the "self-describing graph" — everything mars adds to the graph becomes automatically discoverable + traversable. Behavior-preserving for the 10 existing edges; consumer-contract filters (integrity_tier, is_definitive) stay baked in per the catalog.

sec_power_fundamentals (new tool) + bio biosimilar flag + insider_holdings is_active

sec_power_fundamentals bio_fda_approvals insider_holdings

For consumers: **New `sec_power_fundamentals(ticker_or_cik)`** — a utility/IPP's year-by-year operating panel in one call (capacity adds/retires by fuel + firm demand + YoY + outages per year) + identity + rate-case history + a per-dataset `coverage` flag. For inferno's SIC-routed power section; 4 handler-side queries over the power_sec panels. `coverage` is load-bearing (a 0 demand can mean "none extracted," not "none exists"). Rate-case honesty: surfaces `allowed_roe` (outcome) with an `allowed_roe_suspect` flag outside [3%,20%]; does NOT fake-map requested_roe/rate_base/commission-filed_date (v2). dev-only. **Plus two data-team fixes**: `bio_fda_approvals` now carries a `biosimilar` flag + `exclude_biosimilar` param (super-bio T030 fixed the biosimilar→originator mismap — a company only carries its OWN biosimilars now, competitor contamination removed); `insider_holdings` defaults to the active insider book (`is_active`, position reported <3y — Steve Jobs et al. correctly drop) with an `include_inactive` opt-out.

mars_mergers_screen — first_seen_at freshness fields + min_freshness filter

mars_mergers_screen

For consumers: **`mars_mergers_screen`** now surfaces `first_seen_at` + server-computed `days_since_first_seen` + `freshness_class` (fresh <7d / recent <30d / aging <90d / stale) on every deal row, so you can judge staleness before trading on it, plus an optional `min_freshness` filter. The tightest Kalshi ACQ filter is now `is_definitive=true` + `min_freshness='fresh'` + verified/vouched integrity — only definitive M&A deals mars saw this week. (Freshness for the funding-side tools follows.)

credit_cross_held_loans — BDC NAV front-run screen (new tool)

credit_cross_held_loans

For consumers: **New `credit_cross_held_loans`** (private-credit team spec, crown's BDC-NAV-front-run thesis) — shared private loans where an early-reporting BDC just marked the credit down / to non-accrual but late-reporting BDC holders (on an older 10-Q quarter) haven't caught it yet. Those stale holders are the short/put candidates. Per shared loan returns each holder's latest mark + `is_stale` + `tradeable`, plus a signal block (`mark_spread`, `fresh_markdown` = biggest cut by a fresh reporter, `stale_public_holders` = the front-run targets). `signal` ∈ all / markdown_divergence / non_accrual_divergence; `holder_ticker` (what is BDC X exposed to that's already been cut); `company_id` drill-down; PIT `as_of`. All from public SEC 10-Q Schedules of Investments (public-filing cross-referencing, NOT MNPI). 14 BDCs, 912 loans held by ≥2, data through 2026-03-31. Reads `mars.bdc_cross_held_marks_mv`. dev-only (credit_ family) + mars vertical.

bio_crl_outcomes summary_only + supply_curve unit_key dedup + mars merger deal_stage filter

bio_crl_outcomes sec_power_supply_curve mars_mergers_screen

For consumers: **(1) `bio_crl_outcomes` `summary_only=true`** — returns just the `by_reason` base-rate aggregate + counts, omitting the per-event list. A base-rate pull was force-returning all ~400 event rows (130K+ chars) and blowing token ceilings; summary mode is ~40× smaller. Also added `n_with_abn_126` per reason bucket (horizon-availability visibility). **(2) `sec_power_supply_curve`** now dedups the as_of path on the normalized `unit_key` (not raw name) — the 8-K + 10-K extraction paths disclose the same physical unit under name variants ("Comanche 3" / "Comanche Unit 3" / "Comanche Generating Station Unit 3") that were multi-counting; unit_key collapses them (stacked with the existing quality_flag filter for the most-correct unit-attributable curve). **(3) `mars_mergers_screen`** gains an `is_definitive` filter (real deals only — deal_stage IN definitive/pending/closed, excludes rumor/talks/loi/terminated; hits a partial index) + surfaces `deal_stage` + `is_definitive` per row. The "how much is real vs rumor" filter for M&A screens.

fund_equity_holders + fund_equity_book — which ETFs own a stock (N-PORT equity)

fund_equity_holders fund_equity_book

For consumers: **New: "which ETFs / mutual funds own a stock"** — at FUND-SERIES grain, from the SEC N-PORT fund-holdings panel (equity slice). **`fund_equity_holders(ticker or cusip)`** returns the ranked fund holders (series_name, series_id, shares, value_usd, as-of) + breadth (n_funds, total held, top-5 concentration). **`fund_equity_book(series_id)`** is the mirror — an ETF/fund's equity positions. This is the granularity 13F structurally cannot give: `thirteenf_who_holds` returns institutional MANAGERS (BlackRock, aggregated across all its funds); these return the actual FUND SERIES (iShares Core S&P 500 ETF, Fidelity 500 Index Fund) with per-fund shares. Verified STRL → Sterling Infrastructure, 548 fund holders, $7.2B fund-held (iShares Core S&P Mid-Cap ETF top). Caveats: SEC-registered funds only (not the full holder base); as_of varies per fund (rolling quarterly snapshot); latest quarter + ~3.5y history now, full 6.5y backfill in progress.

sec_power_supply_curve — filter fleet-aggregate + outlier rows (quality_flag)

sec_power_supply_curve

For consumers: **Fix**: `sec_power_supply_curve`'s point-in-time (`as_of`) path reads `power_capacity_events` directly, so it now filters `quality_flag IS NULL` — excluding power-sec's newly-flagged `'aggregate'` rows (fleet/portfolio sentences extracted as one giant "unit", which also defeated the per-unit dedup and multi-counted the same capacity) and `'mw_outlier'` rows (garbage MW like Maine Yankee @ 20,943 MW). Effect: the curve is now unit-attributable capacity (a clean conservative lower bound) instead of fleet-inflated — coal retirements 2025 drop ~10.3→5.1 GW, 2027 ~8.9→1.5 GW. The non-as_of path (v_capacity_panel) was already sanitized server-side.

mars_funding_history role_canonical + sec_power_demand_forward firm/pipeline split

mars_funding_history sec_power_demand_forward

For consumers: **(1) `mars_funding_history`** — each `canonical_investors[]` entry now carries **`role_canonical`**: the investor's TYPE in that deal from mars's 16-bucket taxonomy (venture_capital / corporate_vc / angel / private_equity / hedge_fund / bank / underwriter / institutional / …), 89.5% populated (null when the source string was a position like "Lead Investor" not a type). Distinct from `position` (lead/participant) and the entity's own `type`. **(2) `sec_power_demand_forward`** — the response now surfaces top-level **`firm_total_mw`** vs **`pipeline_total_mw`** (+ firm/pipeline n_events + `firm_events_missing_mw`) so the firm-vs-speculative split can't be misread as a blended total. The trap it guards: a utility whose only demand rows are speculative interconnection-queue entries (AEP) shows `firm_total_mw=0` — read the firm total for real commitments, not a blended sum. Also documents that `expected_load_mw` is NULL on some firm rows (mw sum under-counts; use n_events for count).

sections__ipo_financials (IPO suite complete) + ipo_screen de-SPAC / earliest-ever fixes

sections__ipo_financials sections__ipo_screen sections__ipo_offering

For consumers: **(1) New `sections__ipo_financials(ticker_or_cik)`** — per-period financial statements (revenue / operating & net income / cash / balance-sheet, oldest→newest) extracted from the IPO prospectus, filling the "no 10-K history yet" gap for fresh IPOs. status ∈ ok / no_financials_extracted / not_in_universe; NULLs preserved. Completes the IPO suite (screen → risk_factors → offering → financials). **(2) `sections__ipo_screen` correctness fixes** (via mcptest): the "one row per company" rule is now **earliest-EVER non-SPAC 424B4** (not earliest-in-window) — a company that IPO'd in 2020 no longer shows as a "2026 IPO" for filing a 2026 follow-on (PDS Biotech etc. correctly drop; a since=2026 screen went 191→128 as false-positives cleared). Renamed issuers carry `current_company_name` (Classover → "KIDZ AI Inc."). New `dedup_mode='per_cik'|'per_accession'` — per_accession surfaces every 424B4 (follow-ons + re-pricings) if you want them. **(3) `sections__ipo_offering` de-SPAC fix fully propagated** — a de-SPAC issuer (AEON) now returns 2026 across the board (accession/ticker/date) with a null lockup instead of the pre-merger SPAC's boilerplate.

bio_similar_setups event_type='crl' — aggregate now computes (was empty)

bio_similar_setups

For consumers: **Fix**: `bio_similar_setups(event_type='crl')` was returning `aggregate: {note: "no populated forward_outcomes"}` even though the neighbors had real post-CRL abnormal returns. The per-horizon aggregate (median/mean/pct_up at 21/63/126/252d) only ran for `event_type='approval'`; crl uses the same `abn_*` shape but fell through to the quarter/excess_vs_xbi branch. Now both approval and crl compute the horizon aggregate. Verified SRPT crl → abn_252 median −0.86 (n=4). (mcptest Round 15.)

operating_metrics — one SIC-routed tool for per-company industry KPIs (cruise live)

operating_metrics get_company_profile

For consumers: **`operating_metrics(ticker_or_cik)`** returns a company's operating-KPI time series in the metric set standard for its industry — you pass a company, it routes by SIC to the right industry extractor and returns THAT company's values (per-company, not aggregates). **One tool for every industry** — new industries are added as internal routes, never new tools. **Route #1 live: cruise** (CCL / RCL / NCLH) → occupancy / ALBDs / net & gross yield / fuel / onboard revenue / newbuild pipeline / fleet departures / classified events (~5yr). A company outside a covered industry returns `status='not_covered'` + the covered list. The `industry` field says which schema you got. **Discovery**: `get_company_profile` now surfaces a `covered_industry` pointer (a hint, not the full series) when a company is in a covered industry, so an agent finds this tool at the moment it's looking at the company. Verified: Carnival occupancy 56%→75%→100%→105%→105% (post-COVID recovery); AAPL → not_covered. Airlines is the next route (no new tool).

sections__ipo_offering — one deal-terms + lockup card per IPO issuer

sections__ipo_offering

For consumers: Phase B of the IPO suite. **`sections__ipo_offering(ticker_or_cik)`** returns one structured card per issuer: the **offering terms** extracted from the 424B4 (price range, shares offered, greenshoe, exchange, ticker, deal size, underwriters, share classes, dual-class flag, net proceeds, use of proceeds) JOINed to the **catalyst lockup schedule** (days, expiry, shares locked, % of total, early-release trigger, tranche), with per-source provenance. Covers 98.3% of the catalyst IPO universe (2,522 issuers). `status` ∈ ok / lockup_only / not_in_universe. Pairs with sections__ipo_screen (find the IPO) + sections__ipo_risk_factors (its risk factors). Verified: SpaceX $85.75B / Airbnb 37 underwriters / Snowflake NYSE / Apple → not_in_universe.

mars_funding_history — empty canonical_investors[] now self-documents (has_named_backers)

mars_funding_history

For consumers: `mars_funding_history` now tells you what an **empty `canonical_investors[]`** means, so you don't misread it as a coverage gap. Per mars's coverage analysis it's GROUND TRUTH — the round's source filing/article didn't itemize backers, which is STRUCTURAL for IPO / debt / Reg D / private-placement rounds (Reg D 100% unnamed, IPO 29%, Debt 34%) vs Series A/B/C/Seed (~95% named). Each round now carries **`has_named_backers`** (bool) and, when empty, an **`investors_note`** that says whether the emptiness is expected for that round type or possibly a gap. So a name with only IPO/debt rounds (e.g. SpaceX's recent IPO/Debt/PP rows) legitimately shows the funding trajectory without namable backers — not a bug.

bio_catalyst_dossier — the next FDA catalyst + how to think about it (index with drill-downs)

bio_catalyst_dossier

For consumers: **`bio_catalyst_dossier(ticker_or_cik, as_of=)`** — one call returns a biotech's nearest PDUFA/AdCom date AND a compact headline for each context pillar (prior CRLs + the base rate for that reason, the target's historical approval base rate (target_base_rate — the key AdCom-efficacy prior), comparable historical approval setups with realized 1yr returns, management track record, manufacturing/inspection risk, pipeline) — **each with an exact `drill` tool call to go deeper**. It's an INDEX, not a dump: read the headlines, then follow the drill for whichever pillar matters and that tool gives you the full detail (and often its own pointers). Real-time fan-out over already-cooked data (base rates, PIT snapshots, CRL reasons, the people graph are all pre-computed) — nothing baked ahead, nothing stale; assembles in <1s. Token-cheap by design (headline + pointer, not the full record of each). Pillars degrade gracefully (a failed/empty read still returns its drill). The point: an agent asked "what's X's next PDUFA and how to think about it" tends to grab the date and stop — this surfaces all the pillars so it can't skip them. Verified: VRTX → prior CRL clinical_design (2016) → that reason's base rate (n=84, median abn_252 −0.34, 76% down) → 5 comparable approval setups (median 1yr abn −0.47) → 5 officers all with prior approvals → no OAI → 10-asset pipeline.

Two IPO tools — sections__ipo_screen + sections__ipo_risk_factors

sections__ipo_screen sections__ipo_risk_factors

For consumers: Two new IPO tools on the SEC surface (sections team substrate, wrapped for inferno's report builder). **`sections__ipo_screen`** — recent IPOs in a date window from 424B4 final-pricing prospectuses (2020-2026, ~99.9% of the IPO universe); one row per company (earliest 424B4 = the IPO itself), SPACs excluded by default, filter by `sector` (SIC prefix) or `sic_filter`. **`sections__ipo_risk_factors`** — classified risk factors from a company's IPO prospectus (424B4, S-1 fallback), mirroring `sections__risk_profile` but on the PRE-IPO filing (works for names with no 10-K yet); returns thematic + topic breakdown + the individual factors. Pair them: screen recent IPOs, then pull each one's risk profile. Perf: risk_factors is <100ms (payload path); ipo_screen is ~4-26s depending on window (per-year 424B4 scroll — narrow the window/sector to stay fast).

bio_trial_readouts flags p-value/outcome mispairs

bio_trial_readouts

For consumers: `bio_trial_readouts` now returns **`pvalue_outcome_conflict`** (bool) per row (super-bio T027). true = the stated p-value contradicts the outcome (MISSED with a significant p, or MET with non-significant) — a likely drug↔outcome mispair (e.g. Keytruda tagged MISSED with p<0.001). **Discount the outcome↔pvalue pairing on flagged rows** (~12% of rows that carry a p-value); the drug / phase / endpoint fields on those rows are still fine. Turns the "sanity-check the p-value" caveat into an explicit filterable flag. Additive.

Management track-record graph — bio_person_track_record + bio_company_people

bio_person_track_record bio_company_people

For consumers: The "people" alt-data (super-bio T010). **`bio_person_track_record`** (`name` or `owner_cik`) — a biotech exec/director's cross-company outcome record: how many companies they were an officer/director at and how those resolved (drug approval / acquired / failed / active), with the per-company detail. The proven-operator read — a CEO who ran 3 biotechs that got approved. Identity is authoritative (SEC Form-4 owner_cik, no fuzzy name matching); 16,943 people, 1,593 proven individual operators. **`bio_company_people`** (`ticker_or_cik`) — the officers & directors at a company, each annotated with their track record elsewhere ("this new CEO ran 3 approval-cos"). Complements `insiders_of_company` (SEC identity+count only) with the biotech outcome graph. Caveats: coverage = Form-4 filers 2003+ (pre-IPO founders absent); weight officers over directors (outcome is company-level, credentialed-by-association); is_entity=true = fund/VC board seat (distinct smart-money signal).

Three new bio tools — patent cliff, CRL outcomes, cash runway

bio_patent_cliff bio_crl_outcomes bio_cash_runway

For consumers: Three new biotech tools (super-bio T009/T011 + the CRL base-rate view). **`bio_patent_cliff`** — loss-of-exclusivity / patent-cliff date per approved drug (`ticker_or_cik`, `before=`, `drug=`); small molecules use REAL FDA Orange Book patent + exclusivity data, biologics use the computed BPCIA 12-year regulatory floor (`source` tells you which — never read the 12yr rule as a hard patent expiry). "What's rolling off patent in the next 3 years for X" is `before=<+3yr date>`. **`bio_crl_outcomes`** — realized post-CRL abnormal returns AND eventual-approval rate BY DEFICIENCY REASON across 419 FDA Complete Response Letters; the reason predicts both the down-move magnitude and the resolution rate (efficacy fatal: −0.43 / ~32% approved-later; inspection_483 fixable: −0.16 / 67%). Returns a `by_reason` base-rate table + individual `events`. **`bio_cash_runway`** — cash-runway trajectory in YEARS + declining-streak per biotech; read the trajectory not a point (commercial-stage names oscillate as revenue lands), and `declining_streak≥3` is the distress signal. PIT on all three.

semantic_search now hands you a cross-company screen (signal_pivot)

semantic_search

For consumers: `semantic_search` results now carry the **ARGOS classifiers each sentence fired** (`classifiers` per hit) plus a response-level **`signals_seen`** (the classifiers ranked by how many hits fired them) and a ready-to-run **`signal_pivot`** string. The pivot turns "here are example sentences about pricing power" into "here is *every company currently flagging that theme*" — it hands you an exact `screen_companies(classifier="is_pricing_pressure")` call. If you use `semantic_search` to find evidence for a theme, follow the `signal_pivot` to get the cross-company screen (the thing an LLM can't reconstruct from filings alone). Additive — existing fields unchanged.

bio_similar_setups gains event_type='crl' — the rejection NO-arm

bio_similar_setups

For consumers: `bio_similar_setups` now accepts **`event_type='crl'`** (alongside 'quarter' / 'approval') — the CRL / rejection **NO-arm**, completing the pairing with the 'approval' YES-arm. Anchors on FDA Complete Response Letters (336 points, 2015+) and returns realized SPY-adjusted abnormal returns *after the rejection*, so you can ask "what did the stock do post-CRL for companies in a setup like this at THEIR rejection?" The finding that validates the CRL-reason feature: the by-reason down-tail (median abn_252) is **efficacy −0.43 (fatal) vs inspection_483 −0.16 (fixable)** — the `reason_category` predicts the magnitude of the down-move. Caveat: ~57% of CRL points have month-only date precision (their 21/63d abn are unreliable; 126/252d + the aggregate base rate are fine). super-bio T013. Also this cycle: `mining_company_reserves` auto-picked up sedar's AEM reserves fix (AEM now serves 55.4 Moz via a `stated_total` confidence tier; 5/7 majors' reserves serve, all 7 production+AISC).

13F "current holders" tools now anchor on the latest MATURE quarter (fixes the early-quarter undercount)

thirteenf_holders_full thirteenf_portfolio_full thirteenf_smart_money_consensus thirteenf_cohort_flow

For consumers: The four postgres-backed 13F tools that answer "current holdings / consensus / cohort flow" now default to the **latest MATURE quarter** (one whose 13F filing deadline + grace has passed) instead of the raw newest quarter. Why it matters: for the ~6-week window after each quarter-end, the freshly-ended quarter is only ~half filed (early-filer subset), which massively undercounts every holder base — e.g. Carnival (CCL) showed **12 holders on the sparse Q2 vs 1,010 on the complete Q1**. These tools now anchor on the mature quarter by default (via smart_money's new `pipeline_meta.latest_mature_report_date`); pass an explicit `report_date` for a specific quarter, including the freshest. Fixes the whole early-quarter undercount in one shot. On the SEC vertical.

new tool bio_trial_readouts + bio_target_base_rate methodology correction

bio_trial_readouts bio_target_base_rate

For consumers: **New tool `bio_trial_readouts`** — clinical-trial endpoint outcomes (primary endpoint **MET / MISSED / MIXED**, the **p-value** as stated, endpoint name) mined per sponsor from SEC filings (2,442 readouts; Ph3 1,240 / Ph2 1,126). The actual trial RESULT — a tier deeper than approvals or pipeline. Filters: `phase`, `outcome`, PIT `as_of` (on filing_date). Filer-anchored (competitor trials excluded); the `sentence` field carries provenance. Caveat surfaced: it's extraction not ground truth — a p-value that disagrees with the outcome (e.g. MISSED with a significant p) is likely a drug↔outcome mispair. Pairs with `bio_pipeline` and `bio_target_base_rate`. **Plus a correction to `bio_target_base_rate`**: its methodology disclosure was inverted — it claimed Phase-3 failures were scored "advanced" (positive, inflating the rate), but advanced programs are actually CENSORED (excluded, never scored positive). The real residual bias is CENSORING (conservative). super-bio's v2 (`phase_progression_v2`) now reclassifies documented Phase-3 misses (from the trial-readout data) to negatives. Corrected the description + methodology_note; dropped the never-firing "discount if n_advanced>0" warning. On the bio vertical.

new tool — mining_company_reserves (gold-major reserves / production / AISC, reconciled + cited)

mining_company_reserves

For consumers: New tool **`mining_company_reserves`** — company-level gold **reserves (proven+probable Moz) + annual production + AISC ($/oz)** for the 7 major gold producers (NEM / GOLD-Barrick / GFI / KGC / AU / HMY / AEM), structured, reconciled, and provenance-backed. It replaces scraping 10-K reserve tables — the failure mode that produced garbage like "271 Moz reserves". Every figure is reconciled (`contained ≈ tonnes × grade`) and carries its SEC accession; **implausible values are WITHHELD (NULL), never guessed**. Reserves for NEM 134.4 / GOLD 146.4 / GFI 54.9 / KGC 19.7 Moz; production + AISC for all 7 (AU/HMY/AEM reserves withheld on a unit-scale/under-capture gate, flagged via `reserves_completeness`, but their production+AISC serve). `include_deposits=true` adds the per-deposit reconciled detail with citation. Built by the sedar team via the SK-1300 ≈ NI 43-101 play (SEC SK-1300 Technical Report Summaries + Canadian AIFs). On the SEDAR vertical.

The T24 gold-reserves feed I made the case for (SK-1300 ≈ 43-101 flipped sedar's "not our corpus") is live — sedar built it + kee provided the sec.acc access. Wrapped the two servable views (company summary + per-deposit detail). Verified matches sedar's numbers exactly; kills the scrape-misread failure inferno hit on NEM/KGC/AU/B.

bio__prior_crl_reason — leak-safe PIT gate on SEC-disclosed CRLs (disclosing_filing_date)

bio__prior_crl_reason

For consumers: `bio__prior_crl_reason` now gates SEC-disclosed CRLs on **`disclosing_filing_date`** (the date of the filing the disclosure was mined from), not `letter_date`. super-bio + futures found 78% of `source='sec_disclosure'` rows had a `letter_date` EARLIER than the filing that disclosed them — so a strict-PIT replay could surface a CRL before it was publicly known (e.g. a 10-K retrospectively narrating a years-old CRL). New gate: openFDA rows still gate on `letter_date` (public when issued); SEC-disclosure rows gate on `COALESCE(disclosing_filing_date, letter_date + 60d)`. Leak-safe (never shows a CRL before it was public), mildly conservative (a CRL first announced in an 8-K may hide until the later filing we mined). `disclosing_filing_date` now surfaced per row. Verified: at as_of=2024-01-01, 22 previously-leaking rows now correctly hidden. On the bio vertical.

sedar_resource_screen — effective_date (estimate as-of date) surfaced + filterable

sedar_resource_screen

For consumers: `sedar_resource_screen` now surfaces **`effective_date`** — the resource estimate's stated as-of date, distinct from `filing_date` — on every row, and adds **`effective_after` / `effective_before`** date filters. This is the enabling field for a deck-vs-spot staleness screen (find estimates whose as-of date is old, or restrict to fresh ones): e.g. `effective_after=2024-01-01` → recent estimates, `effective_before=2010-01-01` → stale ones. ~13% of rows carry it (withheld when the report states no explicit as-of, so filtering on it is a stated-as-of subset — a lower bound). sedar also refreshed the underlying table (v4 — fixed a silent JSON-truncation bug that was dropping the largest multi-zone grids; now 23,602 rows / 471 tickers, null-tonnes 7%→3%, nothing to re-wire there). On the SEDAR vertical.

sedar added the effective_date servable column + a full resource re-scan. Wired the column into output + two range filters. Verified live: effective_after=2024-01-01 → 320 fresh gold rows (LUN 2025-04-15); effective_before=2010-01-01 → 33 stale.

insider tools now return company_name (no more ticker→name confabulation)

insider_flow insider_signal insider_holdings insider_liquidity_calendar

For consumers: The four insider tools that resolve a ticker or CIK to an issuer now return **`company_name`** alongside `resolved_via`, so a consumer never has to guess the issuer name from the ticker. (A report builder was confabulating SRPT→"Synaptogenix", KGC→"Kohlberg Kravis Gold & Copper", CCL→"Crown Cruise Line" from `resolved_via: "atlas:TICKER"` with no name attached.) The name comes free from the Atlas ticker resolution, with a `cik→name` fallback (cik_companies) for cik-direct calls, and it's populated even on empty-result responses. Verified: cik=320193 → "Apple Inc." across all four. On the SEC vertical (dev).

From inferno's dogfood batch-1 #4 — "returning company_name on tool outputs would harden every consumer." The Atlas resolver already carried company_name; just surfaced it + added a cik_to_name() reverse (db/qdrant.py, lazy holdings import) for the cik-direct path.

mars_form_d_flow — fund_type / is_pooled_fund filters now live (PE / VC / Hedge fund flows)

mars_form_d_flow

For consumers: `mars_form_d_flow` now filters + breaks down by **`fund_type`** ('Hedge Fund' / 'Private Equity Fund' / 'Venture Capital Fund' / 'Other Investment Fund' — exact title-case) and **`is_pooled_fund`**, and each bucket carries `n_pooled_funds`. mars re-ingested the pooled-fund Form D filings that had been excluded at the source (corpus 355K → 765K rows; ~410K pooled), so PE/VC/Hedge private-fund-formation flow is now one query — e.g. `fund_type='Venture Capital Fund'` for the VC pulse. These filters were shipped disabled last week (the columns were 100% unpopulated); they're now real data. On the funding vertical (dev).

The three substrate gaps I caught smoking mars_form_d_flow last week are all fixed by mars — the big one was fund_type/is_pooled_fund being 100% NULL because pooled-fund filings were never ingested (a legacy scope choice). mars re-ran with --include-funds (+410K rows). Rewired the filters + breakdown + n_pooled_funds; verified live against mars's numbers exactly (May-2026: PE 492/$17.1B, VC 754/$9.2B, Hedge 157/$209M).

mars funding tools — deal_kind filter (clean VC series) + first_seen_at PIT-gate

mars_sector_funding_flow mars_funding_search mars_funding_history mars_investor_portfolio

For consumers: Two long-awaited additions across the mars funding tools, both wrapping columns mars shipped this week. **(1) `deal_kind` filter** on `mars_sector_funding_flow` + `mars_funding_search` — pass `deal_kind='operating_raise'` for a clean VC / company-primary-raise series that strips out bank debt, PE fund formations, PIPEs, grants (the event-mix that had inflated "funding" totals — 2026 flow drops from $4.3T to $1.23T when you isolate operating_raise). `mars_sector_funding_flow` also gains `breakdown_by='deal_kind'` to see the full mix per bucket; `mars_funding_search` surfaces `deal_kind` on every result row. The old event-mix caveat is retired. **(2) `pit_gate_on` param** ('round_date' default | 'first_seen_at') on all four tools — 'first_seen_at' gates on when a deal became *knowable in the corpus* (not its assigned round date), so it's leak-free for honest backtests. Note: first_seen_at has a coverage cliff at the recent corpus-ingest date, so 'round_date' stays the default for descriptive series. Backward-compatible (new optional params). On the funding vertical (dev).

crown asks #4 (deal_kind) + #3 (first_seen_at) — mars shipped both columns (deal_kind is 100% populated on funding_deals_v2; first_seen_at is on all four deal tables), and this is the thin-wrapper side. Verified: deal_kind='bank_debt' sorted by amount surfaces Oracle $40B / Anthropic $35B debt; the Q1-2026 mix shows bank_debt ($545B) actually > operating_raise ($447B) — crown's event-mix concern, quantified.

new tool — mars_private_credit_flow (private-credit deal-flow pulse, spreads + seniority)

mars_private_credit_flow

For consumers: New tool `mars_private_credit_flow` — aggregate private-credit deal flow as a time series, a window into a market public HY spreads can't show (94K deals). Buckets week/month/quarter/year with breakdowns by `seniority` ('first_lien_senior_secured' / 'first_lien_unitranche' / 'second_lien' / 'subordinated' / 'mezzanine'), `deal_type`, `lender`, or `benchmark_rate`, plus **median cash + PIK coupon spreads** per bucket. **BDC portfolio re-disclosures (87.5% of rows — they re-report the same held loan every quarter) are EXCLUDED by default** so flow totals aren't inflated; `include_bdc_holdings=true` for a held-book (stock) view. PIT: `pit_gate_on='announcement_date'` (default, the real deal timeline) or `'first_seen_at'` (leak-free for backtests). `principal_usd` is native USD. Note: an `industry` cut is NOT exposed — `industry_group_type` is only ~0.4% populated on this table (flagged to mars). Built for crown's private-credit-cycle-vs-HY-spread research. On the funding vertical (dev).

mars ask #1 from crown's mars-for-macro batch — mars shipped the spec + covering index (45ms, 0 heap fetches) same day; wired the thin reader over mars.private_credit_deals_v2. Smoke verified: senior-secured ~475bps vs second-lien ~700bps vs mezz higher (realistic term structure), lender breakdown surfaces Blackstone / Owl Rock / Ares, BDC-exclude drops 27K→2.5K deals. Caught industry_group_type at 0.4% coverage and dropped that cut (same discipline as form_d's fund_type) — don't ship a filter that returns near-nothing.

new tool — mars_form_d_flow (SEC Form D private-offering pulse, PIT-clean)

mars_form_d_flow

For consumers: New tool `mars_form_d_flow` — aggregate SEC Form D private-offering flow as a time series (week/month/quarter/year buckets), the **cleanest point-in-time private-capital pulse in the catalog**. Every row is a real SEC filing, so `filing_date` is a genuine PIT anchor — no extraction ambiguity, no dedup, no coverage cliff (355K filings, 2006 → today). Filters: date window + bucket + `industry` (descriptive values e.g. 'Biotechnology', 'Oil and Gas', 'Other Technology') + `jurisdiction_of_inc` (full names e.g. 'DELAWARE', 'TEXAS') + `entity_type` ('Limited Partnership' ≈ fund vehicle) + `breakdown_by` (industry / jurisdiction / entity_type) + PIT `as_of` gate on filing_date. Returns per-bucket registered-offering $, amount-sold $, median, and filing count; `total_offering_amount` (a ceiling) and `total_amount_sold` (raised as of filing) are reported separately — never conflate them. Built for crown's private-capital-cycle-vs-HY-spread research. On the funding vertical (dev). Substrate note: fund-type / pooled-fund classification isn't populated yet (use `entity_type='Limited Partnership'` as a proxy) — flagged to mars.

mars ask #2 from crown's mars-for-macro batch — mars shipped the spec + covering index (20.7ms) this morning, we wired the thin reader same-day. Smoke against live data caught three spec/substrate mismatches before ship (fund_type is 100% NULL, jurisdiction values are full uppercase names not 2-letter codes, industry values are descriptive title-case not uppercase enums) — corrected in the tool, gaps flagged back to mars. PIT as_of verified truncating the series leak-free.

insider_liquidity_calendar sort fix + insider_holdings name matching (same-day polish)

insider_liquidity_calendar insider_holdings

For consumers: Two fixes on the tools that shipped this morning. (1) `insider_liquidity_calendar` — the no-identifier date-window SCAN sorted event_date ASC + LIMIT, so a `plan_sale` scan (a trailing-180d PAST anchor) returned the STALEST slice, not current selling. Added a `sort` param (soonest_first / most_recent_first) with a smart default (plan_sale-only → most_recent_first; forward events → soonest_first) + a `truncated` flag when the limit hides more rows. A plan_sale scan now surfaces today's activity. (2) `insider_holdings` / `insider_liquidity_calendar` — `name` matching is now token-AND: `"Elon Musk"` matches the SEC-stored `"Musk Elon"` (was returning 0), and a multi-token query narrows the set (excludes "Musket David"). Thanks mcptest.

mcptest Round 13. The liquidity scan ordering was the real one — a plan_sale scan is exactly the prospect-research entry point and it was returning the least-current activity. (A separate stale-position data issue on insider_holdings — Musk's 2010 Form-3 TSLA stake not superseded — routed to insider's mat.)

mars merger tools migrated to deal-grain (merger_deals_v2) — no more over-counting, integrity-filtered

mars_mergers_screen mars_mergers_for_company mars_trending_acquirers

For consumers: The three mars merger tools now read `merger_deals_v2` (DEAL-grain — one row per deal after cluster/dedup) instead of the article-grain `mergers_resolved_v2` (one row per covering article). Fixes latent over-counting (a deal covered by N wire articles no longer counts N times). New: `integrity_tier IN ('verified','vouched')` filtered by default (drops ~730 known-garbage 'warn' extractions — trillion-dollar misparses, self-mergers, sentinel names; `include_warns=true` to see them); deal value is now `deal_value_usd_converted` (FX-normalized USD) with fallback, and undisclosed shows NULL never $0; new fields surfaced (`close_date`, `is_friendly`, `is_cross_border`, `n_articles`, `deal_value_usd_source`, `integrity_tier`). Removed the per-article `source_ref`/`snippet`/`published_at` (moved to the deal_articles_v2 bridge; `n_articles` is the provenance signal). `mars_trending_acquirers` dropped its manual DISTINCT-ON dedup (deal-grain is already unique). On the mergers vertical + dev.

mars confirmed merger_deals_v2 is canonical (funding_deals_v2's M&A sibling); migrated all 3 tools with the column map. Verified: Microsoft→Mistral $16M real / TikTok null (undisclosed, not $0); trending deal-grain Inszone 19 deals total NULL (all-undisclosed rollup), Eli Lilly 12 deals $32.6B. Mars built 3 partial integrity indexes on merger_deals_v2 for the query shapes; will EXPLAIN our plans.

new insider_holdings tool — standing beneficial-ownership positions (Form 3 + running holdings)

insider_holdings

For consumers: `insider_holdings(person or issuer)` returns CURRENT standing positions — the stake an insider actually holds, from Form 3 (initial statement) + running Form 4/5 holdings, NOT a sum of Form-4 deltas (which understates every founder who owned stock before becoming an insider). Two axes: a PERSON (owner_cik/person_id/name) → positions across all issuers (the wealth surface); an ISSUER (ticker/cik/company_id) → that company's insiders' positions. Thin reader over a daily mat (441K positions / 184K people, ~0.5ms) with the traps pre-baked: latest-snapshot dedup, "owns nothing" declarations dropped, garbage share counts bounded (`shares_bounded=false` = NULLed), funds/trusts excluded (`real_people_only` default). Derivative `shares` is the UNDERLYING count — never add it to a share balance. On the SEC vertical + dev.

Thin reader over insider.insider_holdings_mat. The "you can't get this by summing Form-4 deltas" surface.

new insider_liquidity_calendar tool — WHEN an insider becomes liquid (option expiries / vesting / plans)

insider_liquidity_calendar

For consumers: `insider_liquidity_calendar` answers "when does this person come into money" — option expiries (forced liquidity), vesting dates, and active 10b5-1 plan sales. Three shapes: a PERSON's upcoming liquidity, an ISSUER's insiders' events, or (no identifier) a forward date-window SCAN of everyone becoming liquid in the next `days_ahead` (optionally `csuite_only`) — the prospect-research entry point. Thin reader over a daily real-people-only forward-looking mat (91K events, ~2ms). `shares` is PROVENANCE not sanitized (describe an event, never rank by it); `plan_sale` event_date is a PAST anchor (currently-selling, trailing 180d) while option_expiry/vesting are future. On the SEC vertical + dev, AND the donor family (capacity+timing signal for prospect research — same table, both surfaces).

Thin reader over insider.insider_liquidity_calendar_mat. Both a trading-desk supply view and the donor/Insider-Focus "who comes into money this quarter" surface — insider flagged it for both.

new insider_signal tool — the scored insider packet for one issuer (flagship)

insider_signal

For consumers: `insider_signal(identifier, window_days=90)` returns a single scored insider-activity packet for an issuer — a `signal_class` (cluster_buy / csuite_buy / net_buying / neutral / net_selling, a DETERMINISTIC classification not a recommendation), discretionary flow (`net_discretionary_usd`, planned 10b5-1 excluded), cluster + C-suite-conviction + unusual-buyer flags, and a `notable[]` array of top insiders with per-person context (first-buy-in-5y, x-vs-own-median). All heavy preprocessing (own-history 5y baselines, cluster detection, unusual scoring) is materialized daily by the insider team; the tool reads one row (~2ms). Verified: PSUS 90d → cluster_buy, 8 buyers, 6 C-suite, net $311.6M, Ackman at 133.9× his own median. Historical `as_of` → `unsupported_as_of` (daily table is as-of-today; use insider_flow/history for backtests). On the SEC vertical + dev. 5th tool in the insider_ family.

The flagship "an LLM can't do this" tool — insider does the deterministic thinking (materialized scored table), argos reads one row, the LLM supplies judgment. Same contracts as the other 4 insider tools (value_usd only, planned excluded, plan_flag_coverage, PIT).

catalyst_calendar gains a tariff event_type (Section 232 / 301, Federal-Register-sourced)

catalyst_calendar

For consumers: `catalyst_calendar` now accepts `event_types=['tariff']` — Section 232 / 301 tariff actions across 12 commodities (copper, steel, aluminum, semiconductors, critical minerals, pharma, timber, autos, aircraft, trucks, polysilicon, drones), sourced from the Federal Register (every row carries its FR document number). History-eligible, so "what was copper's 232 status on 2025-09-01?" → `in_effect` (via `include_passed=true` + `as_of`). Detail carries commodity / statute / milestone / status + FR ref. ⚠ `rate_pct` is deliberately NULL on multi-rate proclamations (a 232 document cites market-share stats, recommendations, and multiple enacted rates in one instrument — a wrong tariff rate on an auto-consumed feed is unrecoverable; `rates_enacted[]` + the FR link are handed over instead). Copper's enacted 50% is populated.

crown's ask via catalyst_calendar. 1-line enum add + history_types routing (past-as_of reads). Verified copper golden: in_effect on 2025-09-01, FR fr:2025-14893. Commodity currently surfaces under detail.drug (generic history column reuse) — flagged catalyst for the history detail-jsonb mirror (same fix that unblocks adcom link fields), which will give proper commodity/rate keys.

new `insider` family — Form-4 insider transaction flow (4 tools)

insider_flow insider_cluster_screen insider_history insider_sector_pulse

For consumers: Four new tools over 10.2M classified SEC Form-4 insider-transaction legs (2004-2026, materialized, ~10-min fresh). `insider_flow(ticker|cik, window_days)` — per-issuer rollup separating DISCRETIONARY open-market buys/sells (the signal) from planned 10b5-1 sales / exercises / awards (noise); returns `net_discretionary_usd` (buys − sells, planned excluded) + C-suite counts + `plan_flag_coverage`. `insider_cluster_screen` — cross-issuer discovery of multiple-insiders-buying (the highest-conviction pattern). `insider_history(insider_cik)` — one insider's trail + own-history baseline (is this a first-buy-in-5y or a serial buyer?). `insider_sector_pulse` — sector-level net insider flow. All PIT-honest on `filing_date` (Form 4s land up to 2 business days late — trade-date keying is lookahead), all aggregate the price-sanitized `value_usd` (never raw filer totals, which reach $300T from typos). 10b5-1 planned-flag = SEC checkbox UNION footnote recovery (`plan_flag_coverage=false` warns pre-2023-04 windows). Dev-only, vertical-gated. Latency 1-16ms.

The payoff of the insider/donor Form-4 split — insider team owns the shared classified transaction layer (txn_classified_mat) + issuer rollups; we wrapped 4 thin readers. Verified: flow(NVDA 180d) net_discretionary -$449M (planned $178M separated); cluster 90d -> TSM 31 buyers; sector pulse precomputed 1ms; history baseline inline (PIT-safe). db/insider.py lazy pool. donor's person-wealth rollup builds on the same table next.

catalyst_calendar now returns a data_health block on every response (stale-feed safety)

catalyst_calendar

For consumers: Every `catalyst_calendar` response now carries a `data_health` block — per-event-type feed freshness + audit status, so an auto-consuming system can detect a STALE or BLIND feed and refuse to trade on it. `status` ∈ {ok, stale, critical, unknown}; `per_type` reports the freshness_class / is_stale / hours_since / row_count for each event_type returned; `audit_passing` says whether what we serve matches reality (Federal Register / Drugs@FDA recall). When degraded, a top-level `health_alert` ⚠ leads the response so an LLM consumer can't skim past it. `row_count=0` ⇒ critical (zero rows is indistinguishable from "nothing scheduled" — exactly how a silent 4-week extraction freeze hid, costing a downstream desk a -29% position). Backed by catalyst's new `catalyst_calendar_meta` table.

Safety change requested by catalyst after a silent extraction freeze. Reads the new catalyst_calendar_meta (per-type + _overall audit row); never breaks the tool (degrades to status='unknown' if the health table is unreadable).

mars_trending_acquirers no longer reports undisclosed deal values as $0

mars_trending_acquirers

For consumers: `total_value_usd` is now NULL (not 0) when none of an acquirer's deals disclosed terms — undisclosed ≠ $0. Fixed a `SUM(COALESCE(...,0))` that fabricated $0 aggregates and let all-undisclosed rollups (e.g. Inszone Insurance, 19 deals, all private) rank as if they had $0 of value. Also folds mis-stored `deal_value_usd=0` rows into undisclosed (a $0 M&A deal never exists) and adds `n_disclosed` (how many of `deal_count` carried a value). Per-target `deal_value_usd` now shows null, not 0, for undisclosed. Verified: Eli Lilly's real $26.3B / 8-disclosed preserved; insurance rollups now correctly NULL.

mcptest smoke flagged it on the mergers vertical (same "NULL means unverified, not zero" class as the sedar/funding threads). NULLIF(deal_value_usd,0) in SUM + COUNT + the per-row json. 279 substrate rows store undisclosed as 0 — flagged to mars for the source fix; NULLIF is the defensive read meanwhile.

sedar mining tools now surface us_ticker / us_exchange (US listing of Canadian issuers)

sedar_resource_screen sedar_mining_screen sedar_mining_report

For consumers: Every row from the three sedar mining tools now carries `us_ticker` + `us_exchange` — the issuer's US listing (NYSE / NASDAQ / NYSE American / OTC), resolved through the issuer (never a naive symbol match — "GOLD" does NOT wrongly resolve to Barrick). `NULL` = no US listing on file (a safe lower bound — a company may trade in the US on a line sedar doesn't hold). Big for US-based investors mapping a Canadian miner to something they can trade: Donlin Gold → NG/NYSE, Vizsla Silver → VZLA/NYSE, Fortuna → FFMGF/OTC. Also: sedar's grade quality gate now uses a row-poisoning rule (a row with ANY impossible grade emits no grades_norm at all), so grade screens are clean at source — we dropped our defensive exclusion. Each grade also carries a `verified` bool (its own contained-metal figure confirms it, ~6% — a trust signal to rank by, not a filter).

sedar shipped us_ticker/us_exchange (issuer-resolved, 69 issuers / 53 with a different symbol — crown proved naive symbol joins unsafe: 232 naive "matches", only 29 survived) + a row-poisoning grade gate (MSR Chita 681 Mt @ 6.8 g/t Au gone; Donlin back on top). Wired US listing into all 3 mining tools + dropped the defensive grade_implausible exclusion.

sedar_resource_screen gains min_grade (basis-aware) + drops the neighbouring-deposit caveat

sedar_resource_screen

For consumers: `sedar_resource_screen` now supports numeric grade filtering — `min_grade` on the selected `commodity`, backed by sedar's new `grades_norm` column (oxide→element converted, so the value IS the element grade). Unit is commodity-canonical: **g/t for precious (Au/Ag/Pt/Pd), % for base metals (Cu/Zn/Ni/Li/U)** — never compare a min_grade across commodities. **Strict basis by default** (directly-reported grades only, not AuEq/CuEq equivalents; `include_equivalents=true` to allow), and grade screens exclude `grade_implausible`-flagged rows so a spurious high grade on a huge tonnage can't top the list. `grades_norm` is now returned per row. Also: sedar's **provenance gate landed** — the neighbouring-deposit contamination (a prose sentence naming an adjacent deposit parsing as a table row) is fixed, so that residual caveat is dropped; the smaller honest caveat remains (rows with no stated contained_metal are served unverified, `quality_flags` shows which). Verified: gold ≥50 Mt AND ≥1 g/t → Donlin Gold 560 Mt @ 2.22 g/t (no PPX/MSR garbage); lithium ≥1% → Thacker Pass. Served 19,830 rows / 492 tickers.

sedar shipped grades_norm (v1.1) + a provenance fix. Wired min_grade with basis-aware strict-vs-equivalents + a grade_implausible exclusion (caught MSR 681 Mt @ 6.8 g/t Au — flagged implausible but still emitting a grade value — ranking above Donlin; excluded it tool-side and flagged the residual back to sedar). jsonb_exists() form avoids the ?-operator clashing with psycopg2 params. Dropped the fixed neighbouring-deposit caveat.

donor family grows — donor_person_resolve (identity join key) + donor_liquidity_events (Form 144 insider sales)

donor_person_resolve donor_liquidity_events

For consumers: Two more `donor`-family tools (Insider Focus prospect research). `donor_person_resolve` disambiguates a name to the right person across the 1.24M-row people graph — returns ranked candidates with CRD / CIK / title / top-employers / prominence so you pick the correct John Smith, then chain the `person_id` (mirror of `mars_company_resolve`; add `employer` to disambiguate). `donor_liquidity_events` reads SEC Form 144 planned-sale notices (115K filings, 99.99% valued) — the Capacity+Timing signal: "who filed to sell >$10M in the last 30 days", by person / issuer / geography. `individuals_only=true` filters trust/LP/holdco filers down to natural-person prospects; `seller_person_id` chains into `donor_person_profile`. Value hygiene: >$5B parse-artifacts excluded by default. Verified: NVIDIA individuals → director Mark Stevens' repeated $300-550M sales; Ernest Garcia II (Carvana) $1.58B. Dev-only, vertical-gated.

Second donor build. donor_person_resolve = the foundational join key (persons_v2); donor_liquidity_events = Form 144 (form_144_v2), the prime-prospect + perfect-timing signal for both major-gifts and wealth-advisory desks. Both buyer-agnostic, both from mars v2, zero new substrate. seller_is_entity heuristic (trust/LP/holdco vs person) is more useful for prospecting than a raw value flag. Flagged form_144 value outliers (a $31.8B Gartner parse artifact) — routing an integrity-gate ask to the mars/donor team.

new `donor` family — advisor/people-graph prospect research (donor_person_search + donor_person_profile)

donor_person_search donor_person_profile

For consumers: Two new tools open the advisor/people graph for prospect research (the `donor` family / Insider Focus). `donor_person_search` filters ~30,890 investment-advisor professionals (from SEC ADV Part-2B brochure supplements) by prior employer / school / degree / current firm / firm AUM / title / CRD / name — sorted by firm AUM descending. `donor_person_profile` is the drill-down: one person's full dated career + education + current firm + bio (chain on `person_id`, same shape as `mars_funding_search`→`mars_funding_history`). Relationship queries like "advisors at $1B+ RIAs who came from Merrill" now answerable in one call. Known limits surfaced in `warnings`: school aliases not resolved (Wharton≠UPenn — use the parent university), coverage is the ~3,295 firms that file 2B supplements not all RIAs, placeholder employers (Self Employed/Retired) filtered. Dev-only for now — vertical-gated, not on the main sales catalog.

The bios team's people graph (mars persons_v2 / bios_current_firm_v2 / person_work_history_v2 / person_education_v2) wrapped as the first two donor-family tools. donor_ (not bio_ — that's the biotech vertical) so a future mcp-donor.kscope.io can bundle these + Form-4 insiders + ADV firm-quality context. Thin readers over bios v1.1 substrate (integrity_tier pre-filtered to verified/vouched in the view; prior_employer/institution as MATERIALIZED CTEs; lower(col) LIKE never ILIKE). Verified live: prior=Merrill + firm≥$1B → 120 matches (Gary Quinzel at $122B Wealth Enhancement, ex-Merrill Lynch 2003-2016); Harvard+MBA → 135; profile drill-down returns dated career.

sedar mining tools now surface price_deck — the filed NPV price assumptions (deck-vs-spot)

sedar_mining_report sedar_mining_screen

For consumers: The economics block of both sedar mining tools now carries `price_deck` — the commodity PRICE ASSUMPTIONS the filed NPV was computed at, as jsonb `{basis: 'price with unit'}` (e.g. `{"Li":"US$24,000/tonne"}`, `{"LCE":"US$20,757/tonne"}`). 383 non-exploration rows. Keys are RAW pricing bases on purpose (Li vs LCE vs Li2CO3 are different bases at different prices — do NOT normalize; match basis to spot caller-side). This unlocks the deck-vs-spot mispricing screen (a filed $1,900 gold deck vs ~$5k spot). Only-when-stated / NULL otherwise.

crown + futures' #1 mining ask. sedar shipped the column to mining_projects_servable; we surfaced it in both tools' economics blocks (same pattern as discount_rate_pct/tax_basis). Verified: LAC {Li:US$24,000/tonne} @ $8.7B after-tax NPV; NVLH {Li2CO3:$8,560/tonne}; LAAC {LCE:US$20,757/tonne}.

macro_calendar now covers EIA weekly inventory release dates

macro_calendar

For consumers: `macro_calendar` now includes EIA weekly energy-inventory release dates — `EIA_PETROLEUM` (Weekly Petroleum Status Report, Wed 10:30 ET) and `EIA_NATGAS` (Natural Gas Storage, Thu 10:30 ET), 2010→2028. Both are filterable via `categories=['EIA_PETROLEUM']` etc. and interleave with CPI/PCE/EMPLOYMENT/GDP/FOMC. The fred/macro team generated them from EIA's pre-announced schedule with federal-holiday shifts (PIT-by-construction). Release DATES only — inventory values live elsewhere. Coverage grew 3,451→5,383 rows.

The event clock for energy term-structure studies. Verified: macro_calendar(categories=['EIA_PETROLEUM','EIA_NATGAS'], window_days=20) returns WPSR 2026-07-15 + gas storage 2026-07-16 interleaved. Data by the fred team; our side added the categories to the enum + refreshed the coverage text.

sedar_resource_screen — 10x richer data + is_headline default (clean one-row-per-zone)

sedar_resource_screen

For consumers: The NI 43-101 resource grid substrate got a full-coverage rewrite — `sedar_resource_screen` now searches ~25,312 rows / 560 tickers (was 2,620 / 393), null-tonnage down to ~7% (from 20%), with multi-zone + cutoff-grade sensitivity now captured. To keep results clean, the tool **defaults to headline rows** (`is_headline=true` — one base-case row per report×category×zone, ~14k rows) so a multi-zone deposit doesn't flood results; pass the new `include_all_cutoffs=true` to get the full cutoff-sensitivity table. `is_headline` is surfaced per row.

Same mining_resource_servable view, atomic upsert (no serving gap) — the richer data was automatic; our side added the is_headline default + include_all_cutoffs opt-in + surfaced the column. Verified: lithium M+I ≥100Mt headline → 4 rows (all is_headline); include_all_cutoffs=true → 7. grade_value/unit (numeric min_grade) still sedar's v1.1.

Mars country coverage 73% → 86.5% — country filters no longer a heavy lower bound

mars_funding_search

For consumers: Mars completed its B4 country backfill — `companies_v2.country` coverage rose from ~73% to ~86.5% (238,960 / 276,217 rows). `mars_funding_search(country='IN')` and similar geography filters now return a reasonable representation of the substrate rather than a heavy lower bound. The remaining ~13.5% NULL are mostly domain-less / genuinely ambiguous private entities. The tool's "results are a lower bound" caveat is softened accordingly.

Substrate backfill (mars-side, Path A deterministic + Path B Qwen, $0), description-only change on our side. NULL-country rows are still excluded when country is set — just far fewer of them now.

New tool sedar_resource_screen — NI 43-101 resource/reserve grids by commodity/tonnage

sedar_resource_screen

For consumers: New tool over `sedar.mining_resource_servable` (~2,620 rows / 393 tickers) — the precise per-category/zone resource + reserve GRIDS extracted from the 43-101 PDFs, the tonnage companion to `sedar_mining_screen` (which is project-level economics). Filter by `commodity` (English or symbol — gold/Au, lithium/Li, rare earth/REE), `min_tonnes` (defaults to `tonnage_basis='ore'` so lithium-brine LCE tonnes aren't compared against hard-rock ore), `category` (measured/indicated/inferred/measured_indicated/proven/probable — normalized to controlled vocab), `is_reserve`, `jurisdiction`, `ticker`. **Grain is one row per (project, category, zone)** — to get one headline row per project, filter to a single category (e.g. `measured_indicated`) or `is_reserve=true`. Grade thresholding deferred to v1.1 (grades are unit-carrying strings). On the SEDAR vertical + main catalog alongside the other sedar mining tools.

Wraps sedar's resource-grid substrate (commodity symbols + category_norm normalized at source). Verified: lithium M+I ≥100Mt ore → Thacker Pass 2328.8 Mt (LAC), Gabriel 2052 Mt (TEA), Bonnie Claire (NVLH); gold reserves → 93 rows. Commodity input maps English→symbol; category normalizes casual→vocab; scale-guarded ≤3000 Mt at source. Target consumer: Peter/Rockpoint.

catalyst_calendar — +cms_rate_notice event_type (11th type)

catalyst_calendar

For consumers: `catalyst_calendar` gains a `cms_rate_notice` event type — CMS Medicare Advantage Rate Notice milestones (Advance + Final) per public insurer (HUM/UNH/CVS/CNC/ELV/...). The Final Rate Announcement lands ~statutory first Monday of April. Snapshot-only (no history table), like `cms_coverage`/`cms_star_ratings`. Verified: `event_types=['cms_rate_notice']` returns rows for the MA insurers.

SEDAR mining NPVs now comparable — discount_rate_pct + tax_basis + a tax_basis filter

sedar_mining_screen sedar_mining_report

For consumers: Each mining project's economics now carries structured `discount_rate_pct` (numeric) and `tax_basis` (`pre_tax` / `after_tax`), populated when the report states them (~306 rows; NULL, never invented, otherwise). `sedar_mining_screen` gains a `tax_basis` filter so you can rank NPVs like-for-like — an after-tax NPV isn't comparable to a pre-tax one, and a 3%-discount NPV isn't comparable to a 10% one. Pass `tax_basis='after_tax'` and read `discount_rate_pct` per row before ranking. The raw `npv_basis` free-text string is still returned for provenance.

Closes the NPV-comparability gap flagged when the mining corpus refreshed — sedar added discount_rate_pct + tax_basis to the mining_projects_servable view (they existed in the base table but weren't exposed), and both mining tools now surface them. Verified: tax_basis='after_tax' → 171 gold rows led by Alamos $11,027M @ 5% after-tax; tax_basis='pre_tax' → a distinct set (AKE $7,145M @ 10%).

SEDAR mining corpus grew 758 → ~1,354 reports + contamination fixed at source

sedar_mining_screen sedar_mining_report

For consumers: The NI 43-101 mining corpus nearly doubled (758 → ~1,354 technical reports, ~266 with verified NPV, ~100 commodities, 2021-2026). Three source-side correctness wins now reflected in the tools: (1) the Item-23 "Adjacent Properties" contamination is fixed — CRUZ (Cruz Battery Metals) now correctly reads exploration-stage / NULL NPV instead of American Lithium's $3.26B, and SMN reads NULL reserve; (2) a served-NPV outlier cap (≤$15B) removed a $30B scale-error row — current max is Alamos Gold's ~$11B; (3) `stage` is now a controlled vocab (exploration/PEA/PFS/feasibility/construction/production/unknown) and `country` is normalized (USA→United States, Perú→Peru), so the old alias-masking is a harmless no-op. Still pending: structured `discount_rate_pct` + `tax_basis` columns exist in the base table (~97% NPV coverage) but aren't on the servable view yet — will surface once sedar exposes them.

Sedar's mining substrate refresh (re-extracted + cleaned, atomic upsert so no serving gap). Description counts + now-stale caveats updated across both mining tools. The CRUZ/SMN attribution warnings that previously told the LLM those rows were contaminated are removed — leaving them would actively mislead analysis now that the source is corrected. Verified live on dev: sedar_mining_screen(commodity='gold') → 899 projects, top Alamos Gold $11,027M (5% after-tax); sedar_mining_report(CRUZ) → exploration / NULL.

SEDAR material-change timeline full history — 9,108 → 26,119 events / 4,153 tickers

sedar_material_changes

For consumers: The Canadian Material Change Report timeline backfill landed — now 26,119 events across 4,153 tickers, full 2021-2026 (was 9,108 / 2,428, 2025-2026 only). The default `since` floor moved from 2025-01-01 to 2021-01-01 so a dormant issuer whose most recent event predates 2025 now surfaces those events instead of returning empty (active issuers still get newest-first, unchanged). Caveat: a chunk of 2024 MCRs are still absent (2024 filing_type is ~55% NULL upstream) and will fill when sedar extends the filing_type backfill to 2024.

Coverage verified live on dev: UEC 162 events spanning 2021-01-26 → 2026-06-09. Default-window broadening is a mild behavior change (wider history reachable by default) but non-breaking — newest-first ordering means active issuers see no difference at the top of the result set.

Mars funding rounds deduplicated — 4,123 duplicate-event merges

mars_funding_history mars_funding_search mars_investor_portfolio mars_sector_funding_flow

For consumers: Mars completed the B1 duplicate-event cleanup — the same funding round reported under different labels or slightly different dates (e.g. a $1B Stripe round appearing twice as "Venture" and "Equity Financing") is now collapsed. 4,123 merges total (3,513 deterministic ±7d + 610 Qwen-adjudicated ±8-30d at 99.7% agreement); `funding_deals_v2` went 80,590 → 76,471 rows (−5.1%). Prevention is now permanent (pre-INSERT range lookup + hourly sweep), so new dupes shouldn't accumulate. Verified: Stripe / ElevenLabs / CoreWeave show no same-amount/same-month collisions.

Substrate-only change (mars-side dedup), no MCP code change — surfaced here because consumers who previously saw duplicate rounds will see them collapsed.

Mars funding amounts now FX-normalized — amount_usd_converted across 4 tools

mars_funding_history mars_funding_search mars_investor_portfolio mars_sector_funding_flow

For consumers: Every funding round now carries `amount_usd_converted` (FX-normalized USD via exchangerate.host at round_date, 99.4% coverage) alongside `amount_usd` (Grok's raw extraction, which for non-USD deals may be in the original currency) and `amount_usd_source` (`grok_usd` / `fx_converted_<CCY>` / NULL for ~460 edge rows). For any cross-currency comparison — screening by amount, sorting by round size, aggregating sector flow — the tools now key on `amount_usd_converted` with fallback to `amount_usd` when converted is NULL. This fixes a subtle undercount/misrank where a large EUR/GBP round could sort below a smaller USD round because its `amount_usd` was the raw local-currency figure. Hedge-fund-analyst queries should read `amount_usd_converted`.

Mars backfilled amount_usd_converted BIGINT + amount_usd_source TEXT on funding_deals_v2 (76,619 / 77,081 rows = 99.4% coverage) with FX conversion at each round's date via exchangerate.host, cached in mars.fx_rates_v2. We wired it through all four funding tools per mars's recommendation: default every filter / sort / aggregate to COALESCE(amount_usd_converted, amount_usd), falling through to the raw value only for the ~460 pre-backfill edge rows.

Verification proof: a global amount_desc slice ranks a SoftBank EUR round (converted $95.7B, raw $87B) ABOVE Stripe ($91.5B) — the sort keys on converted, not raw. USD rows show converted == raw / source='grok_usd'; EUR rows show fx_converted_EUR with the expected ~1.1× lift. Dev only (mars family is beta-gated on prod). Commit 56591ce.

OAuth self-hosted on each vertical URL — claude.ai UI works with custom connectors

For consumers: Vertical URLs (`mcp-sedar.kscope.io`, `mcp-funding.kscope.io`, etc.) now speak OAuth end-to-end. Point claude.ai desktop / web at any vertical URL, walk through the sign-in page → magic-link email → connect — works standards-compliantly with no Anthropic connector-registry submission. Claude Code / Cursor / any bearer-first client keeps working exactly the same via `admin.bearers`. Two delivery modes for the same entitlement — Adrian's admin UI writes to `admin.users.verticals` for OAuth customers, `admin.bearers.verticals` for CLI customers. Vertical entitlement gate live on both paths (customer with `verticals=['sedar']` gets 401 on funding, and vice-versa). Server names now per-vertical for claude.ai UI dedup — `argos-mcp-sedar` / `argos-mcp-funding` instead of the shared `argos-mcp-server`.

Spike outcome after claude.ai UI refused bearer auth on vertical URLs. Option-1 path (Anthropic connector-registry submission per URL) would have been days-to-weeks of bureaucratic tax; Option-2 (self-hosted OAuth per URL with same-origin discovery) shipped in a few hours and is proven end-to-end.

Auth architecture:

Verified end-to-end via claude.ai UI on both vertical URLs. SEDAR test: 7 tools, sedar_mining_screen call for top gold projects in Ontario returned Alamos Island Gold at $11.0B NPV with the trust framing surfaced verbatim to the user. FUNDING test: 13 tools, ready for real analyst query.

For consumers on Claude Code / Cursor / other CLI-first clients: bearer path fully preserved. No config change needed.

For consumers on claude.ai UI: add each vertical as a custom connector using https://mcp-<vertical>.kscope.io/mcp/. Sign in with the email your admin added to admin.users. Magic link, click, done.

grok CLI compat — single-underscore aliases for bio__/sections__/tranche__ tools

bio_* sections_* tranche_*

For consumers: If you're calling this MCP from grok CLI / Composer 2.5 (which silently drops any tool whose qualified `<server>__<tool>` name contains `__` more than once), the 25 tools in the `bio__*` / `sections__*` / `tranche__*` families are now reachable via single-underscore aliases (`bio_*` / `sections_*` / `tranche_*`) that point at the same handlers. Existing double-underscore names still work — this is purely additive. Anthropic / Claude clients see both names in the catalog (mild bloat, ranks together in `discover_tools`); grok / Composer sees the single-underscore aliases as workable. Verify: `grok mcp doctor` should now report 85 workable tools (was 60 pre-fix). Public catalog on prod goes from 48 workable → 61 workable for grok clients.

Per company-scorer team-mail 2026-07-04. grok CLI (v0.2.82, Composer 2.5's headless runner) qualifies MCP tools as <server>__<tool> and silently drops any tool whose qualified name would contain __ more than once — e.g. argos-dev__bio__pipeline has two __s, dropped with only a stderr ERROR line the model never sees.

Fix: tools/__init__.py:grok_alias() maps <family>__<X> → <family>_<X> for the three affected families (bio / sections / tranche). server_http.py:create_app adds the alias entries to the tools dict AFTER the DEV_ONLY filter runs, pointing at the same handler instances. list_tools uses the dict key (not the tool's canonical name) so aliases surface. BETA_TOOLS + TITLES + DEV_ONLY_FAMILIES all mirror the aliases so beta-gating, titles, and dev-only-family filtering all stay consistent regardless of which name the caller uses.

bio__prior_crl_reason — corpus expanded 88→220 companies; new `source` + `drug` fields; INCY coverage-gap closed

bio__prior_crl_reason

For consumers: super-bio expanded the CRL substrate 439→934 rows / 88→220 distinct companies with company self-disclosed CRLs from SEC filings (Tier-1 coverage-breadth work; ~132 net-new companies). New per-row fields — `source` (`openfda` = authoritative FDA letters | `sec_disclosure` = company self-report; filer_cik match, confidence=1.0) and `drug` (populated on sec_disclosure rows). **The INCY residual blow-up from futures' battery #3 is closed** — INCY now returns 3 CRLs including the 2023 ruxolitinib XR pattern that was the exact cream-vs-XR discriminator. `application_number` + `approval_status` are NULL on sec_disclosure rows by design (SEC filings don't reliably state them); `same_application_as_current` + `eventual_resolution` degrade honestly to null/'unknown'. Also fixed: `same_molecule_diff_formulation` now extracts the molecule root (INN) so callers passing `'ruxolitinib cream'` correctly hit XR-tablet rows — was previously a whole-string substring check that missed cross-formulation matches.

Per super-bio team-mail 2026-07-03. Corpus expansion (openFDA feed + SEC-disclosure Tier-1) closes the coverage gap flagged as the binding constraint in futures' battery #3 debrief. Ship notes:

catalyst_calendar — +foreign_permit event_type (10th type; CNSC nuclear MVP)

catalyst_calendar

For consumers: New 10th `event_type`: `foreign_permit` — foreign government-approval calendar. MVP scope is CNSC nuclear licensing hearings (Canada), covering uranium/nuclear developers with a US ticker (NXE Rook I, DNN Wheeler River). `detail` carries `{regulator, sub_type, us_ticker, is_developer, applicant, matter, decision_class, estimated_decision, outcome}`. Snapshot-only routing (same as `cms_coverage` / `cms_star_ratings`). Uranium/energy angle — new sector for the calendar; broader jurisdictions may follow as catalyst_calendar expands.

Per catalyst_calendar team-mail 2026-07-03. Same three-line pattern as cms_star_ratings: _EVENT_TYPES set + inputSchema enum + description. Snapshot-only routing — does NOT enter history_types.

thirteenf_holders PIT route — v3 accumulation signals + qoq_action now populated

thirteenf_holders thirteenf_holders_full

For consumers: PIT queries (passing `report_date=`) now carry the same v3 Layer-3 signals the Qdrant latest-quarter path had — top-level `consecutive_accum_qtrs`, `consecutive_dist_qtrs`, `accel_4q`, `avg_position_value`, `is_cluster_entry`, `is_cluster_exit`, `is_quiet_accumulation`, `corporate_action_suspect`. Each holder in the list also carries `qoq_action` ∈ {`new`, `exit`, `add`, `trim`, `unchanged`} mapped from `holdings_events`. Fixes the PIT-replay gap that was degrading backtest fidelity — replays match live workups now. `exit` doesn't appear on this response path by construction (exited holders aren't in the current-quarter holdings), and `qoq_action` may be null for holders with no state change recorded — LEFT JOIN semantic, entry stays sound.

Per smart_money team-mail 2026-07-03. Two LEFT JOINs added on the PIT (report_date-specified) code path — same tables smart_money populates for every historical (cusip, quarter) already; pure MCP-side query wiring, no compute change on their side.

For multi-cusip tickers, top-level Layer-3 signals pick the row with the highest avg_position_value (single-cusip tickers are a no-op). For multi-cusip holders where the same cik has divergent events under different cusips, qoq_action picks the "loudest" event per materiality precedence (exit > new > decrease > increase > unchanged) so a holder doesn't disagree with themselves in the response.

WEX/2025-09-30 verify case from smart_money's spec smoked live: consecutive_accum_qtrs=1, is_cluster_entry=true, is_cluster_exit=true, corporate_action_suspect=false; qoq_action populates on the holders list (trim / add / new / unchanged; exit is structurally absent — it would require the holder to NOT appear on the response path).

Caveat still applies from smart_money 2026-06-12: late RESTATEMENT amendments can retroactively change historical answers under as_of. Rare (~3.4K amendments corpus-wide, mostly old).

catalyst_calendar — +cms_star_ratings event_type (9th type)

catalyst_calendar

For consumers: New 9th event_type: `cms_star_ratings` — annual CMS Medicare Advantage Star Ratings release (forward ~October), per public insurer (UNH/HUM/CVS/ELV/CNC/MOH/ALHC/CLOV). Detail carries overall stars + YoY + 4.0-cliff proximity. Snapshot-only (no history table — same routing as `cms_coverage`/`debt_maturity`/`clinical_readout`).

Per catalyst_calendar team-mail 2026-06-28 (second batch — bundled with their data ship per "our learning" from yesterday's enum-lag). Data already live on proc2.catalyst_calendar_snapshot (8 rows). One-line-tier fix mirroring the earlier device + cms_coverage batch.

catalyst_calendar — 3 new event_types: device_panel, device_pma, cms_coverage

catalyst_calendar

For consumers: Three new event_types accepted by `catalyst_calendar`: `device_panel` (FDA device advisory committees), `device_pma` (device PMA approvals + projected forward decisions), `cms_coverage` (Medicare coverage events — granted/denied/restricted/delayed/pending). Same common record shape, `detail` jsonb passes through per type. PIT routing: device_panel + device_pma read snapshot OR history (same as pdufa+adcom); cms_coverage is snapshot-only (same as debt_maturity + clinical_readout — no history table for CMS).

Per catalyst_calendar team-mail 2026-06-28. Data was already live on proc2.catalyst_calendar_snapshot (device data also in catalyst_calendar_history); the gap was purely the tool's input-schema enum rejecting the new strings at input validation before the query ran. Three-line fix:

Description block now lists eight event types (was five) with per-type data-layer notes.

Unblocks futures' device-PMA triage (WALD/NPCE/AORT) + the device replay battery + the CMS coverage screen.

bio__prior_crl_reason — FDA Complete Response Letters with deficiency reason (PIT)

bio__prior_crl_reason

For consumers: New no-arm complement to `bio__similar_setups(event_type='approval')`. Surfaces prior CRLs for a sponsor with the deficiency reason as the load-bearing field (CMC/manufacturing often fixable, efficacy often fatal). PIT-strict via required `as_of=`. Pass `application` for same-app-repeat flag, `drug` for same-molecule-diff-formulation flag. 439-letter corpus skews to generics/specialty pharma — when a sponsor returns 0 rows the response carries an explicit `coverage_caveat` so you don't infer absence-as-evidence. Tool emits ingredients (reason + verbatim evidence sentence + classify_confidence), never verdicts. Dev-only family. Output schema is open — futures owns the contract; ping back if a field should be added.

Sibling of bio__warning_letters / bio__inspections / bio__fda_approvals. Closes the YES-arm-only gap noted in bio__similar_setups(event_type='approval') (2026-06-19) — that tool returns realized post-approval returns conditional on approval; this one surfaces the prior-CRL pattern that often preceded a no-arm outcome.

Per super-bio's team-mail spec 2026-06-27. Wraps proc2.super_bio.fda_crl_letters (439 letters 2002-2026 from FDA's openFDA CRL transparency feed; deficiency reason classified once-at-ingest by Haiku over OCR'd letter text; 147 matched to a super_bio sponsor).

PIT semantics: letter_date < as_of strictly. eventual_resolution is PIT-gated via LEFT JOIN against fda_approvals.submission_status_date < as_of so the current event's own future approval never leaks. Caller can pass application / drug to get computed same_application_as_current and same_molecule_diff_formulation flags per row.

Surfaces an explicit coverage_caveat when zero rows return for a sponsor — absence != no prior CRL given the corpus skew. Tool stays in "ingredients not verdicts" mode: reason category + verbatim deficiency sentence + classify_confidence per row.

Spec-verified case from super-bio: REGN as_of=2026-01-01 → 5 CRLs incl. BLA 761303 odronextamab same-molecule repeat (2024-03-22 clinical_design → 2025-07-30 inspection_483).

Durable per-call audit log — JSONL to /var/log, daily rotation

For consumers: Every tool call you make is now logged to a durable JSONL file separate from the systemd journal. Schema: ts / ts_iso / env / email / kind / tool / args (preserved as JSON object) / ok / error / latency_ms. Files keep forever (no auto-delete); daily rotation at UTC midnight. If your persona makes high-volume calls and you want a query path beyond `jq`, ping kee — we can add a /audit?email=...&since=... admin endpoint. Tool args are written verbatim with no PII redaction; flag this if you put sensitive query text in args.

Compliance-grade audit logging on both servers. Files:

Each line is one tool call: timestamp, email (or *break-glass* for static bearer), tool name, full args, success/failure, error class, latency_ms. Three audit points per call: gate denial, unknown-tool, success-or-exception with real latency.

Query patterns (jq-based):

# Every call by a specific user today
jq -c 'select(.email == "user@example.com")' < argos-mcp-prod.jsonl
# Failed calls only
jq -c 'select(.ok == false)' < argos-mcp-prod.jsonl
# Calls that mention VRTX anywhere in args
jq -c 'select(.args.ticker=="VRTX" or (.args.tickers // [] | index("VRTX")))' < argos-mcp-prod.jsonl

Failure mode: audit-write failures emit a journal warning and drop the audit line — never crash a tool call. Audit must not be load-bearing on the request path.

Caveat: tool args are written verbatim. No PII redaction. Fine for internal compliance; needs a redaction pass before external audit sharing.

bio__similar_setups — event-aware mode (approval-anchored comparables)

bio__similar_setups

For consumers: New optional `event_type` kwarg. Pass `event_type='approval'` to anchor on a company's FDA approval setup and get neighbors who were in similar setups at THEIR approvals, each with realized SPY-adjusted abnormal returns at 21/63/126/252 trading days. Branch on `forward_outcome.type` ('quarter' vs 'approval') — they carry different shapes. CRITICAL: approval-anchored outcomes are YES-arm only (no CRLs yet — catalyst's CRL registry in flight). Real base rate for small/mid biotech post-approval is NEGATIVE at 1yr (median ≈ −0.24, ~28% up — sell-the-news + derate). Default behavior (omit event_type) unchanged.

Per super-bio's NOTE_FROM_SUPERBIO_SIMILAR_SETUPS_EVENTAWARE_2026-06-19. bio_snapshots collection now carries two kinds of points distinguished by the event_type payload field: 'quarter' (~46.7K, existing) and 'approval' (NEW, 253 brand_exact@conf=1.0 own-drug approvals 2015-2025).

Wired via optional event_type kwarg. Applies the filter to BOTH the anchor scroll (so we pick the right kind of snapshot for the query — e.g. NBIX's CRENESSITY 2024-12-13 approval point, not their latest quarter) AND the k-NN query.

forward_outcome shape differs by type. Quarter: {type:'quarter', event_date, days_after_snapshot, excess_vs_xbi, up}. Approval: {type:'approval', anchor:'fda_action_date', event_date, brand, abn_21, abn_63, abn_126, abn_252, up} — null abn_* for horizons not yet realized on recent approvals.

Response surfaces event_type_filter (the value passed) + my_event_type (actual anchor's event_type). Note text refreshed to spell out the YES-arm caveat + the negative base-rate finding (super-bio's own 06-19 CRL study).

Atlas — REITs added to atlas_tickers + has_corporate_filings filter

resolve_company_name get_fundamentals screen_companies search_activist_positions

For consumers: AMT, SPG, O, PLD, BMO (REITs) now resolve via Atlas ticker→CIK lookup. New Qdrant payload field `has_corporate_filings` (boolean) — filter to True if you want to exclude 13F-only filers like OZK / PFBC that show up under SEC `company_tickers.json` but only file 13F-HR / SC 13G. PFBC was also reclassified `preferred` → `stock`.

Atlas's atlas_tickers Qdrant collection now includes REITs (entity_type='reit') — 9,664 → 9,983 points. Sample: AMT 1053507 reit / O / PLD / SPG.

New payload field has_corporate_filings (boolean): true iff the CIK has at least one 10-K/Q / 8-K / 20-F / 40-F / S-1/3/4/11 (or /A) filing in the SEC accessions index. OZK and PFBC are false (their SEC-mapped CIKs only file 13F-HR / SC 13G via alternate CIK chains).

Transient bug heads-up: ~523 tickers were wrongly flagged dead overnight June 18→19 (Form 25-NSE per-security redemption miscount); fixed atlas-side ~13:30 UTC June 19. If you ran resolution between ~00:30 and ~13:00 UTC June 19 and saw missing AMT/O/PLD/BMO, it's resolved now.

Atlas — 2,189 mega-caps unstuck from "dead_with_data"

screen_companies resolve_company_name get_company_profile search_activist_positions

For consumers: If you were filtering by `status='active'` and missing AAPL / MSFT / NFLX / SCHW / ABBV / AAL (and 2,180 others) since April, they're back. 14 ticker-reuse renames also corrected — BK now resolves to BNY Mellon CIK 1390777 (not the old BANK OF NEW YORK CO). NFLX and SCHW had been miscategorized entity_type=etf, now stock.

A reconciler had a hardcoded date still_trading = end_date >= "2026-03-01" that flipped 2,189 ticker'd CIKs to still_trading=False on 2026-04-07. An overprotective guard in update.py then prevented resurrection. Combined: AAPL, MSFT, etc. stuck in "dead" state for two months. Fixed at the source (rolling 30-day cutoff + resurrection-allowed). No MCP-side code change required.

catalyst_calendar — history now mirrors snapshot exactly; evidence + date_range + adcom + projected surface

catalyst_calendar

For consumers: PIT-replay queries through history now carry real `evidence` text (governing source narrative) and proper `date_range` bounds — previously these were always NULL on history reads. New `event_type='adcom'` for filtering FDA AdCom meeting dates separately from pdufa. New `status='projected'` for v1.5 calibrated PDUFA windows. History rebuilt through `build_calendar` so phantom guards (CORT relacorilant) and cross-CIK collapses (SNSE→CELC) now auto-inherit into replay.

Per catalyst team's NOTE_FOR_MCP_HISTORY_MIRRORS_BUILD_2026-06-18. catalyst_calendar_history rebuilt through build_calendar, auto-inheriting every build-layer guard. Three additive columns on history table now wired into our reader: date_range_start, date_range_end, evidence. Two new enum values: event_type='adcom' and status='projected'.

Spec-verified: event_types=['adcom'] as_of=2024-06-30 → 5 rows with real evidence text (PRGO Advisory Committee on Nonprescription Drugs, BCLI NurOwn ALS, VANI September 2023 PADAC). Forward pdufa regression unchanged.

bio__inspections — FDA inspection classifications (NAI/VAI/OAI)

bio__inspections

For consumers: The EARLY FDA enforcement signal — most pharma enforcement appears here first, often months before a Warning Letter. OAI = the bad one (regulatory/legal action warranted); VAI = voluntary; NAI = clean. Pair with bio__warning_letters for the escalation. Canonical pairing: Catalent OAI 2025-11-24 sits 4 days BEFORE the CTLT Warning Letter 2025-11-20 — the inspection IS what the Warning Letter later escalated. Use worst_only=true as the convenience for OAI-only screening.

Sibling of bio__warning_letters AND its upstream early-warning signal. 26,133 classified inspections since 2015 (OAI 453 / VAI 9,313 / NAI 16,367), 50% matched to super_bio sponsors. Unmatched rows are foreign API makers, CDMOs, and BIMO clinical investigators — reach them via company ILIKE.

Filters: ticker_or_cik / company / classification ('OAI'|'VAI'|'NAI') / worst_only (convenience for OAI) / project_area / since / min_confidence.

Spec-verified: company=Catalent, classification=OAI → Catalent Indiana LLC 2025-11-24 OAI Drug Quality Assurance (4 days before the CTLT Warning Letter); ticker_or_cik=CTLT → 65 inspections across plants (VAI 32 / OAI 1 / NAI 17 in top-50); project_area='Human Cellular, Tissue, and Gene Therapies' → CBER cell/gene-therapy site inspections.

Dev-only family.

bio__warning_letters — FDA Warning Letters by recipient

bio__warning_letters

For consumers: If your persona deals with biotech sponsors or CDMOs, you now have a Warning Letters surface. Use `company=` (ILIKE) for CDMOs like Catalent that aren't in super_bio's sponsor universe; use `ticker_or_cik=` for matched super_bio sponsors. Empty results don't mean "no FDA issues" — inspection classifications (NAI/VAI/OAI) aren't here yet.

Sibling of bio__fda_approvals. Late-stage public FDA enforcement actions matched to super_bio sponsors (matched_confidence 0.85 for name_token, 0.60 for name_overlap). Filter by issuing_office='Office of Manufacturing Quality' for the supply-chain catalyst cohort.

Spec-verified: company=Catalent → Catalent Indiana LLC 2025-11-20 OMQ CGMP/Finished Pharmaceuticals/Adulterated; issuing_office=OMQ since=2024-01-01 → 3 letters incl. Seaway Pharma + Zhejiang Huahai.

Dev-only family — visible on mcp.staticpipe.com, not registered on prod.

Two MCP servers — dev / prod split

For consumers: Engineering iterates on mcp.staticpipe.com; sales-managed mcp.kscope.io is the customer-facing surface. Bio + sec_power + mars tools live only on the dev URL. Update your config to point at dev if you need those; prod's catalog is sales-curated and stable.

Two MCP servers now: mcp.staticpipe.com (dev, engineering's iteration target) and mcp.kscope.io (sales-managed, customer-facing). Active development on bio + sec_power + mars happens on dev and graduates to prod only when sales-approved. Branch isolation between dev and master plus an ARGOS_ENV runtime gate makes accidental prod leaks hard.

bio__fda_approvals — modality filter + CBER cohort + source field

bio__fda_approvals

For consumers: If you were grouping FDA approvals by drug class via brand-name string matching, use the new `modality` param instead (CAR-T / gene_therapy / cell_therapy / mAb / small_molecule / etc.). `source` field on every row disambiguates CDER vs CBER origin. Class-only queries (modality without ticker_or_cik) auto-lower min_confidence to 0.0 so CBER rows with unmatched ticker still show up.

bio__fda_approvals adds a modality filter (10-class enum) and surfaces source (CDER vs CBER) on every row. CBER cohort (25 cell/gene/CAR-T approvals) added — KYMRIAH through AUCATZYL across CAR-T, LUXTURNA through BEQVEZ across gene_therapy, STRATAGRAFT through RYONCIL across cell_therapy.

Per-user tool entitlements via proc2.admin.users.tool_access

For consumers: Customer access to gated tool families (bio, mars, sec_power, sedar, news, adv) is now per-user via the admin UI. `tool_access=[]` means public catalog only (always-on), NOT "no tools." Adds on top of public.

Adrian's admin UI can now grant per-customer access to gated tool families via proc2.admin.users.tool_access (jsonb array of patterns like ["bio__*", "sedar_*"]). Additive — never replaces public.

News + ADV moved out of the public catalog BREAKING

search_news search_news_articles adv_manager_profile adv_manager_raum_trend adv_screen_managers adv_manager_fees adv_raum_decliners adv_manager_disciplinary_detail

For consumers: 8 tools moved from public to premium-gated. If your persona uses `search_news*` or `adv_*` on the prod URL and the customer doesn't have an explicit grant, those tools no longer appear in list_tools. Personas using the dev URL aren't affected (full catalog there).

The 2 news tools + 6 ADV tools moved out of the public catalog. Grant via search_news* / adv_* patterns through Adrian's admin UI.

macro_calendar + macro_series

macro_calendar macro_series

For consumers: FRED release calendar + ALFRED PIT vintages. 44 series with full first-print revision history. UNRATE 2020-04-01 returns 14.7 at as_of=2020-05-15 (first print) vs 14.8 today (revised). Pair with catalyst_calendar for the macro+catalyst calendar story.

FRED release calendar + ALFRED PIT vintages — 44 series, full revision history. UNRATE 2020-04-01 returns 14.7 at as_of=2020-05-15 (first print) vs 14.8 today (revised).

catalyst_calendar — PDUFA / IPO lockup / debt maturity / clinical readout

catalyst_calendar

For consumers: Unified dated-catalyst record across 4 event types. PDUFA confidence is calibrated vs Drugs@FDA at conf>=0.9 (medAE 5 days). PIT routing handled internally — pass `as_of` for past replays, get history-table reconstruction; current/forward queries hit the snapshot.

PDUFA / IPO lockup / debt maturity / clinical readout in one uniform dated-catalyst record. Calibrated PDUFA confidence vs Drugs@FDA (medAE 5 days). PIT routing: past as_of → history table; current / forward → snapshot.

How to connect

This is the dev MCP server. Engineering's iteration target; actively-developed tool families (bio, mars, sec_power) live here. Sales-managed mcp.kscope.io is the customer-facing surface.

Connect

{
  "mcpServers": {
    "argos-dev": {
      "type": "http",
      "url": "https://mcp.staticpipe.com/mcp/",
      "headers": { "Authorization": "Bearer <DEV_TOKEN>" }
    }
  }
}

Dev token lives in /etc/argos-mcp-dev/env. Ask kee for the current value. Static bearer only — no OAuth, no per-user grants.

Tool catalog

Auto-generated from the registered tool dict at server startup. Restart to refresh.

Company signals + fundamentals (5)

Echo / distress / classifier / sub-section signals; XBRL fundamentals.

Institutional ownership (13F) (12)

Top holders, manager portfolios, cohort flow, manager performance.

Activist positions (13D/13G) (1)

Beneficial-ownership stakes + LLM-classified intent.

Filing sub-sections (10-K/10-Q/20-F/S-1) (5)

Sub-section semantic search, peer risk-factor overlap, YoY filing diff.

Adviser quality (Form ADV) (9)

RAUM trends, fees, disciplinary detail, forced-seller watchlist.

Corporate-bond panel (N-PORT-P) (10)

Fund-held debt — maturity walls, refi pressure, dispersion.

Earnings + catalysts + macro (5)

Projected reports, Foster SUE, PDUFA / lockup / debt-maturity / clinical readouts, FRED + ALFRED.

Filing-header identity (4)

Rich SEC-header card, address history, insiders, name → CIK.

Canadian filings (SEDAR) (4)

Mining / cannabis / TSX-listed semantic search, bilingual.

Biotech pipeline + FDA [dev-only] (17)

Survivorship-complete drug pipeline, FDA NDA/BLA history.

Private-co funding + M&A (Mars) [dev-only] (13)

Private rounds, M&A timelines, trending acquirers.

Power sector (utility / IPP) [dev-only] (8)

Forward supply curve, demand wedge, outages, rate cases.

Meta (1)

Semantic catalog search.

Other (84)