Release notes

presentation-contract is a house rendering standard every guided workflow now loads. It sets three rules for how an answer is laid out: lead with the answer rather than burying it in a paragraph; put any set of three or more parallel figures — firm attributes, cohort percentiles, a ranked shortlist, filing-history changes — in a table rather than a comma-run of a sentence; and collect the caveats (exactly as filed, definitional gaps, the coverage a screen did not walk, not an endorsement) into one block at the end. It governs layout only, and never changes what a tool reported or how it is filed.

It is wired as a dependency into all six guided workflows — Firm Due-Diligence Brief, RIA Acquisition Screen, Peer Benchmark, Custody & Disclosure Triage, Filing Change Monitor, and Verify a Claim — on top of each skill's own output_contract sections. Skills are fetched at call time, so this reached every already-connected client immediately: no reconnect, no tool-surface change.

verify-a-claim takes a public claim about an adviser — a website line, a pitch deck, a recruiter email — and puts it beside what the firm actually filed. It returns one of four verdicts: supported by the filing, differs from the filing, not on the form, or depends on definitions. Dates are attached to every comparison, and definitional gaps are named before any gap is reported.

The skill reports what the filing says and stops. It never characterises a claim as false or misleading, and a gap is a question for the firm rather than a finding.

This is the other half of the design v1.02 was about. Tools are cached at connection time, so adding server_info forced a one-time reconnect. Skills are fetched at call time, so a new guided workflow reaches every already-connected client immediately — no refresh, no reinstall.

MCP clients read the tool list once, at connection time, and most never re-read it. A connection made on v1.0 still sees only the original four tools and none of the guided skills. A new server_info tool reports this server's live version, tool count, and tool names so an assistant can detect the drift itself and say so.

• Cursor: Settings → MCP → toggle ADVmcp off and on, or hit refresh on the connection.
• Claude: quit and reopen the app, or disconnect and re-add the connector.
• Other clients: reconnect the server so it re-runs the handshake and re-reads tools/list.

The endpoint is unchanged — https://advmcp.ai/mcp, no account or key. Skills need no refresh at all: they are fetched at call time, so list_skills and get_skill always serve the latest versions to an already-connected client.

• Tool errors now return a short generic message instead of raw database detail, with the real error logged server-side — nothing about the database's internals leaks to a connected client.
• Every tool's output schema is now explicit: responses describe exactly the filed columns they return (the cohort tools' statistics documents stay flexible by design, since their measure keys vary with the data). Nothing about the returned data changed — it is now precisely described, so clients can validate what they receive.
• server_info joined the tool list (tool #9), reporting the server's live version, tool list, and skills catalog state so an assistant can detect a stale client cache on its own.
• Every tool call now writes an audit record — tool, timing, outcome, and a parameter fingerprint (never the parameters themselves) — giving the server an examiner-grade usage trail.

The MCP surface grew from four tools to seven. Two of the additions serve the skills catalog below; the third makes cohort math a single call.

• list_skills and get_skill — discover the guided workflows this server ships, and load one's full instructions at the moment it is needed. Nothing is installed and nothing is copied; an updated skill reaches every connected assistant on the next call.
• cohort_stats — percentile benchmarking used to require walking a cohort page by page: a 193-firm band took four calls, a national one dozens. This tool computes the distribution server-side in one call — define a cohort by state, AUM band, and retail flag, and it returns the cohort size, summed AUM, custody and disclosure base rates, and p10/p25/p50/p75/p90 for eleven filed or derived measures: AUM, clients, accounts, employees, firm age, AUM per client, per account, and per employee, accounts per client, clients per employee, and discretionary share. Only approved registrants with reported AUM are counted, so exempt reporting advisers never distort a ratio.

The existing tools got richer payloads. search_firms and top_firms now return discretionary AUM alongside total AUM, the custody and disclosure flags, and each firm's registration date — so screening that previously needed a get_firm call per candidate can be done from the search results directly, and a peer list brings firm ages with it.

Skills are written procedures for Form ADV work — what to pull, in what order, and what not to say. A connected assistant reads them from the server via the two catalog tools above:

• Form ADV Literacy — What every field on Form ADV actually means, and which conclusions the filing does and does not support.
• Presentation Contract — The house rules for rendering any Form ADV output — lead with the answer, tabulate parallel figures, keep positions and caveats honest, and collect caveats at the end. Layout only; it never changes a filed figure.
• Firm Due-Diligence Brief — A one-page read on any registered adviser: business shape, fee model, custody, disclosures, and the questions the filing cannot answer.
• RIA Acquisition Screen — Turn an acquisition thesis into a ranked, evidence-backed shortlist of registered advisers — and an honest list of what the ADV cannot tell you about them.
• Peer Benchmark — Place one adviser against a defined cohort as percentiles rather than adjectives — eleven distribution measures and two base rates in a single cohort_stats call, composition against named peers, and five years of filed trajectory (April 2021 – August 2026).
• Custody & Disclosure Triage — Run a roster of advisers against Item 9 custody and Item 11 disclosure answers, and rank what actually needs a human to look at it.
• Filing Change Monitor — Watch a list of advisers and report what actually changed between filings — separating real events from restatements, routine lifecycle, and noise. Filing history runs April 2021 to August 2026: 33 dated comparison points, month-over-month for most of it and daily since 12 August 2026.
• Verify a Claim — Check a public claim about an adviser against what the firm actually filed — supported, differs, not on the form, or a definitional gap, with dates attached and no accusations.

Every skill carries explicit guardrails, published alongside it on the home page. They report figures exactly as filed, never estimate a number a firm did not file, and never characterise a disciplinary disclosure beyond what the filing literally says.

Two skills were rebuilt during this release to use the new tooling. Peer Benchmark (v2.1.0) runs its distribution tier in a single cohort_stats call, gained firm age as a measure, made national cohorts practical, and — with real percentile points to place against — now reports positions as brackets ("between p50 and p75") rather than implying single-digit precision the data does not support. Form ADV Literacy (v1.1.0) picked up the new tool in its reference.

The change log was a fortnight deep at launch. It now spans April 2021 to August 2026 — 33 dated comparison points across 566,393 dated firm records, holding 163,317 filed changes on 27,881 advisers.

Coverage is not continuous, and the record says so rather than implying otherwise:

• April 2021 to October 2022, month over month.
• A single bridging comparison spanning October 2022 to May 2025.
• A single bridging comparison spanning May 2025 to December 2025.
• December 2025 to August 2026, month over month.
• Daily from 12 August 2026.

A change recorded on one of the two bridging comparisons is real — the filed values genuinely differ between those dates — but the timing within the window is unknown. Every change row carries both the date it was recorded and the date it was compared against, so the window is always visible.

The deeper record makes the annual filing cycle measurable. Advisers must file an annual updating amendment within 90 days of fiscal year end, so calendar-year firms cluster in Q1. April 2026 recorded 12,171 AUM changes; April 2022, four years earlier, recorded 11,224, against 565 to 3,864 total changes in every other month of that year — roughly thirty times a normal month, in two independent years. A restatement filed in April means something different from the same restatement filed in July, and the Filing Change Monitor now reads it that way.

Two further patterns emerged at this depth. Of 1,341 registration status changes across the whole record, 1,339 are a newly registered adviser's 120-day provisional status maturing into full approval — routine lifecycle rather than news. Disclosure flips are the rare, high-signal type: 1,237 in five years, against 47,690 AUM changes, and roughly two in five are removals rather than additions.

Roughly 49 monthly rosters since 2020 remain uningested. The SEC published those months as XLSX workbooks whose sheet data runs to about 137 MB uncompressed, which the ingest worker cannot parse inside a single request; a two-phase reader is the planned fix. Two further months are SEC placeholders containing no data.

Several defects were found and corrected:

• firm_changes returned an error on every call. It sorted on a column that does not exist in the table; the tool had never worked. Fixed.
• Unscoped name searches in search_firms could exceed the query timeout. A stored search column with a trigram index brings a representative query from roughly 1,650 ms to 2.4 ms.
• The website field was wrong for 6,983 firms. Form ADV Item 1.I lists every web address a firm controls — staff social profiles, podcast pages, regional messaging accounts — and the first entry was being taken. The initial fix ranked the candidates instead, but its name test was inverted: it asked whether the domain contained the firm’s entire legal name, which is almost never true, so that test never passed and shortest-domain-wins became the effective rule. It left Vanguard on a Reddit page, BlackRock on Douyin, and AllianceBernstein on a messaging-app link. Name matching now runs on significant tokens of the firm name, ignoring industry words such as fund, advisors, and management, and a bare domain is preferred over a deep link. A further 1,793 firms were corrected, including many of the largest advisers in the country.
• get_firm exposed an internal database column in its response. Removed.
• Change detection recorded roughly 13,800 changes that never happened — about 35% of the log — where a field was simply being observed for the first time. Those rows were deleted and the comparison logic corrected.
• The daily refresh was not completing. The pipeline ingests the feed in chunks and expects to be re-invoked until finished, but was being called only once a day, so no vintage had finished since 17 August and most of the directory was serving figures from that date. It now resumes every ten minutes and finishes an in-flight vintage before starting a newer one.
• One change comparison spanned 38 months, compressing three years of amendments into a single dated row and attributing all of it to December 2025. It has been split into two properly bounded windows, and every change row now carries the date it was compared against.
• Duplicate change rows were removed and a uniqueness constraint added. An index on snapshot date brought the change computation back inside its timeout, after repeated runs had been timing out and stacking concurrent copies.

One limitation remains on the website field: about 4,200 firms resolve to a domain that does not contain their own name. Most of those are correct — a firm branded differently from its legal name — but some are not, and 285 firms list no address other than a social profile anywhere in their filing.

The home page now lists the skills catalog, reading live from the server so it cannot drift from what is actually published. The database metrics are live rather than hardcoded. The aggregate AUM figure is relabelled “Sum of reported RAUM” and carries a note that advisers who sub-advise for one another report the same assets, so it is not the size of the industry.

The remaining monthly rosters. 226 of the SEC's monthly adviser files have been catalogued, reaching back to June 2006; 29 are ingested. Completing the roughly 49 months since 2020 needs the two-phase XLSX reader described above, and would deepen the record from five years to twenty.

A cross-firm change feed. The current change tool answers questions about one adviser at a time, so “what moved across the market this week” can be asked of a watchlist but not yet of the whole dataset.

The site is built directly on the SEC Investment Adviser Public Disclosure (IAPD) compilation feed. The current production vintage contains 23,740 firm records, of which 16,865 carry an APPROVED registration status and 10,552 are retail-facing.

Ingestion is fully automated: a streaming parser downloads the daily SEC feed, normalizes each filing, and writes it to a versioned schema with support for diffing and historical snapshots.

• firms — the latest view of every registered investment adviser.
• firm_snapshots — dated point-in-time copies used for change tracking.
• firm_changes — field-by-field diffs between successive snapshots.
• cohort_stats — percentile context for AUM, clients, accounts, and employees.
• zip_centroids — US Census ZIP Code Tabulation Area geocoding data.
• ingest_runs, ingest_warnings, ingest_locks — audit trail and concurrency control.

The public directory at /firms lets anyone search and filter all 23,740+ firms by state, AUM tier, and retail status. Each firm has its own profile at /firms/$crd with Item-by-Item provenance and cohort percentiles.

The site is designed as an MCP-first application. The server exposes four public tools at https://advmcp.ai/mcp with no login or API key required:

• search_firms — find advisers by name, city, or state.
• get_firm — full profile for a given CRD number.
• top_firms — largest advisers by reported AUM.
• firm_changes — recent filing changes for a firm.

Rate limits are enforced at 60 calls/minute, 900/hour, and 5,000/day per caller IP to keep the service stable.

A resumable ingestion pipeline runs once a day at 07:00 UTC via a scheduled cron job. It pulls the latest SEC feed, processes it in chunks to stay within platform limits, and finalizes by geocoding firms, computing changes, and rebuilding cohort stats. If the current vintage is already up to date, the run skips automatically.

A private admin dashboard at /admin shows how many assistants have connected the MCP, how many tool calls have been made, latency distributions, tool usage breakdowns, and daily activity trends. Access is protected by the admin password.

Row-level security policies make firm data publicly readable while keeping logging and rate-limiting functions restricted to the server-side service role. A public security review hardened the schema and verified that the MCP endpoint remains intentionally open for read-only access to public SEC records.

The brand is now ADVmcp, with theme-aware logos, a warm custom palette, and support for both light and dark modes. The footer appears on every page, and legal pages (Privacy Policy, Terms of Use), a sitemap, and robots.txt help search engines index the site.