Every financial AI product now claims accuracy. Some promise guarantees. We took a different route: build the system so that the dangerous failure — a confident, wrong number — is structurally hard, and so that every claim can be checked by the reader against the cited source. This page explains the actual machinery, because “trust us” is not a methodology.
Layer 1 — the numbers never come from a model
Everything quantitative in AnalystBook — financial statements, segment tables, forensic scores, guidance changes, pay figures, ownership stakes — is extracted by fixed rules reading official SEC filings: XBRL for the financials, dedicated parsers for proxy pay tables, Exhibit 21 subsidiaries, Form 4s, 13D/G stakes, comment letters. The same filing always yields the same figures. A language model never decides a number, so a language model can never hallucinate one into your financials.
Layer 2 — the AI reasons over that record, not the web
Ana, our AI analyst, answers through read-only tools over the structured record — verbatim filing sections, period-aligned financials, insider activity, events. No web browsing, no answering from the model's training memory. If the record doesn't support an answer, the correct output is a refusal — and she's built to say “I don't have that in the filings yet” rather than improvise.
Layer 3 — gates check the AI before you read it
Grounding isn't a prompt instruction; it's enforced after the model writes. Every statistic in an answer is compared against the record and dropped if it doesn't match. Quotes get the subtler check — management “quotes” that sound verbatim but were never read from a filing in that conversation — and forces a corrective pass: read the document or drop the narrative. Investigations run under hard time and context budgets, so a long-running ask finishes honestly from the evidence gathered instead of grinding into nonsense.
Layer 4 — published analysis is audited against the SEC
Our Intelligence Briefs are written by AI that reads the actual filings under strict rules (never from memory; a figure not in the source is “N/A”, never estimated). Then every figure the brief states is audited against raw SEC data and a computed year-over-year change report. A brief with a mismatched figure is rejected and never published — the rejection is stored for internal audit. Published briefs carry their gate verdict and are monitored for staleness against the filing calendar.
Layer 5 — measurement, every day
Ana's behavior is pinned by a certified question bank across dozens of research desks — questions whose correct answers have been verified — and evaluated daily. Every answer in the product carries a one-click flag; flags land on an internal health dashboard we actually read. When something slips through, it becomes a new gate or a new certified question.
What we don't claim
We don't claim infallibility — no honest AI vendor can. We claim something more useful: verifiability. Every number is computed from the filing by fixed rules; every AI answer carries its cited sources. You never have to take our word for anything, which is the only trust model that belongs in financial research. For research purposes only; not investment advice.
This is the long version of the six design decisions on the Technology overview.