Research methodology

1. Define the research mandate

A run begins with a supported US ticker and an investment question. The user may choose the full committee or one of four focused teams, then set a short-, medium-, or long-term horizon; core, standard, or deep analysis; standard or strong counterargument intensity; and a decision purpose such as a new entry, holding review, position sizing, or earnings review.

Up to five comparison symbols can be supplied. A company named explicitly in the question may also be recovered as a comparison symbol, but the system does not invent an unnamed peer at this boundary.

2. Collect and seal evidence

The collection layer resolves issuer identity and builds a timestamped evidence snapshot. Depending on availability, it can include SEC filings and company facts, current reports and amendment lineage, insider and beneficial-ownership filings, BLS macro series, US Treasury yield data, and configured provider data such as quotes, price bars, fundamentals, peers, documents, calendars, and categorized news.

Every requested capability is not always available. The run records whether a source family is available, stale, unavailable, or withheld by rights where applicable, and carries provider limitations forward instead of silently filling gaps.

3. Specialist and department analysis

Full-committee research requires eleven accepted specialist memos: three market roles, three company roles, three financial roles, and two risk roles. Each role receives only the evidence needs and decision frame assigned to it, and its claims must reference accepted evidence artifact identifiers.

Each department consolidates its member memos into a department position, including supporting or opposing stances, materiality, contrary observations, unresolved questions, evidence references, and an observable condition that could invalidate a claim. Focused-team research stops at the selected department and is explicitly marked as lacking cross-team review.

4. Challenge, response, and ballot

For full-committee research, the four department consolidations enter a blind challenge round. Department leads challenge assigned claims without relying on the public identity of the original author. The owning department then responds, may revise a claim with authenticated lineage, and records a ballot and remaining disagreement.

5. Structural and semantic audits

The structural audit verifies that the required workflow artifacts exist for the same run and sealed snapshot, that evidence content and locators match their stored hashes, and that claims retain valid evidence and lineage. A full-committee run must present the complete set of eleven memos, four consolidations, four challenges, and four response ballots before this stage can pass.

The semantic audit compares bilingual claims with fixed evidence excerpts and classifies them as entailed, partial, contradicted, or not assessable. Material contradictions can remove or downgrade the affected claim; an audit artifact does not make an uncertain claim true.

6. Chair synthesis and editorial gate

The AI chair receives accepted specialist artifacts, the audited claim register, department ballots, capability limits, and the original question. It must produce a decision with a decisive reason, strongest countercase, observable falsifier, confidence, primary claims, and all four department views.

Before persistence, a deterministic editorial gate evaluates both languages. Hard failures include unsupported numbers, invalid claim ownership, anticipated-question evidence conflicts, excess primary-claim reuse, untyped comparators, and unsafe instructions such as an immediate buy or sell or a guaranteed return. One bounded targeted rewrite may repair permitted fields; unresolved hard failures block publication.

7. Publication and access

A successfully persisted report is stored with its report, run, snapshot, version, source, claim, limitation, and artifact identifiers. Its run ends as completed or complete with limitations. Failed or incomplete stages do not appear as published reports.

The public catalog reads the absolute latest stored version. Eligible reports become indexable after seven full days; before that boundary, non-paying readers see a locked card rather than a crawlable report link.