Methodology

What SEER's citations claim, how much of that is checked by machine, and where the limits are.

The premise

SEER asks the report generator to cite factual assertions with a specific, dated, tier-labeled source document. A citation makes a claim checkable; it does not establish that the claim is correct. When a source URL is recorded, the citation drawer links to the document or a matched passage. Missing links and evidence are identified rather than replaced with a claim of support.

Source tiers

Sources are ranked by how directly they bind the company to what they say. A regulatory filing is a stronger record than a job posting, and a report should never let the second carry a claim the first could not.

T1	Primary filings. The company's own words, filed under legal obligation. 10-K 10-Q 20-F 8-K 8-K/A Form 4
T2	Company-spoken, unfiled. Management on the record but outside a filing. Earnings Call Investor Day Conference Presentation Investor Presentation Shareholder Meeting Guidance Update Special Call Trading Statement M&A Call Operating Results Call Fixed Income Call Interim Update
T2.5	Third-party industry research. Independent, but not the company's own record. Industry Research
T3	Company web properties. Current but undated and revisable, so cited as undated. IR Page
T4	Personnel and hiring signals. Weakest tier; corroborating, never load-bearing. LinkedIn
MKT	Market data. Price and FX, point-in-time, never described as live or current. Close Intraday FX Price History Peer Close
SEER	SEER's own prior published thesis. Analytical context only — never evidence for a factual claim about a company. Thesis

Citation grammar

One canonical form: [TIER, DOC-TYPE, YYYY-MM-DD]. Dates are ISO or the literal word undated — never a retrieval date standing in for a publication date, which is how an undated web page silently acquires today's authority. Several sources for one claim appear as adjacent brackets. The tier and document-type vocabulary above is generated from the same definition the report generator is bound by, so this page cannot describe a grammar the pipeline does not enforce.

Source matching and claim support

The Audit recorded badge means a structured audit summary is stored. It does not certify the report, establish publication approval, or mean every claim was checked successfully.

The historical proven category counts claims with an excerpt located and no mechanical conflict recorded. Matching used exact, normalized or fuzzy text search. Authentic text can be irrelevant to a claim; detecting no conflict does not establish support. The historical judged category is the remaining claim count, not a count of confirmed model verdicts.

Model assessments can disagree with source-matching results and can themselves be wrong. The legacy totals do not establish whether all expected model verdicts were returned, so missing verdicts cannot be read as passes. Recorded conflicts and flags may overlap and should not be added into a factual-error count.

Available counts also identify malformed citations, documents absent from the supplied sources and missing evidence entries. Missing measurements are shown as not recorded, rather than as zero defects.

Original report text and PDFs retain historical auditor wording such as “proven” and “passed.” This display does not rewrite or re-audit those artifacts; those terms are not current verification guarantees.

Legacy reports

A report marked legacy has no readable structured audit summary stored. Historical prose may still contain audit observations. Missing structured results establish neither source support nor successful completion of checks.

Limits