How to read the evidence.

Research workflow

From source to public record

Source hierarchy

Primary evidence first

We prefer the original document appropriate to the question: official datasets and administrative records, filings, technical documentation, original research, or an issuer’s announcement. Source quality depends on the claim: a company page may establish what a vendor says its product does, but independent evidence is stronger for performance conclusions.

1. Official datasets and administrative records
2. First-party technical documentation
3. Regulatory filing
4. Official press release
5. Academic publication
6. Reputable journalism
7. Other secondary source

Source check

A numeric check, not a fact-check verdict

A separate program looks for significant numeric tokens in retrieved source text: numbers with three or more digits, decimal points, currency symbols, or percent signs. It can reuse stored retrieval snapshots, including browser captures. It does not check every word, every small number, or whether a comparison is meaningful.

The run dated 2026-09-17 had usable text for 209 of 210 cited sources, including 20 browser captures. Numeric-presence checks passed for 369 of 370 assertions; 0 had unmatched tokens and 1 remained unchecked. This is the check-run date, not a claim that every source was retrieved again on that date.

A matching number may describe a different thing in the source. Passing this check establishes neither correct interpretation nor truth, freshness, independent corroboration, or human review. The check never changes a claim’s review or publication state.

Three deliberate choices. A source we could not re-read is never counted against an assertion, because a figure missing from text we never saw proves nothing. A page that answers a script with a refusal but a reader with the page is read the second way, and counted apart from the automated ones, because re-running the program does not reproduce it — while a paywall or a bot-verification challenge is not passed by either route, and what sits behind one stays unread and is reported as such. And a figure printed in order to refuse it — a circulating number this index declines to assert — is counted separately rather than checked, because that sentence makes no claim about the source’s contents.

Confidence

A score is not a verdict

Confidence records the verifier’s assessment of evidence support and extraction reliability. It never replaces the source, evidence excerpt, or status. Contradictory claims are preserved and linked rather than silently overwritten.

Publication policy

Verification and visibility are separate

Topic publication requires the founder’s approval. Within a published topic, individual assertions may still be labeled Reported or Unverified; publication is not per-claim human verification. Only the Primary source and Corroborated assessment labels denote completed human review under the evidence rules. Company-profile claims use a separate, stricter review gate. Reading-guide prompts are editorial framing, not additional evidence or causal findings.

Corrections

Corrections retain a record

Corrections retain an internal audit history rather than erasing the earlier record. A public change log and corrections intake are not yet available. The current topic pages show the current assertions and their review labels, not the complete history of revisions.