> ## Documentation Index
> Fetch the complete documentation index at: https://docs.beliefstate.ai/llms.txt
> Use this file to discover all available pages before exploring further.

# Evidence and provenance

> Trace interpreted fields to source evidence and explicit uncertainty.

## Evidence contract

Every interpreted field should resolve to timestamped evidence and a stable source identity.

* Keep source identifiers and exact citations with downstream records.
* Use observation time—not retrieval time—for point-in-time research.
* Keep methodology versions and warnings with calculated outcomes.

## Derived sources

An aggregator summary can document what that publisher represented, but it does not prove the attributed investor's exact wording or original publication.

> Treat aggregator-derived evidence as publisher evidence. Do not label it primary-source or use it to prove originality without independent primary-source verification.

## Author performance and experimental ranking

The ticker brief exposes hypothetical percentage outcomes, sample depth and comparable author cohorts in author\_track\_records.

The experimental score weights benchmark-relative return at 60%, profitable-call rate at 25%, and downside control at 15%. Excess return is mapped from −10 to +10 percentage points onto 0–100; average adverse excursion is mapped from −20% to 0% onto 0–100. Values outside those scales are capped. The weighted score is pulled toward 50 using n/(n+20), where n is the number of comparable, non-overlapping calls.

Five measured calls are required for a score. A rank requires 20 calls spanning 180 days, complete tracked history, next-session-open entries and at least two eligible authors in the same ticker, action, attribution basis, horizon and price source. These policy choices are not calibrated predictions. Negative mean returns remain losses regardless of score or benchmark performance.

Formal Hold recommendations are neutral. A reported holding is long exposure. Sell closes a long; only explicit Short opens a short. Accepted linked exits end affected windows at a verified next regular open. Missing exit prices remain counted gaps. Unlinked articles and partial trims never invent an exit or position size.

> Scores do not verify account profit, combine into portfolio returns, establish global author skill or authorize recommendations. Rankings cover recorded initial beliefs only; revisions and monthly repeats are not new independent calls. Costs and dividends are excluded. Preserve ranking status, sample counts, methodology and citations when displaying results.

## Unknowns

Preserve explicit unknowns so downstream systems can distinguish missing evidence from negative evidence.

Missing evidence stays missing. Clients must not replace nulls, sparse samples, or unresolved provenance with inferred certainty.

* [Inspect thesis evidence](/api-reference/beliefs/find-investor-beliefs-or-inspect-a-full-record)
* [Review source coverage](/api-reference/sources/find-research-sources-by-name-domain-or-source-id)

## Agent feedback

Report unclear, stale, or incorrect documentation through [BeliefState support](https://beliefstate.ai/contact). Include this page URL and the smallest reproducible detail.

*Last updated: 2026-09-18 · API version: v1 (Latest) · OpenAPI publication: 2026-09-13.*


This documentation is built and hosted on [Mintlify](https://mintlify.com), a developer documentation platform.