A Small Provenance Schema for AI-Generated Stock Summaries

An AI summary is easier to audit when each claim carries its source. For developers building research tools, a useful starting point is a small provenance record.

{
  "claim": "Revenue increased year over year",
  "source_url": "<original-company-filing-url>",
  "reporting_period": "<period>",
  "basis": "consolidated",
  "page": "<page-or-section>",
  "checked_by_human": false
}

Why these fields matter

A reporting period prevents a recent summary from silently repeating an old number. A basis field prevents mixing standalone and consolidated accounts. A page reference makes it possible for a reader to reproduce the check without searching a long document.

Validate before displaying

Treat model output as untrusted input. Parse it against a schema, check that the URL belongs to an expected source, and display an explicit missing-source state when the field is absent. A well-formed URL alone does not establish that the document supports the claim.

Keep extracted figures separate from narrative interpretation. Store units and currencies explicitly; do not let the interface imply that a percentage, a currency amount and a per-share value are interchangeable.

A minimal review loop

1. Fetch or upload the original filing.
2. Extract a claim with its passage and document location.
3. Verify period, units and accounting basis.
4. Let a reviewer accept, amend or reject the claim.
5. Preserve the source reference when generating a shorter summary.

This is a suggested design pattern, not a description of a production architecture or a claim of improved model accuracy.

Disclosure: this post is from the StockTool.AI account. StockTool.AI is our stock insights and tracking platform; the pattern above is relevant to tools that organize financial research. AI output should be checked against original disclosures. This article does not provide investment advice.

Written with AI assistance.