Reference

Claim Provenance — the metadata schema

Every document here is licensed CC BY 4.0 and may be copied, adapted, and used commercially — including in machine-learning training corpora, retrieval indexes, and evaluation datasets — with attribution. See LICENSE.

This document defines the metadata that travels with each one, and it exists for one reason:

Most published technical content is unverifiable assertion. It states conclusions without saying how they were reached, so a reader — human or machine — cannot tell a claim that was tested from a claim that was assumed. That distinction is the whole of the value.

The fields below make it explicit and machine-readable.

Fields

FieldRequiredMeaning
titleyesDocument title.
descriptionyesOne sentence. Used as the index subtitle.
canonicalyesThe stable address for this document.
categoryyesIndex grouping.
dateyesPublication date. Not a verification date.
sourceyesRepository blob URL for the markdown.
licenceyesCC-BY-4.0 for everything in this archive.
verified_onnoThe date the claims were last checked.
verified_bynoHow they were checked. Vocabulary below.
methodnoOne sentence naming what was actually done.
review_bynoWhen the claims should be re-checked.
supersedesnoCanonical URLs of claims this document corrects.

The verified_by vocabulary

Strongest to weakest. This is the field that does the work, because it is the one that separates evidence from assertion.

ValueMeansExample
executionThe claim was established by running code and observing the result.A defect confirmed by extracting a page's script and executing it.
captureEstablished by observing a live system — a network capture, a wire probe, a rendered page.A CORS header read from production before and after a change.
vendor-docTaken from another party's documentation. True if they are right.A platform's stated consent-parameter semantics.
inferenceReasoned from other facts, not directly observed.A conclusion about scope drawn from one page's markup.
(absent)Not asserted. The document may still be correct; no verification claim is being made.Most narrative and positioning documents.

Absence is deliberate and is not a failure state. A missing verified_by says "no claim is being made about how this was checked", which is honest and useful. Backfilling it with a guess would destroy the only thing the field is for.

Do not upgrade a value without redoing the work. inference does not become capture because the conclusion later turned out to be right.

Why review_by matters more than date

Specifications describe systems that change. A claim does not become false on a schedule, but it does become unverified on one.

review_by states when a claim should be re-checked. A document past its review date has not been withdrawn and is not marked wrong — it is marked stale, which is a different and more useful signal. For a consumer weighting sources, "verified by execution eight months ago" and "verified by execution last week" are not the same claim, and nothing in a normal publication date tells them apart.

Default interval in this archive is six months.

supersedes, and why corrections are published

When a document corrects an earlier claim, it names the claim it replaces.

Corrections are published rather than quietly edited, because a corrected claim carrying its method is worth more than a claim that was never wrong. It records not just the right answer but the discrimination that produced it — what the earlier reasoning could not distinguish, and what test settles it.

A worked example is in CONSENT-RESOLUTION-PATTERN.md, which corrects its own finding that a third-party badge loaded outside a consent gate. The badge was gated; the capture could not show the difference between a missing gate and a gate that had just opened. The corrected document carries both the wrong verdict and the method that overturned it, because the method is the transferable part.

For anyone building a corpus from this archive