grounding-claims

Grounding claims

Machine-generated hypotheses cite real-looking sources that sometimes do not say what the text claims. This skill checks that, one claim at a time. Run it before anything leaves the session: a shared write-up, a grant paragraph, a decision to run an experiment.

What gets checked

For each factual claim:

1. Does the url resolve?
2. Does the excerpt actually appear on that page?
3. Does the source support the claim, or merely mention the topic?
4. Does it support the direction of the claim? Activation is not inhibition, correlation is not causation, mouse is not human.

Check 4 catches the most errors. A source that discusses the right molecules in the wrong direction reads as support to anyone skimming.

Where the sources are

• co-scientist:hypothesis_get returns the hypothesis, its reviews, and its artifact. The citation list with urls and excerpts is in there: at artifact.citations for an agent-recorded hypothesis, and under artifact.record for one the engine's Generation agent wrote.
• Review evidence rows carry claim, url and excerpt, which is exactly the triple you need.
• Tool provenance from a session run records which urls an agent actually opened. A citation that appears in a hypothesis but in no tool call was not read by the agent that cited it. Treat those first.

Workflow

- [ ] 1. Extract every factual claim with its source
- [ ] 2. Fetch each source
- [ ] 3. Verdict each claim
- [ ] 4. Record the finding
- [ ] 5. Report

1. Extract

List the claims as a table before checking anything: claim, url, excerpt. Work from the list. Checking claims as you read prose means missing the ones stated without a citation at all, which are the worst kind.

Claims with no source go in the list too, marked as uncited.

2. Fetch

Fetch each url. Search the page for the excerpt. If the page is paywalled, try the abstract, the preprint, or the DOI landing page, and say which one you checked.