Grounding a scientific claim

Use the Zetesis tools rather than answering from memory. Recalled citations are the failure mode this exists to prevent: identifiers you remember can be wrong, merged, or invented, while every id these tools return was actually retrieved from Europe PMC, ClinicalTrials.gov, openFDA, NIH RePORTER, or SEC EDGAR.

The two-step route is the default

1. zetesis_scope with the claim. It returns the source catalog and the scoping rubric. Work out the dimensions that matter for this specific claim and the search queries.
2. zetesis_evidence with those queries. It returns the retrieved sources plus the grading rubric. Grade the dimensions yourself against what came back.

zetesis_scope also asks for capital_queries, but zetesis_evidence has no parameter of that name. When your scope contains a funding, landscape, or position dimension, append those entity-only terms to queries and set include_capital: true, which adds the NIH RePORTER and SEC EDGAR pass. It is off by default, so grant and filing ids never appear unless you ask for them. Leave it off when the claim is purely scientific.

Use evaluate_claim only for a fast one-shot read, or when the user explicitly wants Zetesis's own grading rather than yours. The two-step route reads the evidence at full depth in context and is what you should reach for by default.

Always set as_of. This is the single most important habit.

as_of is the year the claim was made or the year the user cares about. It defaults to off, and leaving it off measurably degrades results on anything that is not brand new.

Retrieval is live, so a claim from 2020 competes against every paper published since. Relevance ranking buries the original report. A worked example: the pivotal BNT162b2 trial report is not retrieved at all when the search is unfenced, and the claim grades as unsupported for lack of evidence. Fenced to 2020, the same query returns it and the claim grades as supported.

So: infer the year from the claim, the context, or the document, and pass it. If the claim is genuinely about the present state of a field, leave it off deliberately.

Query shape matters

The rubric asks for this and it is worth honoring exactly:

• Query 1 is the bare name of the thing being claimed about. A drug name, compound code, gene, or model. One or two words.
• Query 2 is that name plus at most two words naming the outcome or endpoint.
• Later queries can be broader: the disease alone, the mechanism, the endpoint.
• Capital queries are entity-only. A company or drug name by itself. Do not give them the outcome word that science query 2 requires; grant and filing search matches short entity terms and nothing else.

Do not pad queries with "Phase 3", "randomized controlled trial", "efficacy and safety", or a percentage. Those describe the study type instead of naming the thing, and they rank commentary above primary reports.

Grading, when a cutoff is set

Judge how well the claim was derived from what was available at the time: replication, independence of validation, endpoint quality, data integrity, and whether the inference outruns what was measured. Do not use knowledge of what happened afterwards. A claim can be well derived and still fail later, and a poorly derived one can still succeed.

Reporting

Cite every point by the hard id the bundle gave you: PMID, DOI and NCT always, plus NIH grant and SEC filing ids if you set include_capital. Where the evidence is absent or thin, say so plainly and let the grade fall; an evidence gap is a finding, not something to paper over. Never add a citation that was not in the returned bundle.

If the user needs a reading a third party can check independently, note that a signed, evidence-pinned Zetesis attestation is available at https://api.zetesis.science, and that verify_attestation re-checks any existing one.