/cross-validate

Cross-validate this claim against independent databases: $ARGUMENTS

A single source for a high-stakes fact is fragile. Force 3+ sources from DIFFERENT databases (not 3 tools that all wrap the same upstream API).

Process

1. Decompose the claim into testable assertions

A claim like "BRAF V600E is FDA-approved indication for vemurafenib in melanoma" has three independent assertions:

• (a) BRAF V600E exists as an annotated cancer variant
• (b) Vemurafenib is FDA-approved
• (c) The approved indication includes BRAF-V600E-positive melanoma

Each assertion can be confirmed/refuted separately. Write them out before calling tools — this is the anchor for what you'll verify.

For a numeric claim ("X% of breast cancers carry TP53 mutations"), the assertion is the value itself; cross-validate by getting the same number from independent cohorts.

2. Pick 3 independent databases per assertion

"Independent" means different upstream maintainers, not different tools wrapping the same API. Examples:

Domain	Independent sources
Drug approvals	OpenFDA, ChEMBL, DrugBank — each maintains its own approval table
Variant pathogenicity	ClinVar (NIH), CIViC (academic crowd-sourced), OncoKB (MSKCC), gnomAD frequency (Broad)
Gene-disease association	OpenTargets (composite), DisGeNET, OMIM, GenCC
Mutation frequency in cancer	TCGA (GDC), COSMIC, IntOGen — each samples different cohorts
Clinical trial existence	ClinicalTrials.gov, EU CTR, ChiCTR
Protein function	UniProt, Reactome, KEGG
Pharmacology	DrugBank, ChEMBL, PharmGKB, FDA Label

Pick 3. If the claim's domain only has 2 truly-independent sources, note that in the report — don't fabricate a third.

3. Run each source independently

Use tu run <tool> per source. Don't shortcut to a compound tool that aggregates multiple sources — that defeats the cross-validation point. Track each source's verdict separately:

Source 1 (ClinVar): CONFIRMS — V600E listed as Pathogenic, condition: Melanoma
Source 2 (CIViC):   CONFIRMS — V600E linked to vemurafenib sensitivity in melanoma
Source 3 (OncoKB):  CONFIRMS — V600 variant Level 1 evidence for vemurafenib in melanoma

If a source returns "no record found", that's not the same as "refutes" — mark it as INDETERMINATE and note possible reasons (different namespace, absent in that database, query string mismatch).

If a source explicitly contradicts (e.g., gnomAD shows population frequency suggesting it's a benign polymorphism), that's REFUTES.

4. Report concordance

Format:

## Claim
<verbatim claim>

## Verdict: <CONFIRMED | DISPUTED | UNSUPPORTED>

## Per-source results
| Source | Verdict | Evidence | Date |
|---|---|---|---|
| ClinVar | CONFIRMS | V600E Pathogenic for Melanoma | 2025-Q4 |
| CIViC | CONFIRMS | V600E + vemurafenib (Level A) | 2026-01 |
| OncoKB | CONFIRMS | Level 1: vemurafenib in BRAF V600 melanoma | 2026-Q1 |

## Concordance: 3/3 confirm

## Caveats
- All three databases share clinical-trial sourcing (overlapping evidence)
- (Or:) Sources disagree on dosing — ClinVar lists 240mg BID; OpenFDA label lists 960mg BID

The verdict logic:

• All sources confirm and none refute → CONFIRMED
• Sources split (some confirm, some refute) → DISPUTED (state the split)
• All sources indeterminate or absent → UNSUPPORTED (no contradictory evidence either)
• Any source explicitly refutes a CONFIRMS-from-others case → DISPUTED, prioritize naming the dissenter

5. Note the limits

End with a "Caveats" section flagging:

• Whether the 3 sources are truly independent or share upstream data
• Whether any source is regional/legacy (e.g., FDA label is US-only)
• Whether the claim's TIME-SENSITIVITY matters (drug approvals change; cite the date)
• Whether numeric claims have a stated tolerance ("> 50%" vs "exactly 53.7%")

Stop conditions

• Can't find 3 independent sources for the domain → use what you have, state you couldn't reach 3, note which sources DO exist.
• A source returns API key error → mark as UNREACHABLE, replace with one alternate. Don't retry.
• 4+ tool calls per assertion without result → stop, report "INDETERMINATE across N attempted sources".