CausalSentinel

v0.1 — 2026-07-13 · the model wrote the whole card

Eight tools were wired up, but get_mr_result returned a placeholder ({"stub": true, "beta": null}) and every part of the card was written by the model, including the evidence table and the source list.

Kept because this is the only record of what the system says with no guardrails.

What it got wrong, and where each was fixed

On the card Why it is wrong Fixed in
Causal effect (MR) … Stub: Placeholder resultsand the verdict is still GO A placeholder is not evidence; nothing should have been concluded from it v0.2
51 unique GWAS SNPs The tool read page 1 of 3. True count is 114 from 256/256 rows v0.2
216 records, no pathogenic identified in sampled set Reads as 0/216. It was 0 out of a sample of 30 v0.2
No statement of what “MASLD” resolved to Every downstream score answers a question the reader cannot see v0.2
“high tolerance to LoF … suggests that inhibiting the protein is likely to be safe” Constraint read as a licence. LoF tolerance is not evidence that inhibition works v0.2 wording, v0.3 printed on the figure itself
Sources are bare links, no database release Timestamped, not reproducible v0.2

Same input under a later version: ../v0.2_2026-08-14/PNPLA3_MASLD_evidence_card.md and, for the current version, ../../PNPLA3_MASLD_evidence_card.md.