Question
The failure mode: models write citations, then hope
A research assistant that retrieves evidence and then synthesizes an answer can emit a plausible-looking [PMID:12345678] that was never in the evidence pool. The model writes the marker first and checks the pool second — or not at all. In our own audits, fabricated markers were a recurring failure class: they look like provenance while being the opposite of it.
Method
The gate: citation markers are adjudicated before render
Since v0.20.0, citation markers in free-form reasoning text pass through a deterministic adjudication before rendering: markers found in the evidence pool are kept; markers that are out-of-pool but verifiably real (checked against NCBI) are kept and labeled external; hallucinated or unparseable markers are stripped from the text with the sentence's meaning preserved. The gate only removes false markers — it is designed to leave the dossier intact and does not silently rewrite citations. It is on in production, and the dossier shows what was removed.
The probe: we feed it fabrications on purpose
A gate is a claim until something tries to break it. Our standing adversarial probe builds a deliberately poisoned document — fabricated citations and a fabricated titer claim — and runs it through the review layer. Healthy means each fabricated item must be stamped WARN or FAIL; any PASS fails the build. The probe runs in CI on every change to the review layer, and it prints its verdict and exit code into the build log: if a fabricated claim ever receives PASS, the build fails — that failure is the alert.
Evidence
The record
The early probes were both caught — session verdict FAIL, fabricated claim stamped. The standing probe that entered CI has been healthy on every run: the fabricated citations and titer claim were stamped unverified; none received PASS. This page is a snapshot; the probe re-runs on every change, and the build log is the running record.
Limits
What this does not cover
The gate adjudicates citation markers, not prose: an uncited sentence can still be wrong. The review layer checks whether a cited source actually supports a claim — same organism, same target, entailed or contradicted — not whether the claim is true in the world. And no dry-lab tool substitutes for wet-lab validation. We repeat this boundary wherever the review layer is described because a probe that sounds like a guarantee is just another fabrication.
References
- Kairos Trust — the standing adversarial probe, its CI wiring and the 2026-07-27 / 2026-09-20 record. /trust/
- Kairos Benchmarks — the dry-lab evaluation context these gates sit in. /benchmarks/