AI Fact-Checking in Biotech: How to Verify Model Output

The most common fact-checking pattern is to ask the model whether its own answer is correct. It rarely helps. The second pass inherits the same reasoning, the same missing context, and the same confident tone, so it usually ratifies the first answer. Confidence scores drawn from token probabilities have the same defect: they measure fluency, not truth.

Real verification requires an independent path that is allowed to disagree. Three are practical today: retrieval against a fixed corpus, recomputation by a deterministic engine, and consensus across models with different training. Each can produce a hard "no" that overrides the prose.

An answer is grounded when each factual claim and citation resolves to a passage in a fixed, versioned corpus. In regulated work that corpus is specific — 21 CFR parts, ICH guidelines, USP chapters, FDA guidance — not the open web, because a citation that changed last week invalidates a study file that referenced it.

The check itself is mechanical: extract every citation the model emitted, resolve it against the index, and fail the answer if any citation has no match. Fabricated references are the single most common failure mode in AI biotech output, and they are also the easiest to catch, because a hallucinated regulation simply does not exist in the index.

Most quantitative biotech questions are not open-ended. Power and sample size, allometric dose scaling, assay acceptance ranges, shelf-life regression, and bioequivalence bounds all have closed-form or simulation-based answers. That makes them verifiable in the strong sense: the number the model produced can be recomputed independently and compared.

The rule that matters is what happens on mismatch. If the recomputed value differs beyond a tight tolerance, the answer is wrong — the narrative does not get a vote. Quandra recomputes numeric output with a numerical verifier and a symbolic solver, and rejects the answer on drift above 2%, sending it back for a retry rather than shipping it with a caveat.

The goal is not an answer to every question. It is a truthful boundary. A system that answers everything at 95% accuracy forces a reviewer to re-derive all of it, because there is no way to tell which 5% is wrong. A system that refuses when it cannot ground or recompute converts an unknown error rate into a known, finite review queue.

Calibrated refusal means declining for a stated, machine-readable reason: no corpus support, verifier disagreement, out-of-scope clinical judgment, or detected patient identifiers. Refusals should be logged and reviewable, because the refusal log is the fastest map of where the corpus needs to grow.

Verification that is not recorded did not happen, as far as an auditor is concerned. For an AI-assisted analysis to survive review, the record needs the prompt, the model and version, the corpus version, the verifier and numeric-check results, the final output, and the identity and timestamp of the qualified human who signed off — written so that later modification is detectable.

Two-person review belongs in the same record. The person who generated an analysis should not be the person who approves it, and the system should enforce that rather than trust it.

• Can the system show me the source passage behind every citation?
• Is the corpus versioned, and is that version pinned into the output?
• Is every number independently recomputed, with a stated tolerance?
• Does the system refuse — and say why — when it cannot verify?
• Is there a tamper-evident record of who reviewed and approved the output?
• Are patient identifiers blocked before they reach the model?

Can an AI model fact-check itself?

Not reliably. A model asked to grade its own answer reuses the same flawed reasoning. Useful verification comes from an independent path — a deterministic recomputation, a symbolic solver, or a separate model with different training — that can disagree with the original answer.

What is groundedness in AI fact-checking?

Groundedness means every factual claim and citation in an answer maps to a passage in a fixed, versioned source corpus. If a citation cannot be matched to a real indexed document, the claim is ungrounded and should be withheld rather than shown.

How do you verify statistical output from an AI?

Recompute it. Sample-size, power, dose-scaling, and stability calculations have closed-form or simulation-based answers, so the number the model produced can be recomputed by a numerical engine and compared. A drift above a small tolerance means the answer is wrong regardless of how confident the prose sounds.

Why does refusal matter more than accuracy in regulated work?

An accurate answer 95% of the time with no signal about the other 5% is unusable in a regulated workflow, because every answer must then be re-derived by hand. A system that refuses when it cannot verify converts an unknown error rate into a known, reviewable queue.

What has to be recorded for an AI answer to be auditable?

The prompt, the model and version, the corpus version, the verifier results, the numeric checks, the final output, and the identity and timestamp of the human who reviewed it — stored so that later tampering is detectable.