Method

How Priors grades a claim.

One fixed evidence rubric, applied to the primary literature behind each claim, with every paper and every reason shown.

Compared with commercial AI How a claim is graded Expert review

Compared with commercial AI

AI gives answers. Science needs to know how certain they are.

	Commercial AI	Priors
Evidence	Weighs text, not evidence: how often a claim is written, not how well it has been tested.	Grades each claim against one fixed evidence rubric, 0 to 100.
Uncertainty	Agreeable, and rarely says “we don’t know”. On 43 questions where Cochrane reviewers found the evidence too uncertain to call, a frontier model still gave a definite answer 40 times.	States how uncertain the evidence is, and withholds a grade when the evidence cannot carry one.
Stability	Unstable. Asked the same question twice, a frontier model changed its answer 35% of the time.	Stable. Re-graded on the same papers, a grade typically moves by about 3 points out of 100.
References	Fabricates references. 7% of the references a frontier model cited did not exist.	0%. Every reference is checked against the DOI registry before it can appear.

Sources: Polzak et al., ICLR 2026; Priors’ own measurements on 40 Cochrane claims published after the model’s training cutoff, September 2026.

How a claim is graded

One rubric, applied the same way to every claim.

As GRADE does for Cochrane and calibrated language does for the IPCC, a fixed rubric makes a grade in one field mean the same as a grade in another.

Step 1

Extract the major claims

Each claim is stated precisely using PICO.

Step 2

Assemble the evidence

Primary papers from Consensus, OpenAlex and PubMed; every DOI resolved before it can appear.

Step 3

Grade against the rubric

Four reviewer roles, one a dedicated sceptic, score consistency, robustness, independence, fit to the claim and completeness.

Step 4

Decline when it cannot

If the evidence does not bear on the claim, it is withheld with the reason published.

Standing	Certainty	Meaning
Established	≥ 85%	Strong, consistent evidence. Unlikely to change.
Likely	70–84%	Well supported, with some gaps or indirect evidence.
Uncertain	50–69%	Mixed or limited evidence. Genuinely open.
Doubtful	30–49%	Little support; the weight of evidence leans against it.
Refuted	< 30%	The evidence contradicts it. Confidently false as stated.

Expert review

Experts decide what counts as evidence.

Independent experts check each field for a missing load-bearing paper. They correct the evidence, never the score, and the rubric re-grades. A reviewed field carries a dated label naming the release reviewed.