Seven in ten researchers have failed to reproduce someone else’s result.

Kelvara does not predict, design, or discover. Each agent answers one narrow question about whether published evidence supports something you are about to rely on: a reagent, a sequence, a target, a claim.

That is a smaller job than most tools in this field claim, and it is the job we can do honestly without a wet lab behind us.

Every agent reports what it could not establish. Which sources were unreachable, which dimensions had no evidence either way, and what the verdict does not cover. The limits block is mandatory and never empty.

Accuracy is published with confidence intervals against named benchmark datasets, including the cases we get wrong. If an agent cannot answer, it says so rather than guessing.