Vancouver and London · barglabs.ai · cejel.dev

Proof for code you didn't write. AI agents now write code and report on their own work. The people who must accept that work get the report, not evidence they can check. We build the evidence.

A fixed-scope review of one repository that answers one named question, with evidence your customer, buyer or acquirer can rerun and get the same result. Two parts, kept separate: an evidence review with Cejel, and a senior engineer's adversarial reading of the code against your own written claims, including whether a pass rests only on tests the author chose and whether new code is actually called. Two to three weeks. Details at barglabs.ai.

A deterministic, offline trust certificate for source code. Point it at a repository and it scores eleven criteria against a published rubric, binds every criterion to a specific file, line and content hash, and issues a certificate. No network call, no model call, no telemetry, and it never uploads source. When the recognised source is too thin to support a verdict it returns insufficient_source and declines to score rather than guessing.

There is no AI in Cejel. Same repository, same revision, same release, same rubric: same result, byte for byte. That is what makes the output evidence rather than opinion.

npx @cejel/cejel .

Also available as signed binaries, a container image, a GitHub Action, a Homebrew tap, and on the MCP Registry. AGPL-3.0.

Checks the claims in a coding agent's report of its work (files changed, tests run, issues closed) against the pull request's own record, and reports the verdict as a check run. No model decides the verdict, and anyone can rerun it from the record it writes. On npm as dunstan, as a GitHub Action and as a Claude Code plugin. AGPL-3.0.

Preregistered, published calibration of other people's AI judges, reviewers and scanners against known answers. Two so far: a hosted judge and an open-weights judge. With the evidence attached, both still passed most false reports.

Rubric witan-rubric-v17-2026-07-24, decision date 25 July 2026, on a preregistered 200-repository holdout the scanner had never run against. Thresholds locked 22 July, before any result was visible. Three independent blind reviews.

• Finding precision 96.43% (95% lower bound 94.16%)
• Detection precision, cycle 13: 64.3% (95% CI 38.4-87.6%). A different measure from finding precision; the gap between them is not yet explained.
• Worst-case false-positive rate 0.66% (95% upper bound 1.10%)
• Rubric-agreement recall 95.64% (lower bound 92.23%)

Rubric-agreement recall measures agreement with blind reviewers on the same bounded evidence. It is not detection recall: it does not mean Cejel finds 95.64% of real defects. Detection recall is measured separately and is lower: on cycle 13, v17's in-scope detection recall is 41.1% (95% CI 20.6-67.8%). What each figure measures. This is finding-level calibration of one rule set on one population: not a security guarantee, not vulnerability detection.

The run before the passing one was a terminal no-go, published in full, failing on inappropriate abstention at 12.70% against a 10% cap. A separate free LLM pack sits at 0.0 recall against a 0.65 floor: zero of thirty-four known defects found, three findings all false positives, published with the root cause named as a detector-architecture failure.

Our public leaderboard scores 24 repositories including our own, which sits mid-table at 2.8/4.0. We dogfood our own tools before we ask anyone else to run them.