Two layers that measure different things and don't cover for each other. This page explains what each one does, what it costs, and where its ceiling is.
It runs over your git history metadata and a temporary read-only clone. 26 checks across five pillars: documentation, agent instructions, executability, history and structure. It calls no model: it's file reading and arithmetic.
Its ceiling: it measures that things ARE there, not that they tell the truth. A well-groomed shell scores well.
The rubricWe write the probes; your own agent answers them on your machine or in your CI, with your credentials. Is this claim in the docs true against the code? Can you find where each responsibility lives? It contributes observations citing file and line; we do the arithmetic. We don't trust the judge: we trust the instrument.
Between 190,000 and 230,000 tokens and 6 to 11 minutes per run, almost regardless of your repository's size: what governs it is how many probes there are. Measured, not estimated.
A negative control. We cloned our own repository and falsified its documentation without touching a single line of code: paths that don't exist, invented commands, architecture described backwards. The files were still there, the same size and with the same live links.
Δ 0
773 lies · 90 documents · 2026-08-25
87→ 60
Autonomy bandL3 → L1
503 lies · 59 documents · real judge, 2026-08-12
Every failed check brings its playbook: the principle, what to do, how it adapts to your repository's conventions, and the criterion by which it counts as closed. Your agent runs it, you approve the changes — the gate is human, always — and on the next analysis the check passes and the card closes by itself. Nobody declares victory by hand.
Claude Code, Codex, Cursor, and any client that speaks Model Context Protocol. The connector exposes the score, the findings, the remediation playbooks, and the deep layer's probes as tools your agent already knows how to use.
Every finding has a path, and the path ends when the re-analysis confirms it.
GitHub or Bitbucket. We read your git history metadata; the code stays where it is.
Five pillars, each one expandable down to the check that lowered it, with the evidence next to it.
Over MCP we hand it the playbook for each failed check, adapted to your repo's conventions. You approve the changes: the gate is human, always.
On the next analysis, if the check passes, the card closes by itself and says so. Nobody declares victory by hand.
You connect the MCP client you already use and the deep analysis runs there, with your credentials. We send the questions; what comes back are observations citing file and line.
On your machine or in your CI. Nothing to upload anywhere, and no read access to hand us for the deep half.
Claude Code, Codex, Cursor, or any client that speaks Model Context Protocol. The connector exposes the score, the findings and the playbooks as tools your agent already knows how to use.
Not a promise of good will: the connector's tools refuse file contents. From code we accept path:line and your description, never the literal quote.