Claims are bound to code
The Claims-Ledger (docs/CLAIMS.md + check_claims in CI) rejects any
public claim that does not resolve to evidence. The anti-example is the
unbacked headline number we refuse to ship.
agent-config
A governed skill/rule/command suite for AI coding tools. Every public claim is machine-checked against the code — including "zero runtime".
Why this is different
Section titled “Why this is different”We sell honesty, so the selling is machine-checked. Every capability or number
in our public docs binds to resolvable evidence — a test, a benchmark artefact,
or a file:line — and CI fails if an unbacked claim sneaks in.
We publish the nulls
The discipline benchmark shows a measured lift on a weak host and a published null on a strong host — the bound, not just the win.
Zero runtime
Static artifact projection into host agents. No daemon, no vector DB, no auto-write memory — a claim you can check in the source.
Try one thing in 30 seconds
Install the production-validator wedge alone — read-only, single-purpose,
carries the anti-hallucination identity.