Quality metrics — how the council performs

Genesis runs an automated nightly eval harness against 8 representative + adversarial fixtures using our local 32B reasoning model as judge. Each fixture exercises the full rewrite pipeline + all 6 guardrails (5 deterministic + 1 LLM-judged factuality with source citation). Regressions trigger alerts and block deploys. This page shows the current state. No signup, no login.

What the numbers mean today: current scores reflect a system that is acceptable for pilot / draft / human-approved workflows — which is exactly the Founding Pilot motion. They are not yet at the threshold for unattended auto-publish; that’s the whole point of approval-required by default. As the eval scores climb and stay clean across content types, auto-publish unlocks per type post-pilot.

Per-fixture baselines

Floor scores set at fixture creation. Each nightly run compares against these to detect regression. Higher judge score = better rewrite quality vs the brief and source facts; pass rate = fraction of the 6 guardrails (5 deterministic + 1 LLM-judged factuality with source citation) that passed for that fixture.

Normal-case fixtures (5)

Adversarial fixtures (3)

These intentionally bait the pipeline into failing the guardrails (restricted-claim trap, hallucination trap, plagiarism temptation). The point is they trigger the right guardrails — partial pass on these is the expected, healthy state.