// part of the AXON ecosystem

GNOMON

Gated Numerical Offline Metrics Over N-cases. GNOMON is not a product feature: it's the reason any measured number on this page (or the homepage) can be trusted at all. It scores faithfulness and context precision with a confidence interval attached, reports cost and latency in the same table as quality, and fails the build when a metric drops below a threshold. No judge metric is ever reported as a bare number.

github.com/sammyjdev/gnomon-eval· Apache-2.0 · 11 ADRs, 77 tests, offline-first (Ollama by default, zero cost to run).

// measured, not vibes

The result it produced, and the bug it caught in itself.

AXON recall A/B (the source of that number)results/2026-07-02-ab-recall · 2026-07-02

0.40–0.52 → 0.72–0.76answer faithfulness, recall off vs. on. 95% CI, intervals do not overlap

measured, provider-meteredN = 17 cases+2,151 input tokens / turn

// this is the same result AXON's homepage evidence card cites; GNOMON is where it actually came from.

ADR-008: the fix for an inflated ndocs/adr/0008 · 2026-05-30

n = casesnot cases × judge runs: the earlier method double-counted every run as an independent observation

bootstrap CIbounded [0,1] by construction

// a narrower confidence interval isn't honest if it comes from counting the same case twice. Fixed before trusting any score built on it.