Nerviq publishes several kinds of evidence — a public benchmark corpus, a human beta evaluation, case studies, a reproducible reference corpus, and a before/after proof repo. This page ties them together in one place, explains what each one proves and does not prove, and states clearly what the full chain still does not add up to.
Six artefacts, from broad (corpus-level robustness) to sharp (one worked example).
Continuous, automated evidence that the audit itself is stable across real-world diversity.
The standing corpus Nerviq's CI gate runs against on every push — 61 real public GitHub repositories across Claude, Cursor, Codex, Copilot, Gemini, Windsurf, and Aider.
Proves — stable behaviour on real-world diversity, strict FP rate under 5% per certified platform.
Does not prove — customer adoption.
A separate public repo (nerviq/nerviq-reference-benchmark-corpus) packaging five archetypes — Node API, Python ML, Go payments, Flutter mobile, platform monorepo — with a rebuild script and baseline score tables.
Proves — scoring is reproducible by an outsider.
Does not prove — correctness of the scoring, only reproducibility.
Curated repositories reviewed by actual evaluators, with published findings.
Eight repos audited by human evaluators across a spread of stacks. Scored the experience, surfaced product gaps, filed raw observation packets.
Proves — useful output on real repos per real humans. Flagged gaps all subsequently closed.
Does not prove — production usage.
Five of the eight evaluations written up with before / after scores, raw observations, and a clear scope line per study. Linked from /case-studies.
Proves — evaluation process is transparent and auditable.
Does not prove — customer testimonials.
Four deterministic archetype fixtures — Flutter mobile, iOS Swift, mature Python ML, FastAPI — used to prove the framework-native verification pass actually closed the mobile + mature-Python gap the evaluation reported.
Proves — average +11.25 point uplift across the four archetypes after the fix.
Does not prove — every mobile repo gets +11.25 points.
One real multi-agent repo, shown in both states.
A single public repo (nerviq/nerviq-multi-agent-before-after) showing a real multi-agent codebase before Nerviq (messy config, drift, missing guardrails) and after (aligned, scored, governed). Raw reports committed alongside the code.
Proves — concrete end-to-end uplift you can read yourself, commit by commit.
Does not prove — that every repo sees the same uplift. One worked example, not a statistical claim.
We are explicit about the limits because an evidence chain is not the same as customer adoption.
Without this page, our numbers had to be pieced together from several places.
"20+ real repos" in Harmony docs, "8 evaluated repos" in case studies, "5 published studies", plus the benchmark corpus and the before / after repo — each number accurate in isolation, but together sounded inconsistent unless a reader did their own archaeology.
This page is that map. If something here still leaves a question, run nerviq feedback and tell us.
A note on repo ownership: the proof repos linked above currently live under DnaFin (founder account), not under the nerviqGitHub organization. That's an explicit choice — see MEMO-09 for the rationale and the explicit triggers for re-evaluation.