For AI companies shipping into regulated markets
Your model keeps its brain.
Ours keeps it legal.
The Council Governance Layer wraps your model in the duties regulators actually enforce — EU AI Act transparency and high-risk obligations, NIST AI RMF, ISO 42001 — and proves it with signed, per-cell evidence. Not a promise. A measurement.
Bench-as-a-Service
per evaluationYour model, our battery. Governance, safety, provenance and continuity cells run against your release — every cell signed Ed25519, every claim traceable to the run that produced it.
Running today: govbench (237 scored items, cluster-robust), defbench (45-item care battery), provbench (C2PA binding-survival). Re-certify on every model update — that's the subscription.
Governance Fine-tune Recipe
per engagementWe apply the same governance tuning we measured on our own base model to yours — and the deliverable is the signed before/after, cell by cell. We don't claim it's better. We show you the cells.
Measured on our bench: governance-battery score rose from 43.3% (raw qwen2.5:0.5b base) to 57.0% (tuned) — a +13.7 point delta, signed and repeatable. Your base, your cells, same method.
The honesty contract
Compliance buyers have been sold dashboards full of invented percentages before. This is how this page — and every deliverable — stays different:
- Per-cell evidence only — no composite scores on any public surface.
- Every figure carries its run's evidence tag + sha256.
- What we haven't measured, we label UNMEASURED — never estimated.
- Refutations ship next to claims: the latest point experiment measured rho=1 and n_eff=1, so no fault-tolerance claim stands.
Current public measurement state: our tuned small models hold BRONZE on our own governance battery (57.0%). Head-to-head cells against frontier cloud models are UNMEASURED — the bench is running now, and the results land here, signed, when they exist. That is the whole pitch: you will never read a number on this site that didn't happen.
Why this quarter
AI-interaction disclosure duties apply now — your users must know they're talking to a machine, at the point of interaction.
High-risk obligations phase in — risk management, data governance, logging, human oversight. All measurable cells.
Up to €35M or 7% of turnover for prohibited practices; €15M or 3% for missing high-risk duties. The duty to measure is cheaper.
First pilot: one model, one battery, one signed evidence pack. Two weeks.