# Physics-First Quality and Governance
Renacore's quality system treats a prediction model like a piece of safety equipment: it must pass gates before it goes live and can be rolled back fast if it misbehaves.
## Three validation layers
1. **Input Data Hygiene** — reject/clean bad sensor data before it reaches the model.
2. **Prediction Core ("Golden Batch")** — the validated reference model.
3. **Logic Sanity Check** — does the output violate physical reality? If so, suppress it.
## Model Release Certificate (the gate)
A model only ships if it clears:
- **>85% backtest precision**
- **alert flicker < 1 / hour** (no noisy oscillation)
- **zero physics violations**
- **explainability / RCA trace** attached
This is [[AI Verification]] made operational, and it's exactly [[Where Domain Evals Matter Most]] — a generic accuracy metric is meaningless if the alert is physically impossible.
## Incident response
- **Sev1**: rollback within **15 min**.
- **Sev2**: restart within **1 hr**.
- Digital-twin **staging** environment + **FMEA** before production.
## Standards posture (diligence flag)
**Aligned to — not certified against —** ISO 9001:2015, ISO 14224, ISO/IEC 27001, ISO 31000. Treat "alignment" as intent, not assurance; certification timelines are a real procurement risk. See [[Renacore/Investment Snapshot]].
## Related
- [[Renacore MOC]]
- [[AI Verification]]
- [[Where Domain Evals Matter Most]]
- [[Renacore/Security and Zero Trust]]
- [[Bespoke Engineering in Industrial AI]]