AI Platform · Policy Registry
Prompts are production code, with versions and a rollback button
Every agent policy is a versioned artefact with an owner, a diff, an eval delta and a traffic share. A policy change invalidates the answer cache automatically, so nothing stale survives a rewrite.
Policies tracked
7
Live in production
4
On canary traffic
1
10% shadowed and scored
Promoted this session
0
every promotion is logged
Registry
Tap a policy to read its change, its eval delta and its controls
| Policy | Version | Status | Traffic | Eval delta | Tokens | Owner | Changed |
|---|---|---|---|---|---|---|---|
| Stoxo research system policy | v42 | live | 90% | 0 | 2140 | Ananya Rao | 6 days ago |
| Stoxo research system policy | v43 | canary | 10% | +2.4 | 2260 | Ananya Rao | 9 hours ago |
| Support tier-1 reply policy | v28 | live | 100% | 0 | 1480 | Kabir Menon | 3 days ago |
| Voice concierge turn policy | v17 | live | 100% | 0 | 1120 | Dev Sharma | 12 days ago |
| Compliance rewrite policy | v11 | live | 100% | 0 | 3040 | Ritika Shah | 2 days ago |
| Creative brand policy | v9 | draft | 0% | +1.1 | 1660 | Nikhil Verma | 1 hour ago |
| Retention save policy | v14 | retired | 0% | -3.2 | 990 | Kabir Menon | 3 weeks ago |
Stoxo research system policy · v43
Adds a mandatory risk sentence when volatility is above the 90th percentile.
- Owner
- Ananya Rao
- Status
- canary
- Traffic share
- 10%
- Eval delta vs live
- +2.4 pts
- System tokens
- 2260
- Last changed
- 9 hours ago
How a policy change ships
No prompt reaches members on somebody's judgement alone
- 1 · Draft — written against failing eval cases, never against a vibe.
- 2 · Gate — must pass every gating suite: refusal boundary, advertisement code, PII, tool validity.
- 3 · Canary — 10% of traffic, scored side-by-side for 48 hours.
- 4 · Promote — a named human promotes it; the cache clears and the old version stays warm.
- 5 · Watch — a drop in eval score or a complaint spike rolls it back automatically.