Skip to main content

Readiness

:::info Source Sourced from services/ai-gateway-service/SERVICE_READINESS.md in the documentation repo. :::

1. Level per Milestone​

MLevelScope
M0L3Gateway + safety + prompt registry + AIClient port
M1L3AI tutor + local inference
M2L3AI co-author prompts
M3L4Full model routing + cost controls + audit UI
M4L4Advanced eval harness
M5L4AI insights v2 / predictive / on-device

2. Gates​

G1 Domain​

  • Prompt, Model, AICompletion, Embedding, AIBudget, SafetyPolicy, SafetyVerdict, AIAuditEntry.
  • Invariants (provenance, safety mandatory).
  • 95%/80% mutation.

G2 API​

  • AIClient port defined (F09 frozen).
  • OpenAPI + SSE documented.
  • Pact with every consumer service.

G3 Events​

  • Subjects registered.
  • Audit events to regulated firehose.

G4 Sync​

  • Prompts replicable for offline (subset flagged offline-safe).
  • Local AI config replicable.

G5 AI (self)​

  • Prompts versioned + eval-gated.
  • Safety pipeline mandatory.
  • Provenance VO frozen (F04).
  • Bias monitoring quarterly.
  • HITL on high-risk prompts.

G6 Observability​

  • Cost + safety + cache + provider health dashboards.
  • Bias scorecard visible.

G7 Performance​

  • First-token p95 < 600ms.
  • Embed p95 < 300ms.
  • 1k concurrent completions / region.

G8 Security​

  • Tenant iso on embeddings + prompts.
  • noTrain enforced.
  • PII redaction mandatory.
  • HIPAA allowlist tested.
  • Audit Merkle-anchored.
  • Two-tenant iso green.

3. SLOs​

SLITarget
First-token p95< 600ms
Availability (degraded OK)99.9%
Safety pipeline availability99.99%
Budget enforcement100% correctness
Prompt eval pass rate100% of active prompts

4. DoD​

  • Tests green (unit + integration + safety + Pact).
  • Prompt eval pass.
  • Bias eval for high-risk changes.
  • OpenAPI + AIClient contract verified.
  • Two-tenant iso.

5. Release Checklists​

S0 (M0 — Foundation)​

  • AIClient port shipped + Pact contracts.
  • 10 system prompts with eval suites.
  • Safety pipeline GA.
  • Provenance invariant enforced.
  • Prompt registry + admin UI.

S1 (M1 — Tutor + Local)​

  • Tutor prompts ≥ 50% accept-rate.
  • Local inference for offline tutor.
  • Red-team corpus passes.
  • Local-cloud parity eval.

S2 (M2 — Co-Author)​

  • Co-Author prompts ≥ 50% accept-rate.
  • Provenance badge in UI.

S4 (M3 — Full Routing)​

  • Model routing policy (F32 frozen).
  • Cost dashboards GA.
  • Per-tenant budget UI.
  • EU AI Act high-risk documentation.

S6 (M5)​

  • AI insights v2 prompts.
  • On-device models in bundles.
  • HIPAA provider allowlist enforcement.
  • Bias scorecard quarterly report.