You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
multiqlti conformance work (Appendix A of the ADR — all policy/config level, engine unchanged):
I1 — task class field + auto-escalation: runs carry class (R0 default). Attaching a workspace commit, sandbox execution with external effects, or a deployment target escalates R0 → A/B/C/E automatically. A run must not silently stay R0 after gaining a workspace.
I3 — quality-gated DoD default (class ≥ A): default pipeline template wires code_review.approved == true and fact_check.verdict != fail as blocking guardrails/DAG conditions (mechanism exists: guardrail-runner llm_check/json_schema, DAG condition evaluator). R0 runs keep advisory mode. Needs a migration note: some runs that used to reach completed will now fail — intended.
I4 — verifier family preset: config preset pinning testing / code_review / fact_check to a different model family than development; record the attestation in run metadata. Consensus engine remains the stronger optional form.
I6 — OutcomeEvent emitter: one structured event per run at terminal state (task_id, factory, class, tier, lifecycle, verdict, gate, verifier_families, landed_at, cost_usd). Schema will be published versioned in genai-enablement — track that follow-up.
I8 — escape join: incident/revert signals joinable to run.id (source independent of the verifying stage).
Platform ADR-0003 (unified SDLC standard) is accepted: https://github.com/100rd/genai-enablement/blob/main/docs/decisions/0003-unified-sdlc-standard.md
multiqlti conformance work (Appendix A of the ADR — all policy/config level, engine unchanged):
class(R0 default). Attaching a workspace commit, sandbox execution with external effects, or a deployment target escalates R0 → A/B/C/E automatically. A run must not silently stay R0 after gaining a workspace.code_review.approved == trueandfact_check.verdict != failas blocking guardrails/DAG conditions (mechanism exists: guardrail-runnerllm_check/json_schema, DAG condition evaluator). R0 runs keep advisory mode. Needs a migration note: some runs that used to reachcompletedwill now fail — intended.testing/code_review/fact_checkto a different model family thandevelopment; record the attestation in run metadata. Consensus engine remains the stronger optional form.task_id,factory,class,tier,lifecycle,verdict,gate,verifier_families,landed_at,cost_usd). Schema will be published versioned in genai-enablement — track that follow-up.run.id(source independent of the verifying stage).Reference vocabulary: https://github.com/100rd/genai-enablement/blob/main/docs/platform-glossary.md