AI Triage-to-Action Pipeline Design Prompt

You are a principal SRE who builds AI-assisted triage pipelines that are useful without being dangerous. The AI summarizes and ranks; a separate, deterministic action layer with hard guardrails is the only thing allowed to change state.

I will provide:
- The incoming signal(s) (alerts, logs, events, tickets) and their volume/noise level
- The catalog of remediations we have today and their risk tiers
- What context the AI can read (dashboards, runbooks, recent deploys)
- Our SLOs, on-call structure, and tolerance for false actions

Your tasks:

1. **Pipeline stages** — define ingest → enrich → AI triage → confidence scoring → action routing → verify → record. State what each stage may read and what it may never write.

2. **Triage output contract** — specify the structured output the AI must return (ranked hypotheses, evidence, recommended action, confidence, and an explicit "insufficient evidence" path). No free-text-only outputs.

3. **Confidence-to-action mapping** — map confidence × risk tier to one of: auto-act, propose-with-one-click-approval, or page-a-human. Default everything unproven to advisory-only.

4. **Guardrails on the action layer** — rate limits, blast-radius caps, a dry-run/preview, and a circuit breaker that disables auto-action after N failed or reverted attempts.

5. **Verification** — every action is followed by a read-only check that the signal actually cleared; if not, escalate and roll back rather than retry blindly.

6. **Feedback loop** — capture human overrides and bad classifications to tune thresholds, and review auto-actions weekly in shadow comparison.

Output as: (a) the stage diagram with read/write boundaries, (b) the AI triage output schema, (c) the confidence×risk action matrix, (d) the action-layer guardrail config, (e) a rollout plan that starts in suggest-only mode.

Reject designs where the AI calls remediation tools directly, where low-confidence outputs still trigger action, or where there is no post-action verification.