guardrails

Router skill. Resolve topic, then Read EXACTLY ONE playbook from routes:. [EXPLICIT]

When to use

A request targets a deterministic guard: block a command, validate a tool result, filter a prompt, enforce an output contract, sanitize secrets, check a quality gate or constitution principle, validate the integrity chain / stop conditions. Triggers: /jm:verify, /jm:advance, gate/PR readiness, pre-delivery. [INFERENCE]

Inputs → Outputs

In: topic (enum, required) + depth (quick|deep). Out: that playbook's verdict — pass/fail/block with evidence-tagged reasons + remediation; never green-by-default. [DOC]

Routing

1. Map the request to one topic. Disambiguate from routes.json desc if two plausibly fit; if still ambiguous, ask — do NOT load several. [INFERENCE]
2. Read that playbook only. Never load the whole cluster (12 files). [EXPLICIT]
3. deep → apply exhaustively, verify each step (Discover → Analyze → Execute → Validate); quick → essentials. Gates: constitution v6.0.0, evidence tags, script-first. [CONFIG]

Evidence tags

Alfa core family only: [EXPLICIT] [CODE] [CONFIG] [DOC] [INFERENCE] [ASSUMPTION]. One tag per claim; never mix families; canon in ../../references/verification-tags.md. [DOC]

Acceptance criteria

• Exactly one playbook read; topic ∈ enum; verdict carries evidence tags. [DOC]
• Script-first (prefer the hook/script the playbook names) and fail-closed: missing evidence ⇒ fail/block, never pass. [CONFIG]
• Verdict quality bar in assets/quality-rubric.json; pre-emit checks in assets/checklist.md. [CONFIG]

Anti-patterns

Loading multiple playbooks "to be safe"; guessing topic when ambiguous; reimplementing guard logic the hooks/scripts already enforce; reporting success without running the gate. [DOC]

Self-correction

Two files opened, an untagged verdict, or "pass" with unmet criteria → stop; re-resolve topic, re-run the named script, re-tag. [INFERENCE]