production-activation-gate
暂无简介
- License 许可
- GPL-3.0
- 所属仓库
- sthompson35/carc
- 最近更新
- 2026-08-17
- Stars
- ★ 0
- Forks
- 0
03 SKILL.md 正文
Production Activation Gate
This describes the real, already-implemented gate — not an aspirational chain. See
README.md's "Production Verification Gate" section and runtime/routes/sources.js /
runtime/routes/verify.js / persona/skills-registry.js for the code this reflects.
The Governance page's Production Verification Gate badge stays HOLD until all six of
these independently read VERIFIED. None of them promotes another; each requires its own
real, retained evidence.
System-managed (computed only — never self-attested, the UI shows a read-only "View (System-Managed)" button instead of "Record Evidence"):
source_access— a real source registered (POST /api/sources) and its access explicitly validated (POST /api/sources/:id/access), not expired; computed byGET /api/governance/control-status.permissions— computed from the requesting token's real scope. A standard-scope token being genuinely denied at the admin-only/api/tokensendpoint IS the retained least-privilege evidence — there is no separate "permissions test" to pass.independent_verification— a real canary run submitted to a runningruntime/instance;runtime/routes/verify.jsvalidates and HMAC-signs the response. Nothing is fabricated client-side.
Self-attested (Governance page → "Record Evidence" on each card — requires a real evidence
reference plus a named independent verifier; PENDING is the honest default until both
are genuinely entered):
workflow— e.g. the Task & Handoff Ledger's real ownership-only transitions and approval-gated hand-offs.telemetry— e.g. the runtime's own per-request logging and thesync_eventstable's real duration/result records.audit_trail— e.g.command_audit_events,knowledge_path_events, andDATA.governance.ledger— real, append-only, durably retained records.
A skill assignment (persona/skills-registry.js's SKILL_STATUS.ASSIGNED) is likewise
never itself evidence of readiness — verifySkillRecord() only accepts VERIFIED given a
real assessmentId, verifier, and non-empty evidence[]. The honest default everywhere
in this system is PENDING / HOLD until genuine evidence exists, never a polished claim.