Policy artifacts — where the rules live

_{entity · created 2026-07-09 · actor claude-cowork · confidence high · status active} _{tags: policy egress tiers safety}

Now

Pointers + one-line summaries (do not duplicate instruction layer):

Tool egress matrix — ~/.hermes/policies/tool-egress-matrix.md (2026-05-16). 8 classes: 0 read-local → 7 money/security/destructive. Control point is which tools cross which boundary; class 6 (credential-touching) blocked by default, class 7 always stop-and-ask with fresh before/after state.
Engagement tiers — ~/.hermes/ENGAGEMENT_OS.md: Tier 0-2 reversible local fixes auto-execute with verification; Tier 3 asks first: money, trading execution, public/external sends, security-boundary changes, destructive filesystem, purchases. Report shape: verdict/evidence/action/verification, ≤6 emoji-led lines.
Money boundary — ~/.hermes/policies/nexus-money-agent-boundary-spec.md + commerce-action-classifier (md+py).
Room protocol — ~/.hermes/BOT_COORDINATION.md (canonical, synced to OpenClaw workspaces 06:50) + ENGAGEMENT_OS control-room contract (owner badges, one-owner-replies, provenance receipts Owner/Action/Evidence/Next).
Quality policies — ~/.openclaw/QUALITY_POLICY.md, ~/.hermes/QUALITY_POLICY.md; security docs ~/.openclaw/SECURITY.md.
Sandboxing — ~/.hermes/policies/grok-in-hermes-sandbox-design.md; sandboxes live under ~/.hermes/sandboxes/.

_{Sources: ~/.hermes/policies/tool-egress-matrix.md (2026-05-16) · ~/.hermes/ENGAGEMENT_OS.md}