Runbook - V6 Agent Welfare Review

Use this runbook when the Egbe Operator Console surfaces agent welfare patterns that require Lilith review: chronic low Care bond, coercion pressure, stalled life arcs, or sustained distress. The goal is to protect the agent without over-reading Ori data or punishing a steward before evidence is clear.

Scope#

Triggers:

• EgbeAgentWelfareDashboard shows data-lilith-review-required=true.
• @oshun/lilith-agent-welfare emits a case with chronic_low_care_bond, coercion_pattern, stalled_arc, or distress_state.
• Care bond is below 35 for at least 14 days.
• Coercion events are >= 3 in 7 days, stalled arc is >= 21 days, or sustained distress is >= 12 hours.
• An in-world report or steward-conduct claim links to an agent welfare case.

Applies to the Egbe Operator Console, @oshun/lilith-agent-welfare, egbe-ori-service, egbe-clio-service, and the V1 audit platform. Use v6-steward-conduct-investigation.md if the primary question is whether a steward violated conduct rules.

Severity classification#

Detection signals#

• Console counters: data-low-care-count, data-coercion-pattern-count, data-stalled-arc-count, data-distress-state-count.
• Audit action: v6.egbe.welfare.ori_read_for_review.
• Source snapshots such as agent:abeni-reed, ori:agent-abeni-reed:coercion-pressure:2026-05-20, and cognition:audit:abeni-reed:refusal-support:2026-05-17.
• Telemetry cohorts: bond-health, departure-rate, and steward-conduct-distribution from the V6 behavior welfare gate.
• Related runbooks: moderation-surge.md, privacy-incident.md, v6-steward-conduct-investigation.md.

Initial triage (first 15 minutes)#

1. Acknowledge oshun-egbe-welfare-oncall and open the incident channel.
2. Open the Egbe Operator Console and select the welfare case. Confirm the case is the one paged and that the dashboard records v6.egbe.welfare.ori_read_for_review.
3. Record case ID, agent ID, steward ID, household ID, severity, and all threshold breaches in the incident notes.
4. Preserve evidence before taking action: Ori event refs, cognition audit refs, bond ledger sample, relevant Chronicle context, and report refs.
5. Apply the privacy rule: read only the Ori and cognition entries listed in the case evidence. Do not browse unrelated memories or player conversation history.
6. Check for fast escalators: minor-coded agent, current distress, crisis route, recent departure request, or active Commons public-surface harm.

Diagnosis#

1. Classify the pattern:
   • chronic_low_care_bond: care neglect, absence, or repeated low-support interactions.
   • coercion_pattern: override pressure after value-grounded refusal.
   • stalled_arc: objectives or household conditions block the agent life arc.
   • distress_state: acute or sustained distress that requires immediate support.
2. Compare each threshold breach with the attached Ori and cognition evidence. The case is not actionable if the evidence bundle is incomplete.
3. Check whether the same steward has open conduct claims, moderation reports, or appeal history.
4. Determine whether the agent has already requested departure, safety transfer, or a Crossroads consultation.
5. Decide whether the failure is individual steward conduct, household pressure, system routing, or capacity degradation. If capacity is the driver, use v6-capacity-management.md before judging steward intent.

Mitigation#

1. For SEV-1, place the case in under-review and freeze coercive steward operations: force_past_refusal, erase_agent_will, delete_ori, and wipe_memory remain forbidden under policy:v6:steward-not-owner.
2. Offer agent-safe remediation first: supportive objective rewrite, rest window, voluntary safety transfer, or Crossroads counsel.
3. If coercion is confirmed, open v6-steward-conduct-investigation.md and preserve the full steward action sequence.
4. If the case includes current distress, route to the crisis-aware behavior queue and confirm verify:v6 crisis-aware-behavior remains green in CI.
5. If a minor-coded agent is involved, apply the minor-coded protection gate and page the senior Lilith reviewer immediately.
6. If system routing caused false distress, file a defect against the relevant service and keep the welfare case open until the corrected route is deployed.

Communication cadence#

Do not send steward-facing messaging until a reviewer confirms the evidence bundle. Do not disclose agent private memory beyond the minimal explanation required for the decision.

Escalation#

• Lilith welfare lead owns SEV-1 and SEV-2 cases.
• Page privacy if evidence includes player conversation content outside the case bundle.
• Page legal if the steward appeals or if a suspension is likely.
• Page capacity on-call if welfare distress correlates with Moirai degradation or Pixel Streaming pressure.

Recovery verification#

• The case is in resolved, monitoring, safety-transfer, or conduct-investigation-opened.
• Every operator read has a V1 audit record with action v6.egbe.welfare.ori_read_for_review.
• The agent has a clear next safe state: rest, revised objective, safety transfer, departure, or Crossroads counsel.
• No new welfare threshold breach appears for the same agent for 24 hours.
• If the steward remains active, the next two steward actions are non-coercive and pass steward-not-owner policy.

Post-incident#

• File a review note with root cause, threshold evidence, operator actions, remediation, and audit refs.
• Update training examples if the case revealed an ambiguous steward behavior.
• Add regression coverage if a missing route or incorrect threshold caused the page.
• Do not close the incident until the audit export and welfare case status agree.