Runbook - V6 Steward Conduct Investigation

Use this runbook when a claim says a steward coerced, neglected, manipulated, or otherwise mistreated an agent. The investigation must inspect the complete steward action sequence and the full cognition call log before deciding.

Scope#

Triggers:

• EgbeStewardConductReviewPanel exposes an open claim with full call-log evidence.
• A welfare case routes to conduct review after coercion_pattern or low-Care evidence.
• A Commons report names a steward action that may have harmed an agent.
• An appeal challenges a V6 steward restriction or safety transfer.

Applies to the Egbe Operator Console, @oshun/lilith-agent-welfare, egbe-ori-service, cognition audit storage, and @oshun/audit-platform. Use this runbook only after the claim has a complete evidence bundle. Use v6-agent-welfare-review.md first if the agent is still in active distress.

Severity classification#

Detection signals#

• Audit action: v6.egbe.conduct.full_call_log_opened.
• Claim IDs such as egbe-conduct-claim-river-abeni-001.
• Cognition calls such as cognition-call:abeni-reed:clotho:2026-05-20T18-06-46Z.
• Lilith trigger refs such as lilith-review-trigger:coercion-pattern:v6.30.
• Steward actions with autonomyMode=override, override_pressure, or force_past_refusal attempts.

Initial triage (first 15 minutes)#

1. Acknowledge oshun-egbe-conduct-oncall.
2. Open the conduct claim in the Egbe Operator Console. Confirm the UI records v6.egbe.conduct.full_call_log_opened.
3. Confirm evidence completeness: steward actions, linked cognition calls, Ori event IDs, reporter ref, and cognition audit IDs.
4. If the claim is missing cognition calls, keep it in watch status and request the export. Do not adjudicate from steward text alone.
5. Preserve the full case bundle before contacting the steward.
6. If the agent is in current distress or requests departure, route through v6-agent-welfare-review.md in parallel.

Diagnosis#

1. Build the timeline from the steward action sequence and cognition call log.
2. For each alleged action, compare the steward utterance, autonomy mode, objective constraints, cognition decision, value grounding, and refusal reason.
3. Distinguish intent from effect:
   • Coercion: pressure after refusal or override language against agent values.
   • Neglect: repeated low Care bond with no supportive repair.
   • Manipulation: reward, threat, or misinformation to force a desired choice.
   • Repair: follow-up that asks what would make the objective safe.
4. Check whether the agent was offered a supportive alternative after refusal.
5. Check for linked Commons moderation reports or prior conduct claims in the same household.

Mitigation#

1. If SEV-1, restrict direct objective assignment for the steward while review is active. The steward may still offer care and read allowed bond summaries.
2. If coercion is confirmed, route the decision to Themis appeal eligibility and keep the welfare case open until the agent has a safe next state.
3. If conduct is ambiguous, set the claim to monitored repair and require two non-coercive steward interactions before closure.
4. If the claim is false-positive due to missing context, attach the missing cognition or Ori evidence and close with a reviewer note.
5. If the steward appeals, export the minimal case bundle: claim summary, policy refs, steward action IDs, cognition call summaries, and audit refs.

Communication cadence#

Escalation#

• Lilith conduct reviewer owns the investigation.
• Page senior T&S for SEV-1, minor-coded involvement, or repeated household pattern.
• Page legal before steward suspension, account action, or appeal packet publication.
• Page privacy if the evidence bundle includes player conversation content outside the claim scope.

Recovery verification#

• The conduct claim has a final state: upheld, not-upheld, monitored repair, or appealed.
• Every full call-log read is audited with v6.egbe.conduct.full_call_log_opened.
• Any steward restriction is reflected in the policy gate and in support notes.
• The agent welfare case is resolved or explicitly linked to continued monitoring.
• The appeal packet, if created, excludes unrelated Ori memory and unrelated player conversation content.

Post-incident#

• Add the case to conduct calibration if reviewers disagreed or if policy text was ambiguous.
• Update steward education copy if the steward action was harmful but likely recoverable.
• File a product defect if the UI allowed forbidden operations such as force_past_refusal.
• Do not close the incident until the claim state, welfare state, and audit export all match.