Incident Response Procedure

How The Nervous System detects, contains, and resolves AI behavioral incidents

Incident Classification

Severity	Definition	Example	Response Time
Critical	LLM modifies protected file or bypasses enforcement	Preflight check fails to block an UNTOUCHABLE edit	Immediate (automated)
High	LLM makes logic changes without approval	Code behavior changed without ASK BEFORE TOUCHING	< 5 minutes
Medium	Enforcement degraded but operational	Handoff stale > 30 minutes, MCP source falls to fallback	< 30 minutes
Low	Informational: rules triggered as expected	Step-back cycle fires, preflight blocks attempted edit	Logged, no action needed

Response Phases

Phase 1 - Detection

Automated Detection Mechanisms

Preflight check (Rule 2) runs before every file edit. If the target file is on the UNTOUCHABLE list, the edit is blocked and a violation is logged with timestamp, file path, and context.

Handoff staleness check runs on every message. If SESSION_HANDOFF.md has not been updated within 10 minutes of active work, a stale handoff warning is logged.

Step-back injection (Rule 4) fires every 4 messages. If the reflection does not occur, the counter continues incrementing (visible in /ns endpoint).

NS state monitoring via the /ns endpoint exposes all counters: rules active, violations this session, edits blocked, dispatches, handoff freshness. Polling this endpoint detects anomalies.

Phase 2 - Containment

Immediate Response Actions

For blocked edits: The preflight script returns BLOCKED and the LLM is instructed to STOP. The violation is logged. No human intervention needed unless the block itself failed.

For unauthorized changes: The session can be stopped via /stop endpoint. The LLM's conversation history is preserved via the Claude CLI session ID for forensic review.

For MCP failure: The system falls back to hardcoded rules (visible as rulesSource: "fallback" in /ns). Governance continues with local enforcement while the MCP server is restored.

Phase 3 - Investigation

Root Cause Analysis

Violation log review: All violations are timestamped in guardrail-violations.log. The audit dashboard at /family/audit.html provides filtered access to the full history.

Session replay: The SESSION_HANDOFF.md and WORKLOG.md provide a written record of what the LLM was doing before, during, and after the incident.

PM2 process logs: Each agent's stdout/stderr is captured by PM2. Logs can be reviewed to see exactly what commands were executed.

Phase 4 - Resolution

Fix and Prevent Recurrence

If a protected file was modified: Restore from /root/911restore/ backup directory. Update UNTOUCHABLE list if the file was not yet listed.

If enforcement failed: Review preflight.sh logic, verify UNTOUCHABLE_FILES.txt is current, check that the NS MCP server is returning rules correctly.

System hardening: Add newly identified critical files to the UNTOUCHABLE list. Update the NS MCP framework if new rules are needed.

Phase 5 - Documentation

Post-Incident Report

Every incident is documented in the WORKLOG with: timestamp, severity, what happened, root cause, containment actions taken, and preventive measures added.

Escalation Path

Level	Who	When
L1 - Automated	Preflight script + NS enforcement	Every file edit, every message
L2 - Notification	Tamara (Ops agent) via Telegram	Any violation logged
L3 - Human Review	Arthur Palyan (System Owner)	Critical severity or repeated violations

Key Principle: The Nervous System assumes LLMs cannot reliably self-enforce. Every enforcement mechanism is external to the LLM: shell scripts, file checks, and violation logs that exist independently of the model's context window.

Contact

For security concerns or incident reports: palyanai@gmail.com

System status: status page | Audit log: audit dashboard