Brain System Maturity Assessment

Last assessed: 2026-04-18 Assessor: Self-reviewed after the brain-keeper first-class agent sprint Overall maturity: ~70% (revised 2026-04-18, was ~80%) Status: Core functional, production-usable, self-observing. Not yet hardened against failure.

Executive verdict

The nervous system is ALIVE and auditable. The operator can dispatch brain-keeper from any interface (Telegram via a router agent → AgentiBridge → brain-keeper, or direct LiteLLM model: brain-keeper, or run_agent from another session) and receive a structured markdown + CSV report that auto-mirrors to Google Drive within 60 seconds as a native Google Doc + Google Sheet. The reports are readable from a phone, sortable chronologically by filename (brain-report-YYYY-MM-DD-HHMM.{md,csv}), and brain-keeper itself can audit its own environment and flag misconfigurations the operator made.

That closed loop — dispatch → audit → artifact → operator read → apply recommendations — is the 70%. The remaining 30% is polish, resilience, and self-flagged follow-ups.

Breakdown by area

Area Weight Maturity Notes
Core pipeline (read/write/amygdala/replay/extraction) 25% 88% Blocks 1-5 shipped. Nuclear halt deferred (B3). First real /replay E2E never executed (B4). 5/6 planned signal sources live (missing: ArgoCD webhook, k8s informer).
Observability (OTel, dashboards, spans) 15% 100% 5-layer OTel pipeline shipped. 27-panel Grafana dashboard across 6 biological regions. brain.inject/delivery/marker_write spans in ClickHouse. Langfuse trace correlation.
Agent surface (brain-keeper, A2A, LiteLLM model, reports→Drive) 20% 92% brain-keeper first-class ops oracle. LiteLLM model + AgentiBridge dispatch. Drive reports (Doc+Sheet). Daily test+triage crons live in ops.yaml. Gap: Opus activation pending (values.yaml override).
Self-healing (cadence tuning, noise cleanup, I/O profiling) 15% 42% Detection loop running: reasoner_feedback.py (every 6h), heal.py 7-point audit (daily 05:53), brain-triage-daily (daily 03:31), brain-heal-daily, brain-feedback-hourly. Cadence tuning, noise purge, marker_write I/O profiling still pending.
Resilience (vault backup, outage tests, concurrent writes) 10% 15% brain-backup-daily cron live (02:43 UTC, S3 target). No chaos testing. No DLQ for brain-ops failures. No concurrent write stress test.
Advanced (profile broadcast, bare claude wrapper, /replay E2E, custom MCP unit) 15% 40% Profile activation broadcast designed but unshipped. Bare claude -p wrapper half-shipped. First real /replay E2E never executed. brain-keeper-tools-prod unit exists but not consumed.

Weighted total: 0.25×88 + 0.15×100 + 0.2×92 + 0.15×42 + 0.1×15 + 0.15×40 = 69.2% → ~70%

What the 70% means in practice

What works end-to-end today

  • Dispatch surface: LiteLLM POST /v1/chat/completions with model: brain-keeper OR AgentiBridge run_agent(brain-keeper-0, task).
  • Command surface: tick, test, triage, enrich <arc>, replay <arc> --to <target>, extract clusters, dashboard. Free-form prompts interpreted loosely.
  • Report pipeline: Markdown + CSV → artifact-store REST PUT → S3 (<your-artifacts-bucket>/miscellaneous/brain-keeper/) → drive-sync Lambda → Google Drive (misc/brain-keeper/) → auto-convert .md → Google Doc and .csv → Google Sheet.
  • Timestamped naming: brain-report-YYYY-MM-DD-HHMM.{md,csv} (UTC minute granularity). Chronologically sortable in Drive.
  • Self-audit: brain-keeper queries its own environment via ClickHouse, writes a 6-region report (Frontal/Amygdala/Broadcast/Hippocampus/Pineal/Hook Observability), emits PASS/WARN/FAIL per check, recommends fixes.
  • Cost per run: ~$0.15 (Sonnet 4.6, ~40 seconds wall time). Opus target would increase this 3-5×.
  • Observability: every brain action emits OTel spans → ClickHouse → Grafana. No more agent interrogation to verify anything.

Validated runs on 2026-04-13

  • brain-report-2026-04-13-2207 — first run, 4268 bytes markdown, 12 PASS / 1 WARN / 2 FAIL, uploaded S3
  • brain-report-2026-04-13-2320 — second run, 5069 bytes md + 2835 bytes csv, new naming convention, both formats uploaded to miscellaneous/brain-keeper/, drive-sync confirmed within ~60s

Known self-diagnosed issues (fragility)

Brain-keeper audits its own environment and reports issues. Single source of truth: KEEPER.md § Known Issues. Do not duplicate the list here.

Path to 100%

Quick wins (~70% → ~80%, low effort)

Action Effort Impact
Fix AGENT_MODE_MODEL: opus in values.yaml 1 line, 5 min Brain-keeper reasoning tier unlocked
Loosen cadence (BROADCAST_MAX_PER_PROMPT=2, MIN_INTERVAL_SEC=300) 2 lines + pod env, 10 min Throttle drops below 50% (F.11 fixed)
Daily brain-test cron (workflows/crons/brain-test-daily.yaml) 1 file, 15 min Drive report every morning, self-audit loop closes
Validate agenticore + agentibridge image rebuilds 2× CI watch, 10 min Live patches become durable (F.9 + AgentiBridge wait fix)

Medium effort (~80% → ~90%)

  • Profile marker_write NFS I/O, identify contention source, apply fix (F.12)
  • Purge noise arcs + auto-demote zero-edge hot arcs in brain_keeper.py (F.13, F.14)
  • First real /replay E2E against a reproducible arc + verify replayed_from edge creation
  • Bare claude -p wrapper for non-Agent-tool dispatched sessions
  • Consume brain-keeper-tools-prod custom unit (add entry to agentihooks-bundle/.claude/.mcp.json + key env)

Hard / operator input needed (~90% → ~100%)

  • Vault backup — operator decides destination (restic to second disk? rclone to S3? Drive as tertiary?)
  • Chaos testing: NFS outage, Redis outage, concurrent write race conditions
  • Arc count scaling beyond 200 arcs (measure quick-refresh + full tick latency)
  • Ship atomic-drifting-catmull.md profile activation broadcast plan

Cross-references