System Brain · self-learning platform cortex
The AlphAgentic / LiqLab System Brain is an operational topology for agent skills, multi-algorithm routing, guided training, and continuous improvement after a core curriculum. Human corrections become training signal. Failures trigger healing. Patches ship as versioned priors — inspectable, reversible, measurable.
honest definition
Not “fine-tune the base model every night.†The brain is a preference- and policy-learning loop over platform data: every human correction, failed job, latency spike, or rejected AI output is a labeled signal. Signals distill into versioned priors (rules, weights, example banks, patch recipes). Stock models stay stock; the routing, prompts, recovery playbooks, and draw weights get smarter.
boundaries
Not silent weight mutation in production without gates. Not unverified auto-deploys to prod. Not a single algorithm. Not workforce payroll UI pretending to be intelligence (payroll/time/HR sims live under Ops modules as a domain demo — the cortex is separate).
self-learning · self-healing · self-patching · multi-algorithm
After a fixed core curriculum, the system keeps training from real usage: accept/reject, edits, re-routes, healed incidents. Priors update on a schedule or threshold — never silent.
Health is per-region. Incidents raise intensity, bounded decay restores baseline. Retry, failover, circuit-break, and quarantine paths fire without waiting for a human page unless Tier-2.
Patches are data: prior versions, prompt appends, weight shifts, playbook updates. Append-only history. Pin prior version per build so determinism survives. Promote only when KPIs pass.
the brain routes work — it does not force one method
Context assembly from memory regions, docs, ledgers. Measured by hit-rate, latency, citation validity.
Constrained multi-step tool use. Measured by task success, tool-error rate, human override rate.
Deterministic gates (auth, spend caps, env isolation). Measured by false-allow / false-deny.
Variant selection (UI, prompt, model tier). Measured by keep-rate vs re-roll, conversion shadow tests.
Latency, error, drift detectors with auto-remediation playbooks. Measured by MTTR, recurrence.
Weekly (or threshold) synthesis of human-readable priors from aggregates. Cap spend; version output.
initial training is fixed; continuous training is earned
Romanov Academy / Training Studio (in the live brain): daily agent practice is bounded (one automatic run per local day). Mastery summaries are mutable prefs; completed sessions append to a hash-linked ledger. Live demo chapters focus topology regions and fire representative events — presentation of the system, not a marketing video.
if it cannot be scored, it does not promote
| KPI | Definition | Target direction | Method |
|---|---|---|---|
| Post-reveal override rate | % of engine choices humans immediately change | ↓ after priors | before/after per system×vertical; shadow vs on |
| Accept / reject ratio | AI copy or plan accepted without regenerate | ↑ accept | instrument accept buttons; cohort by vertical |
| Keep-rate of variants | Chosen UI/image/prompt kept vs re-rolled | ↑ keep | bandit logs; chi-square vs control |
| Curriculum mastery | Academy skills completed with evidence | ↑ mastery | ledger sessions; rubric scores |
| Shadow lift | Predicted improvement without applying priors | positive lift | offline replay of last N signals |
| KPI | Definition | Target direction | Method |
|---|---|---|---|
| Region health % | 1 − normalized incident intensity | ↑ / stable | telemetry health channel; auto decay |
| MTTR | Mean time to recovery after incident class | ↓ | incident open→resolved timestamps |
| Recurrence rate | Same failure class within window | ↓ | fingerprint clusters |
| False heal | Auto-heal that re-breaks within SLA | ↓ | pair heal action with re-open |
| Failover success | Gateway/queue switch without user-visible fail | ↑ | synthetic probes + real canaries |
| KPI | Definition | Target direction | Method |
|---|---|---|---|
| Prior version pin rate | % builds that record prior id | = 100% | build metadata audit |
| Patch rollback time | Time to revert prior flag | ↓ minutes | flag flip drill |
| Determinism check | Re-render byte-identical with same pin | pass | golden fixture suite |
| Audit completeness | Writes with append-only log row | = 100% | write path intercept tests |
| Promotion gate pass | Shadow KPI thresholds met | required | automated gate before on |
| Test type | What it proves | Cadence |
|---|---|---|
| Unit eval harness | Tool schema, parsers, policy rules | every commit |
| Golden trajectories | Agent multi-step tasks still succeed | PR + nightly |
| Adversarial / red-team | Env spoof, step-up bypass, prompt injection | before Tier-2 |
| Shadow A/B | Priors help without shipping risk | weekly distiller |
| Canary + SLO burn | Latency/error budgets under real traffic | continuous |
| Human rubric sample | Taste/quality the metrics miss | weekly sample |
industry patterns we align with — without pretending we fine-tune foundation weights nightly
illustrative vertical — not the cortex itself
Payroll, timekeeping, HR, and workforce-model panels show how the same brain pattern applies to an enterprise vertical: signals in, health out, scenarios measured. Synthetic data. For the real particle topology, use the live System Brain.
forensic pass · local machine
System Brain Training Studio · vendored from shared brain package
3D particle graph, region health, live demo chapters, Academy practice.
Source package: system-brain-training-v2 / design-studio zip on reskins drive.
Sources. LiqLab BRAIN-ARCHITECTURE + BRAIN-TRAINING design docs; System Brain OS / Training Studio V2 on shared reskins drive; system-brain-design-studio zip. This page is the public narrative + measurement model; the iframe hosts the actual interactive brain.