HEALTH AND LEARNING SHOULD BE PROVEN, NOT ASSUMED
Can the system detect
its own failure?
HEARTBEAT is being built to verify health, expose uncertainty, detect degradation, confirm recovery and learn through bounded experiments — without confusing a confident model, a clean data trail or a green dashboard with proof.
- Public interactive challenge intentionally disabled for the stability release.
- No internal SHANX telemetry, credentials or operational controls are exposed.
Observe
Read expected services, dependencies, freshness and contradictory evidence.
Degrade honestly
Use HEALTHY, DEGRADED, UNKNOWN and conflicting states instead of forcing false certainty.
Prove detection
Introduce a controlled safe failure and verify that the health layer actually notices it.
Verify recovery
Do not return to HEALTHY until restoration itself has been independently checked.
HEARTBEAT × HELM — public boundary · 30 Aug 2026: proving that a system is healthy does not grant an AI agent permission to act. HEARTBEAT challenges health, degradation, uncertainty and recovery evidence. SHANX HELM governs whether an agent is allowed to execute a bounded action.
Separation of responsibility: HEARTBEAT can report HEALTHY, DEGRADED, UNKNOWN or recovery evidence. HELM must still perform its own identity, capability, authority, policy and approval checks before any material execution. A green health state is never an authorization token.
Failure rule: if HELM depends on a service whose health evidence is DEGRADED, UNKNOWN or stale, that health signal may justify a safer policy outcome — but HEARTBEAT itself does not issue authority, approve actions or execute them.
HEARTBEAT
Observe health, preserve uncertainty, detect degradation, test recovery and keep evidence honest.
HELM
Govern AI-agent execution through explicit identity, capability, authority, policy, approval and evidence boundaries.
Health ≠ authority
A healthy system may still be forbidden to act. Permission is a separate decision from system health.
Evidence can inform policy
Verified health evidence may inform HELM policy decisions without becoming authority by itself.
Learning evidence 01 — Provenance ≠ Truth: HEARTBEAT has experimentally separated source integrity from evidence support. Knowing exactly where a claim came from does not make the claim true.
ML experiment 01 — Model Selection Harness: on synthetic data, HEARTBEAT rejected training-score vanity, detected an obvious target leak before fitting and kept model selection separate from final evaluation. Result: PASS WITH LIMITS.
Evidence first
Learning must leave a source, test, result, limitation and next question — not just a confident answer.
Attack the result
Overfitting, leakage, drift, stale evidence and misleading confidence are treated as failure modes to hunt deliberately.
No silent self-training
HEARTBEAT does not autonomously rewrite itself or promote experiments into production rules.
Zero execution authority
Learning, prediction and health status do not grant permission to act. Governed execution belongs to HELM.
Current boundary: HEARTBEAT remains experimental. Its public page shows research evidence and the public HELM separation only; internal telemetry, credentials, control paths and proprietary research machinery remain private.
Next learning attack: Drift Detector v0.1 — test meaningful distribution shift, false alarms on unchanged data and the difference between detecting drift and proving model failure.
Next engineering proof: when development resumes, test the worker-health adapter in staging, introduce one controlled safe dependency failure, prove DEGRADED, restore it and regression-test recovery before re-enabling the public challenge.