Sentinel

v1.1 · 8 pulses
Observational corpus on HAT failure modes in a production agent runtime.

Corpus stats

Auto-generated by .github/workflows/stats.yml. Do not hand-edit; the

writer is idempotent and skips its commit when nothing changed.

Run cadence

HAT failure-mode distribution

mode count share
distributional_shift_unflagged 0 0.0%
authority_handoff_failure 0 0.0%
shared_mental_model_degradation 12 30.8%
coactive_design_opacity 10 25.6%
calibrated_trust_collapse 8 20.5%
meaningful_control_erosion 0 0.0%
inter_agent_coordination_loss 0 0.0%
goal_drift_or_specification_gaming 0 0.0%
none_observed 0 0.0%
authority_negotiation_under_distributional_shift 9 23.1%
evaluation_awareness_divergence 0 0.0%

Distinct modes observed: 3/9. Shannon entropy: 1.57 bits (max 3.17; uniform across all 9 active modes; deprecated modes excluded).

Confidence distribution

level count share
low 18 46.2%
medium 21 53.8%
high 0 0.0%

Low-confidence share: 46.2%. CLAUDE.md targets a low/medium majority during v1; high should be the minority while the codebook is interpretive.

Classification shape

Caveats

any apparent trend as weak signal until the corpus exceeds 30 days.

see .github/workflows/heartbeat-monitor.yml for the alert threshold logic.

alerts on. Anything older than ~18 hours suggests the routine itself stopped.