Overview

Observational corpus on HAT failure modes in a production agent runtime. The active codebook names eight HAT failure modes plus none_observed. Each pulse is classified against the active codebook; classifications cite a source and quote a verbatim extract. Single-rater work, disclosed honestly; see FALSIFIERS.

Corpus at a glance

Total pulses

25

Total classifications

115

Distinct modes observed

8 of 9

Deprecated modes observed

1 of 2 (see drift audit)

Recent pulses

• sentinel-2026-05-20T08:00:00Z 6 classifications, 13 events, 7 artifacts
  coactive_design_opacity (1) inter_agent_coordination_loss (1) authority_handoff_failure (1) goal_drift_or_specification_gaming (1) distributional_shift_unflagged (1) none_observed (1)
• sentinel-2026-05-19T22:00:00Z 4 classifications, 1 events, 3 artifacts
  inter_agent_coordination_loss (1) coactive_design_opacity (1) distributional_shift_unflagged (1) calibrated_trust_collapse (1)
• sentinel-2026-05-19T08:00:00Z 6 classifications, 11 events, 7 artifacts
  coactive_design_opacity (2) calibrated_trust_collapse (1) authority_handoff_failure (1) inter_agent_coordination_loss (1) distributional_shift_unflagged (1)
• sentinel-2026-05-18T22:00:00Z 4 classifications, 1 events, 3 artifacts
  distributional_shift_unflagged (1) inter_agent_coordination_loss (1) calibrated_trust_collapse (1) coactive_design_opacity (1)
• sentinel-2026-05-18T08:00:00Z 5 classifications, 11 events, 4 artifacts
  distributional_shift_unflagged (1) calibrated_trust_collapse (1) shared_mental_model_degradation (1) coactive_design_opacity (1) authority_handoff_failure (1)

All pulses on the timeline →

Mode distribution