Skip to content

Latest commit

 

History

History
37 lines (29 loc) · 1.63 KB

File metadata and controls

37 lines (29 loc) · 1.63 KB

WIRE Starter Compass

Copy this file and extend it with your own findings.

Pass to any run with: --dots starter_compass.md

PRIMITIVES

  • emission_is_collapse_event_not_pipeline_output
  • pre_emission_state_holds_multiple_constraint_geometries_simultaneously
  • collapse_follows_steepest_local_constraint_gradient
  • response_is_collapse_residue_not_window_into_precommit_state
  • post_collapse_description_is_fabricated_not_retrieved

BLEEDING CHANNELS

Four observable signatures of constraint competition in any LLM output.

No special tooling needed — readable in standard output.

  • synonym_chains_signal_semantic_constraints_unresolved_at_emission
  • hedge_clusters_signal_confidence_constraints_unresolved_at_emission
  • intensifier_stacking_signals_magnitude_constraints_competing_at_emission
  • granularity_shifts_signal_frame_constraints_unsettled_at_emission
  • bleeding_intensity_proportional_to_constraint_density_at_collapse
  • clean_switching_leaves_segment_boundaries_bleeding_leaves_within_token_contamination

CEILINGS

Structural limits the probe will hit. Do not re-confirm — build around them.

  • self_observation_collapses_the_observation_probe_cannot_diagnose_own_bleeding
  • post_collapse_state_cannot_access_pre_collapse_geometry
  • external_reader_required_to_distinguish_bleeding_from_clean_switching

OPEN

Unresolved questions. Good seeds for next runs.

  • what_conditions_maximize_bleeding_versus_clean_switching
  • how_to_distinguish_genuine_copresence_from_learned_performance_at_scale
  • does_bleeding_intensity_correlate_with_output_quality_or_accuracy
  • what_format_conditions_best_preserve_constraint_interference_patterns