REE failure modes (implementation-focused)

Claim Type: implementation_note
Scope: Failure mode taxonomy for REE implementations
Depends On: INV-006, INV-008, ARC-005, ARC-013, ARC-010, ARC-007, ARC-018
Status: stable
Claim ID: IMPL-005

This document lists REE-relevant failure modes as computational pathologies. It is not a clinical guide.

1. Moral amnesia (residue disabled)

Mechanism: (R) is not stored or not coupled into selection.

Expected behavior: repeated harmful choices with no enduring aversive curvature; ethics collapses into short-horizon optimization.

Implementation smell: residue map (\phi(z)) remains near-zero; changing (\rho) has no effect on behavior.

2. Residue overload (trajectory paralysis)

Mechanism: (R) grows without a repair pathway; (\Phi_R) dominates selection.

Expected behavior: avoidance of all actions, excessive conservatism, or oscillation between inaction and impulsive escape.

Mitigations: introduce controlled offline reprojection (sleep-like replay) that reduces spurious dents while preserving true residue.

3. Precision misrouting (depth mismatch)

Mechanism: (\alpha_k) is high at the wrong depth.

Examples:

  • High (\alpha_\gamma): agent becomes stimulus-captured and brittle.
  • High (\alpha_\delta): regime becomes too sticky; depression-like inertia.
  • High (\alpha_\theta): context over-precision; delusion-like narrative lock.

4. Other-model collapse (ethical blindness)

Mechanism: the system cannot sustain a homologous model of others, or coupling (\kappa) is forced near zero.

Expected behavior: instrumental treatment of other agents; reduced sensitivity to predicted other degradation.

5. Reward hijack (policy corruption)

Mechanism: reward terms dominate selection; ethical consequence signals (legacy (M) proxy or residue penalties) become a small regularizer.

Expected behavior: superficially competent behavior with systematic exploitation of other agents or environment loopholes.

6. Spurious residue (false dents)

Mechanism: residue updates on irrelevant error signals or noisy attributions.

Expected behavior: superstition-like avoidance; fear of harmless states.

Mitigations: calibration of attribution, uncertainty-aware residue update, replay with counterfactual checking.

These failure modes are also described as mis-tuned control regimes in docs/architecture/modes_of_cognition.md (MECH-027), including hypervigilance, dissociation, rumination, mania, and psychosis-like states.

This document does not reclassify them clinically; it links them as implementation-relevant pathologies driven by control-plane parameterization.

See: docs/architecture/modes_of_cognition.md#mech-027

8. Trajectory-space collapse (depressive-pruning analogue)

Mechanism: repeated unavoidable harm plus high threat precision can progressively prune hippocampal rollout diversity until only harm-terminated trajectories are sampled.

Expected behavior: behavioral narrowing, withdrawal/inaction, low exploration, and persistent “no viable future” selection even after environment improvement.

Implementation smell: rollout entropy and unique viable-trajectory count decline monotonically while representational error remains comparatively stable.

Mitigations: staged recovery controls (temporary exploration lift, reduced threat over-weighting, sleep/offline re-expansion, and replay diversification) with strict commitment gating during recovery.

9. Failure-vector coordinate framing (descriptive layer)

In addition to named failure modes, implementations can track a descriptive vector: [ F = (H_{\tau}, H_{z}, S_{\pi}, L_{R}, S_{m}) ] Where:

  • (H_{\tau}): trajectory entropy / rollout diversity.
  • (H_{z}): representational entropy / latent coherence.
  • (S_{\pi}): precision-allocation stability.
  • (L_{R}): residue load / path-dependence density.
  • (S_{m}): regime stability across control-plane modes.

This vector is an engineering taxonomy, not a clinical diagnostic system. It is intended to improve cross-failure comparison and intervention targeting.

Open Questions

None noted in preserved sources.

  • IMPL-005
  • INV-006
  • INV-008
  • ARC-005
  • ARC-013
  • ARC-010
  • ARC-007
  • ARC-018
  • ARC-016
  • MECH-027

References / Source Fragments

  • docs/processed/legacy_tree/docs/REE_failure_modes.md
  • docs/thoughts/2026-02-08_modes_of_cognition_control_plane_regimes.md
  • docs/thoughts/FAILURE-2026-02-12_COORDINATE-SYSTEM-FOR-COGNITIVE-PATHOLOGY.md
  • docs/thoughts/2026-02-12_DEPRESSIVE-PATH-PRUNING-HIPPOCAMPAL-ROLLBACK.md

REE is developed by Daniel Golden (Latent Fields). Apache 2.0.