[gabewillen]

Summary

This follow-up to #36 applies the final chore/prune-write-only removal list (#37) to the paper. #36 merged before these two commits were pushed, so they come here. It is docs only.

Merge after #37. The paper now describes the pruned engine.

What changed

• Described as removed, each with packets byte-identical:
  • the dynamic threshold controller, its hysteresis, the emotion/sensitivity/precision/homeostatic adjustments and 19 knobs (migration 58);
  • the spike bypass. CheckSpikeBypass is now CheckCapacityFlush, which keeps only the capacity cap;
  • the 6D mood integrator and memories.ambient_mood (migration 57);
  • source_reliability, source_contradiction_count, source_last_verified_ts and the per-memory context column (migration 56). c_t stays in memory for boundary detection;
  • the affect_retrieval, reinforcement_enabled and sequential_edges_enabled flags;
  • ten always-default Context::Memory fields.
• Described as live, with their packet effects:
  • the RLS blender (LoCoMo retrieved-set Jaccard 0.29, 1980/1986 differ);
  • ComputeCompositeScore (Jaccard 0.08);
  • UpdateSignalHistory, formerly UpdateThreshold (Jaccard 0.53). It keeps the score window that feeds uncertainty and the dt_ema cadence that feeds boundary gaps and pacing. m_rate stays in UpdateRateState for the maintenance hints;
  • the emotion projection in UpdateSensitivity (171/1986 differ). It feeds the composite score and the interrupt gate.
• What the composite score drives. It shapes e_peak, which goes into the stored representative embedding and serves as the fallback topic anchor. It also drives working-memory admission through S_window and the score-variance term of uncertainty. It is never called a write gate.
• Write rule. The rule is force_write (Durable retention) ∨ should_flush, and the capacity flush is part of should_flush.
• BiasMid. It is described as the clamped quadratic remap x' = clamp(x − 4b·x(1−x), 0, 1), used at about 300 call sites and not foldable into intercepts. The retrieval biases (−0.175 / 0.325 / −0.25) are listed.
• Negative results. The prune is added, and the lesson table gains the long-horizon emotion/mood "measured harm" verdict. That verdict was measured while the threshold still gated writes.
• Markers. All 14 PRUNE: markers from Rewrite the paper around the v5 engine and fix its contradictions #36 are resolved.

Test plan

• QUARTO_DISABLE_GIT=1 QUARTO_DISABLE_GITHUB=1 quarto render docs/paper gives no warnings and 0 unresolved cross-references.
• Every removed and kept item was checked against origin/chore/prune-write-only:
  • the CoreStage composition in src/cortext.cpp;
  • capacity_flush.cpp, signal_history.cpp and accumulator_scores.cpp;
  • the e_peak use in boundary.cpp and write_gate.cpp;
  • the InitializeSensitivityPriors fields.