MECH-111 per-candidate E1-prediction-error novelty (PARKED design)

Status: PARKED — design-complete, NOT implemented. Do not treat as a live substrate. No code, no config, no contracts, no claims change exist for this. Authored: 2026-06-09T22:06Z Maps to: MECH-111 (E1-prediction-error novelty drive) rendered per-candidate; equivalently the deferred MECH-314b Phase-2 refinement (“E1 LSTM forward variance over candidates”, structured_curiosity.py:24-25).

Why this is parked (read first)

This design was worked out for infant_substrate:GAP-13 (V3-EXQ-590 novelty Goldilocks calibration), then stood down after the 2026-06-09 GAP-13 re-adjudication (REE_assembly commit 1906f95a0d) established that:

  1. A per-candidate novelty routing channel already landed 2026-06-07 (MECH-314a Phase-2, curiosity_candidate_source=e2_world_forward), and V3-EXQ-648a proved it carries variance into E3 (C2 load-bearing PASS; consumed candidate z_world spread 0.149; curiosity bias range 0.0206). So “the per-candidate channel does not carry variance” is no longer true.
  2. The real residual for GAP-13 is selection authority — whether a per-candidate-varying bias actually moves the committed argmin — which is owned by the shared behavioural-diversity frontier (modulatory-bias-selection -authority / V3-EXQ-643a, MECH-341 / V3-EXQ-660, V3-EXQ-604c), NOT by adding another novelty channel.

This mechanism (E1-prediction-error novelty) is a scientifically distinct signal from the existing spatial-distance channel (MECH-314a, Wittmann-2008 striatal novelty = distance to visitation/residue centres). It is the faithful rendering of MECH-111’s actual claim — novelty = E1 prediction error. But:

  • It is not what GAP-13 or its unblocked claims (DEV-NEED-003, MECH-314) need — those are served by the existing channel.
  • It would hit the same selection-authority wall: a varying E1-PE bias still has to win the argmin against primary harm/goal scores, exactly what V3-EXQ -643a gates. Until that frontier lands contributory, this channel is as selection-inert as any other modulatory bias, so it would be non-validatable.
  • MECH-111 is currently a dead, non-priority claim (the broadcast branch was deleted 2026-05-25; novelty_bonus_weight is a no-op knob).

Resume trigger: revisit only if (a) MECH-111 or MECH-314b Phase-2 becomes a governance priority in its own right, AND (b) the selection-authority frontier (V3-EXQ-643a + 604c + 660) has landed contributory so a validation experiment is runnable rather than vacuous.

Problem this would solve (if ever live)

MECH-111’s claim is novelty = E1 prediction error. Its only V3 instantiation was a broadcast scalar E3Selector._novelty_ema (EMA of E1 prediction-error MSE) subtracted uniformly from every candidate score — which is argmin -invariant (a uniform shift never changes selection) and was deleted 2026-05-25 (evidence/planning/v3_exq_571_root_cause_2026-05-25.md). The result: novelty_bonus_weight is a dead knob, and V3-EXQ-590’s 5-arm sweep produced byte-identical coverage across all arms.

The existing per-candidate channel (MECH-314a) carries spatial-distance novelty (distance of each candidate’s predicted z_world to the visitation buffer / active residue centres). It does not carry the E1-prediction-error signal that defines MECH-111. This design adds that signal per-candidate.

Core mechanism

A small learned forward-variance head nu on E1DeepPredictor that predicts how surprising E1 will find a state, evaluated per candidate:

nu : z_world -> scalar   (expected E1 prediction error at that state)

Why a learned head rather than re-running E1 over each candidate rollout: a hypothetical candidate rollout has no ground-truth next state, so the true prediction error is uncomputable at selection time. nu(z_world) is the standard predictive surrogate (Pathak 2017 ICM forward-error; Daw 2006 frontopolar uncertainty) and is O(1) per candidate vs an LSTM re-roll.

Design (implement-substrate Step 3 form)

Config (all no-op defaults; gated on novelty_bonus_weight > 0 AND use_e1_novelty_head)

Param Type Default Purpose
use_e1_novelty_head bool False master; builds nu head on E1
e1_novelty_head_hidden int 32 head width
e1_novelty_target_ema_alpha float 0.1 EMA-smoothing of the regression target
novelty_bonus_weight float 0.0 revived live per-candidate scale (the 590 sweep knob)
novelty_bonus_bias_scale float 0.1 clamp, mirrors curiosity_bias_scale

Data flow 

TRAIN (sense / training tick):
  realized E1 PE = compute_prediction_loss()   # [z_self, z_world] MSE; harm-excluded
  target_ema <- EMA(realized PE)
  L_nu = MSE( nu(z_world_prev.detach()), target_ema )   # detached input: no grad into encoder

SELECT (select_action):
  cand_summaries[K, world_dim] <- _candidate_world_summaries(...)   # e2_world_forward (ARC-065 GAP-A)
  per_cand_novelty[K] = nu(cand_summaries).detach()                # genuinely per-candidate
  mech111_bias[i] = -clamp( novelty_bonus_weight * per_cand_novelty[i], +/- bias_scale )
  dacc_score_bias += mech111_bias   -> e3.select(score_bias=)      # argmin shifts
  • The trainable nu head lives on E1DeepPredictor (it estimates E1’s own error; sits beside the existing schema_readout_head, e1_deep.py:170).
  • StructuredCuriosity stays pure-arithmetic: its 314b leg gains an optional per_candidate_uncertainty argument that, when supplied, replaces the broadcast e3._running_variance scalar with nu(cand_summaries) per-candidate. So the bias composition reuses the existing curiosity path; only the source of the 314b signal changes.

Backward compatibility

Master off / novelty_bonus_weight = 0.0 -> head not built, no channel -> bit-identical. The ~30 existing callers passing novelty_bonus_weight=0.0 are unaffected.

Phased training (REQUIRED — the correctness gate)

An untrained nu outputs noise -> per-candidate spread is noise -> re-vacuates exactly like the e2 / encoder cases.

  • P0: warm E1 + nu (regress realized E1 PE) until nu’s per-candidate spread over K exceeds a floor.
  • P1: freeze E1; nu reads frozen z_world; channel active.
  • Validation non-vacuity precondition: nu_per_candidate_range > floor AND cand_world_pairwise_dist > floor (requires SD-056-trained e2 + candidate_summary_source="e2_world_forward"), else substrate_not_ready_requeue. Same guard pattern as V3-EXQ-648a / 649.

MECH-094

Applies. nu training and candidate reads are waking-only; the bias compute takes simulation_mode and returns zeros under replay (mirrors 314a/b/c). Handled by call-site scoping + the simulation_mode argument.

GAP-4 stochastic-attractor safety (binding constraint from the closure node)

SD-048 interoceptive noise is the primary irreducible stochastic attractor; a forward-PE novelty signal will otherwise chase irreducible noise (Burda 2018 noisy-TV). Mitigation is structural: nu regresses the harm-excluded E1 prediction error — compute_prediction_loss (agent.py:6008) is MSE over the [z_self, z_world] sequence only (z_harm / z_harm_a are not in E1’s input), and z_world does not carry harm_obs_a (GAP-4 audit). So the autonomic-noise attractor cannot enter nu. Deliberately NOT e3._running_variance (which SD-048 inflates and which the broadcast 314b leg currently reads).

ML/AI engineering notes

  • Moving-target instability (BYOL / target-network lesson): the regression target (E1’s own PE) drifts as E1 learns -> EMA-smoothed target + .detach() on the z_world input.
  • Target = z_world / z_self prediction error, not the E2/E3 running variance — that departure is the whole point of the mechanism.

Validation (when unblocked)

V3-EXQ-590b: the same 5-arm Goldilocks over novelty_bonus_weight in {0.1..1.0}, on this channel — P0 trains e2 (SD-056) + E1 + nu; non-vacuity precondition guards re-vacuity; the sweep then measures whether different weights produce different mean_coverage / H_pos (the inverted-U Goldilocks point). Queue via /queue-experiment. NOTE: this is the E1-PE-channel variant; the re-adjudication’s sanctioned 590b runs on the spatial MECH-314 e2_world_forward channel — these are different experiments testing different novelty signals.

Relationship to existing substrate

  • MECH-314a (live): spatial-distance novelty (Wittmann 2008). Distance of candidate z_world to visitation/residue centres. Different signal.
  • MECH-314b Phase-1 (live): broadcast e3._running_variance scalar. This design is its deferred Phase-2 per-candidate refinement.
  • MECH-111 broadcast (dead): deleted 2026-05-25; argmin-invariant.
  • This design: the per-candidate, learned, E1-PE rendering — instantiates MECH-111 per-candidate and closes MECH-314b Phase-2 in one head.

References

  • evidence/planning/v3_exq_571_root_cause_2026-05-25.md — broadcast deletion + the E2-bottleneck root cause + follow-on #5 (always-populating visitation buffer).
  • evidence/planning/infant_substrate_plan.md — GAP-13 node + the 2026-06-09 re-adjudication (1906f95a0d).
  • ree-v3/CLAUDE.md — SD-056, ARC-065 GAP-A, MECH-314a Phase-2 amend (V3-EXQ-648a), MECH-314 / MECH-313 / MECH-320 entries.
  • Pathak et al. 2017 (ICM forward-error intrinsic motivation); Daw et al. 2006 (frontopolar exploration uncertainty); Burda et al. 2018 (noisy-TV).

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