Critical-Period Crystallization: Phase-3 Plasticity Injection + EWC Residue Write-Protect

Implements: INV-074 (necessity), MECH-333 (open phase, option E subset), MECH-334 (closure / crystallization), ARC-075 (infant-curriculum plasticity magnitude asymmetry) Subject: learning.developmental_critical_period_closure_crystallization Status: IMPLEMENTED 2026-05-17 Registered: 2026-05-17 Depends on: ARC-062 Phase 1 (gated_policy), ARC-046 (infant curriculum GAP-9 scheduler), ARC-065 (diversity stack), ResidueField (SD-005) Blocks / enables: V3-EXQ-543h (ARC-062 GAP-B falsifier with crystallization ON); the resolution path for Q-052 (magnitude vs structural)

Problem

V3-EXQ-543g (2026-05-17, FAIL / weakens ARC-062) ran the canonical 543d 2x2 factorial (use_gated_policy x use_dacc) on the SP-CEM substrate with the GAP-B option-2 first-action one-hot head input. Result:

Arm use_gated use_dacc mean reef fraction
ARM_0 OFF OFF 0.278
ARM_1 OFF ON 0.174
ARM_2 ON OFF 0.444 (best)
ARM_3 ON ON 0.243 (regresses BELOW gated-only)

D2 (ARM_3 - ARM_2 >= +0.10) FAILED with delta = -0.200. The dACC training signal, added on top of the gated policy, destroys the discrimination ARM_2 established. This is the experimental signature of shared-parameter gradient interference: the gated policy is never crystallized before dACC is introduced, so dACC’s gradient overwrites the discrimination (the heterosynaptic-depression analog, O’Dell 2025, cited in INV-074’s literature anchors – the dominant pathway actively depresses competitors).

This is the substrate signature of the missing critical-period closure mechanism (MECH-334). The infant substrate (ARC-046 / GAP-9 scheduler) has four phases but implements ZERO weight crystallization – all weights stay fully plastic through every phase; phase transitions only change scalar params and env kwargs (ARC-075: temporal scheduling alone is architecturally incomplete; Bengio 2009).

Solution

Two coupled mechanisms, both behind one master toggle (REEConfig.crystallize_at_phase3, default False, bit-identical OFF), fired once at the infant-curriculum Phase 2->3 transition.

1. GatedPolicy plasticity injection (Nikishin et al. 2023 NeurIPS)

At Phase-3 entry, GatedPolicy.crystallize():

  • Sets requires_grad=False on head_0, head_1, discriminator params – the discrimination ARM_2 established during the open window is crystallized; routed diversity gradient can no longer overwrite it (MECH-334 closure analog for the policy heads).
  • Adds a fresh plastic expansion MLP with the SAME input contract as the scoring heads, last-Linear zero-initialised so the module’s output is bit-identical at the instant of crystallization (the expansion contributes exactly 0 until routed gradient grows it – no transition discontinuity).
  • forward() becomes frozen_gated(x) + expansion(x.detach()). The .detach() is essential: it blocks the expansion’s gradient from flowing back into the (frozen) heads’ upstream feature producers, in addition to the heads themselves being requires_grad=False.

expansion_parameters() exposes the plastic params for the experiment’s post-Phase-3 optimizer (mirrors the SD-033a GAP-D lateral_pfc.bias_head_parameters() pattern). crystallize() is idempotent; reset() does NOT un-crystallize (developmental closure persists across episodes).

Why option (E) and not (A)-(D) of MECH-333: the EXQ-571 audit showed F accounts for 88-89% of E3 scoring variance with all diversity bias_fractions at 0.000. A learning-rate schedule (A) or magnitude rescale (B) does not stop the structural overwrite once F dominates; gradient blocking (C) is the strongest isolation but plasticity injection (E) achieves the same isolation (a clean plastic gradient channel that cannot reach the consolidated weights) while preserving the consolidated knowledge – and it is the most directly validated technique for this exact RL failure mode (Nikishin 2023).

2. ResidueField EWC write-protect (Kirkpatrick et al. 2017)

ResidueField.snapshot_ewc_anchor() captures a detached copy of the rbf_field centers/weights at Phase 3 plus an established-basin diagonal Fisher proxy:

Fisher_i = |weights_anchor_i| * active_mask_i

ResidueField.ewc_penalty() returns ewc_lambda * sum_i Fisher_i * [ (w_i - w_anchor_i)^2 + sum_d (c_i,d - c_anchor_i,d)^2 ], differentiable w.r.t. the current params (anchor + Fisher detached). The experiment adds this to its loss.

This is not a hard freeze: residue geometry keeps adapting, but established basins resist overwriting in proportion to their accumulated strength – the biologically faithful MECH-334 reading (“high resistance to overwriting established basins”), not a uniform L2 pull. Returns a 0.0 scalar when disabled / lambda 0 / no anchor, so the experiment adds it unconditionally.

Phase-3 hook plumbing

InfantCurriculumScheduler.__init__(on_phase3_entry=...) takes a zero-arg callback fired EXACTLY ONCE the first tick the scheduler is in Phase 3 (guarded by _phase3_hook_fired). Kept as a caller-supplied callback so the experiment-harness helper stays ree_core-free; the experiment passes a closure calling agent.gated_policy.crystallize() + agent.residue_field.snapshot_ewc_anchor().

Pre-check: F-error dependence of routed diversity signals (ICM self-defeat)

Q-052’s literature anchor flags the Pathak et al. 2017 ICM self-defeat mode: an F-prediction-error-based diversity signal collapses to zero as F consolidates. If a routed signal is F-error-dependent, plasticity injection cannot rescue it – the signal is already gone before Phase 3 fires. Audit of the MECH-313 / MECH-314 / MECH-320 implementations (ree-v3/ree_core/policy/{noise_floor,structured_curiosity,tonic_vigor}.py, agent feed at agent.py):

Signal Source F-error-dependent? Behaviour by Phase 3 Route to expansion?
MECH-313 noise_floor max(baseline_T + alpha, min_T) constant temperature lift No constant (no decay) Yes – robust
MECH-314a novelty min RBF distance to active residue centers / mean-norm No (Wittmann 2008: striatal novelty independent of RPE, by design) tracks residue coverage, not F accuracy Yes – robust
MECH-314b uncertainty reads e3._running_variance (per-tick PE-MSE accumulator) Yes decays toward low/stable as F consolidates No – self-defeated
MECH-314c learning-progress EMA of |PE_t - PE_{t-K}|, fed e3._running_variance Yes – canonical Pathak 2017 collapses to ~0 as PE plateaus No – self-defeated
MECH-320 tonic_vigor primary = slow EWMA over realised E3-score-receipt (avg reward rate); secondary gate_pe reads _running_variance No for primary (Niv 2007 avg-reward-rate is reward-rate, not PE); only the non-load-bearing gate_pe modulator is F-adjacent primary robust (a competent agent keeps receiving good scores) Yes – primary robust
dACC / MECH-260 anti-recency per-candidate bias from recent action-class frequency No (state-dependent recency, not F-PE) robust Yes – robust; this is the 543g corrupting signal under test

Implication for the experiment. Routing MECH-314b/314c to the post-crystallization expansion layer is pointless: both signals will have self-defeated to approximately zero before Phase-3 crystallization fires (F consolidates through Phases 0-2; 314c is the textbook ICM self-defeat case; 314b decays with the same _running_variance). They cannot establish competitive weight on a fresh expansion layer because their gradient magnitude is already gone. The crystallization hypothesis is therefore tested with the F-robust signals – dACC/MECH-260 (the exact signal 543g showed corrupts the gated weights, and it is not F-error-dependent, so the hypothesis is cleanly testable), plus MECH-313 / MECH-314a / MECH-320-primary. V3-EXQ-543h holds MECH-314 in novelty-only configuration (314a ON, 314b/314c OFF) in the routed arms so the test is not diluted by a confounded null channel. This narrows but does not weaken the 543h test: if ARM_3 still regresses with crystallization ON and only F-robust signals routed, the problem is isolated to dACC’s own architecture, not gradient interference.

Architecture context

  • MECH-333 (open phase) is only partially implemented here: option (E) plasticity injection provides the post-window plastic channel. The pre-window F-attenuation (ARC-075 requirement 1: reduce F’s gradient contribution during Phases 0-1) is NOT implemented in this landing – it is the natural next substrate increment if 543h shows crystallization alone is insufficient. This landing is the MECH-334 closure half plus the plastic channel that closure presupposes.
  • Aton et al. 2013 two-stage gate: waking dynamics set gate 1 (consolidation instructions), sleep executes gate 2. Crystallization here operates on the waking-side policy weights; it does not replace MECH-120 (SWY) / MECH-165 (replay diversity), which are post-closure maintenance.
  • Relationship to SD-033a GAP-C/D: the crystallized discriminator’s gating_weight still routes to LateralPFCAnalog (GAP-C) unchanged; the expansion adds a parallel plastic bias, it does not alter the discriminator output contract.

What this enables

  • V3-EXQ-543h – the GAP-B falsifier re-run with crystallization as a third factor (2x2x2). Pre-registered prediction: with crystallization ON, ARM_3 (gated+dACC) >= ARM_2 (gated alone), i.e. D2 PASSes, because dACC gradient is sunk into the expansion layer rather than overwriting the crystallized heads.
  • Q-052 resolution path: 543h discriminates magnitude (A) vs structural (B). If crystallization rescues D2, the F-dominance failure is structural (B) and MECH-333/334 are the correct architecture. If ARM_3 still regresses with crystallization ON and F-robust signals routed, the defect is in dACC’s own architecture (a narrower diagnostic), not gradient interference.

INV-074 (plasticity crystallization necessity), MECH-333 (critical-period open phase – option E subset), MECH-334 (critical-period closure / crystallization – primary), ARC-075 (infant curriculum plasticity magnitude asymmetry), Q-052 (magnitude vs structural), ARC-062 GAP-B (the falsifier lineage this unblocks), ARC-065 (diversity stack: MECH-313/314/320), MECH-260 (dACC anti-recency: the 543g corrupting signal under test).

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