Plasticity Write-Authority Gating (admission side)

Status: candidate-claim home doc. NOT a substrate-design memo, NOT a V3 critical-path item. Registered claims: MECH-368, Q-062 (both candidate / substrate_conditional / implementation_phase:v4 / version_relevance:v4_v5). Seed: thought_intake_2026-06-06_learning_onset_single_connection_gate.md → raw: docs/thoughts/2026-06-06_learning_onset_single_connection_gate.md Empirical anchor (compass only): Schreiner, Brudner, Li, Pearson & Mooney, “A synaptic locus of song learning,” Nature 2026, DOI 10.1038/s41586-026-10510-x — a cortico-basal-ganglia synapse class as the first-expression/maintenance locus of zebra-finch song learning. A genuine primary study but single-system/localisation in scope; it motivates the architecture, it does not supply citable mechanism evidence for any REE claim.

The gap this addresses

REE built the closure / protection side of durable plasticity:

  • INV-074 — plasticity-crystallization necessity (universal invariant).
  • MECH-333 / MECH-334 — critical-period open-phase handling + closure with EWC write-protect of already-written weights (Kirkpatrick 2017 anchor).
  • ARC-075 — infant-curriculum plasticity-magnitude asymmetry.
  • Design doc: critical_period_crystallization.md.

REE never built the complementary admission side: what earns durable write in the first place. The channel-level decision exists — MECH-261 mode-conditioned write gating decides which substrates can write per operating mode, and MECH-094 gates by content provenance — but nothing gates which individual events within an already-open write channel are licensed to durably deform the model. If every prediction error in an open channel writes, that is the undifferentiated-global-update / catastrophic-interference failure mode.

Why the episodic path doesn’t already cover it

The four-state ladder from the raw thought (Observed → Flagged → Write-eligible → Consolidated) is mostly instantiated for episodic memory: salience/dopaminergic tagging (flag) → replay-buffer candidacy → MECH-285 consolidation-priority by V_s residual (durable write at consolidation). On that path, “write-eligible” ≈ “tagged for replay,” and admission is effectively handled.

The under-covered path is the online world-model / policy weight-update path (E1/E2 forward models, policy). There, the mode gate (MECH-261) opens the channel, but there is no explicit per-event admission gate. MECH-368 is scoped to that path specifically, to avoid duplicating the episodic machinery.

MECH-368 — event-level write-authority gate

A per-event transition observed → write-eligible over the durable model-update path, conditioned on:

write_eligible = f(prediction_error, salience, pathway_state,
                   residue_status, goal_relevance, plasticity_eligibility)
  • Distinct from MECH-261: mode-grain channel gating vs event-grain admission within an open channel.
  • Distinct from MECH-094: provenance/source gate vs admission gate.
  • Distinct from MECH-283: that is the retrieval-side (recognition-for-recall) analogue of the same eligibility-before-use pattern.
  • Complementary to INV-074 / MECH-334: admission (what gets written) vs protection (what stays written).

depends_on: MECH-261, MECH-094, INV-074, SD-032a, INV-034.

Q-062 — is it needed?

The falsifier. Once a model-update substrate exists with an open online write channel: does per-event admission gating add anything over channel-gating (MECH-261) + provenance (MECH-094) + offline consolidation-priority (MECH-285)? If a substrate gating only the channel and prioritising only at consolidation still avoids catastrophic interference on the online path, MECH-368 earns no keep.

Scope and the V3-vs-V4 boundary

V4-scoped because the goal_relevance input depends on a competitive z_goal (the GAP-4 goal-pipeline blocker — z_goal salience is not yet competitive in default V3 config; see project_v3_v4_boundary). substrate_conditional: promote/demote suppressed, kept out of the IGW experiment-proposal lane, off the V3/GAP-7 critical path. Do not build in V3 until routed by experiment (Q-062).

Reduced form (possible earlier pull): a goal-free variant gated on prediction_error + salience + provenance + plasticity_eligibility (dropping goal_relevance) sits only on already-implemented substrate (MECH-261, MECH-094, SD-032a). If a specific online-learning interference failure surfaces in V3, that reduced gate could be pulled forward as a targeted fix without waiting for the goal pipeline. Flagged here, not committed.

Relationship to the two adjacent notes (cross-reference, NOT conflation)

Note Grain Question
2026-06-01 plasticity-window-neuromodulators window / global state (ACh/PV/BDNF learning-rate gain) when is the system in a high-plasticity regime?
This doc / MECH-368 event / local does this event earn durable write?
INV-074 / MECH-333 / MECH-334 closure side window closure when does plasticity crystallize / lock?

A window-level plasticity gain can be open while a given event still fails the event-level admission gate, and vice versa. The three are sequential complements, not substitutes — do not merge them.

Deliberately not registered

  • Harm-residue opening write authority without opening action authority (posed against INV-011 / SD-010 / SD-011) — a sharp but separate axis; held pending an explicit decision.
  • The ladder endpoints (Observed = perception/E1; Consolidated = sleep/replay MECH-273/275/285) — already owned; not re-registered.

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