MECH-189: Super-Ordinal Goal-Anchor ContextMemory Writes Substrate

Claim ID: MECH-189 Subject: development.super_ordinal_goal_formation Status: IMPLEMENTED (substrate) 2026-06-09 — v3_pending until the validation EXQ PASSes Registered (substrate doc): 2026-06-09 Closure node: infant_substrate:GAP-11 (EXQ-ISEF-002 transient-benefit-patch z_goal seeding) Depends on: INV-041, SD-016, INV-037, INV-038, MECH-117, MECH-112 Blocks / unblocks: DEV-NEED-006, DEV-NEED-024, the MECH-189 retest (V3-EXQ-588 successor with a NEW letter)

Problem

MECH-189 asserts: “During the child phase, high-salience benefit contacts under high contextual complexity are written to persistent ContextMemory as super-ordinal goal anchors that bias adult z_goal seeding across novel episodes.”

The V3-EXQ-588 (EXQ-ISEF-002) FAIL was reviewed non_contributory for MECH-189 (failure_autopsy_V3-EXQ-588_2026-05-19, confirmed): the transient-benefit-patch env feature (GAP-3) works (C2/C3 confirmed), but the run measured the within-episode GoalState attractor (MECH-112 / DEV-NEED-006 infant gate), not the child-phase ContextMemory super-ordinal write path the claim actually describes. The autopsy and developmental_needs_register.md:136 both conclude:

“MECH-189 needs cue-indexed persistent goal-anchor writes and an adult z_goal seeding readout before it can be gated.”

V3 had no such substrate. The IncentiveTokenBank (SD-057) is per-object and per-episode (GoalState.reset() clears it); the ghost-goal bank (MECH-292) is over hippocampal anchors. Neither is the cross-episode super-ordinal z_goal store MECH-189 needs. GoalState itself resets z_goal each episode. So the mechanism the claim describes — childhood-formed meta-goals that persist across adult episodes — had no place to live.

DEV-NEED-024 (Super-ordinal goal formation) carries the open question this substrate must answer: “What contextual-complexity threshold triggers a super-ordinal write?”

Solution

A new agent-owned, cross-episode-persistent cue-indexed store SuperOrdinalGoalMemory (ree-v3/ree_core/goal.py, sibling to IncentiveTokenBank), wired into REEAgent.update_z_goal. NOT reset by per-episode agent.reset() — cross-episode persistence is the defining property of a super-ordinal goal hierarchy versus an episodic z_goal.

Store

Each slot is a (key = z_world context [world_dim], value = z_goal anchor [goal_dim], strength) triple — cue-indexed: the key is the context cue, the value is the stored super-ordinal goal. Pure stateful tensor store + cosine arithmetic; no nn.Module, no trainable parameters, no gradient flow.

WRITE (child phase only)

After the within-episode GoalState.update(), the current z_goal is written as a super-ordinal anchor keyed on the current z_world context iff the MECH-189 conjunction holds:

(a) high salience — salience = benefit_exposure * (1 + drive_weight * effective_drive) >= super_ordinal_salience_threshold. The drive-modulated “large benefit spike”; routine low contacts do not clear it, transient-benefit patches do.
(b) high contextual complexity — complexity >= super_ordinal_complexity_threshold. Pluggable (super_ordinal_complexity_mode):
- "novelty" (default, self-contained): complexity = 1 - max cosine(z_world, occupied anchor keys). Empty store → 1.0 (first contacts bootstrap the hierarchy); a context already covered by an anchor → low complexity → no write (the spec’s selective neoteny — adult-routine stability).
- "external": complexity supplied by the caller, so an experiment can drive it from E1 cue-context entropy / prediction-error without coupling the substrate to those optional channels.
The mode/threshold is the DEV-NEED-024 adjudication target — to be settled by the validation EXQ and a follow-on lit-pull, not hard-coded here.

Literature verdict (DEV-NEED-024, 2026-06-10 — directional, not a settle). The follow-on lit-pull (targeted_review_connectome_mech_189; combined view mech_189_evidence_overview.md) finds the gate biologically justified but the default 'novelty' form questionable on two counts: (1) reference set — biological novelty is comparator-mismatch against the whole world model (Lisman & Grace 2005), not 1 − cosine to the goal-anchor keys; (2) functional form — the signal should be prediction-error magnitude (Aberg 2017) and likely signed/nonmonotonic (Quent 2022 U-shape: schema-congruent and schema-incongruent both encode), not a monotonic dissimilarity. Recommendation: prefer super_ordinal_complexity_mode='external' fed by an E1/E2 prediction-error / surprise signal; if 'novelty' is retained, reference the world model rather than anchor keys and consider a signed form. Transfer caveat: the cited PEs are value/RL prediction errors, whereas REE’s E1/E2 PE is a forward-model error — the validation EXQ should not assume they are interchangeable. Lit signal only; does not change MECH-189 status/confidence.

Writes are permitted only while write_enabled (the child phase). The curriculum freezes writes at the child→adult transition via REEAgent.set_super_ordinal_write_enabled(False) (the MECH-334 on_phase3_entry crystallization precedent).

Allocate vs reinforce (gate (b) governs formation only): gate (a) high salience is required for any write. A contact within merge_similarity of an existing anchor key reinforces that anchor (EMA-blend key+value toward the current z_goal, raise strength) — and this proceeds on salience alone, regardless of complexity, because a recurring high-salience context should strengthen its existing super-ordinal goal toward the matured z_goal. Gate (b) high contextual complexity governs only the allocation of a NEW anchor (a genuinely novel/rich context warrants a new meta-goal). Without this split the anchor would freeze at the tiny z_goal captured at its first contact and never learn the matured childhood meta-goal — surfaced and fixed by the V3-EXQ-588c smoke test, which went from anchor norm 0.019 (frozen-at-first-contact) to 0.373 (matured) after the correction.

READ (adult z_goal seeding readout)

Each update_z_goal tick, when the live z_goal norm is below super_ordinal_seed_below_norm (default 0.4 — matching the DEV-NEED-006 gate, i.e. the agent has no strong episodic goal of its own), the store is queried with the current z_world; if the best-matching anchor clears super_ordinal_seed_match_threshold, z_goal is pulled toward the retrieved super-ordinal anchor via GoalState.cue_pull (no benefit pulse). This is the “stored z_goal anchors bias z_goal seeding in adult episodes even in novel contexts” readout.

Config (all no-op defaults; bit-identical OFF)

On GoalConfig (mirrored through REEConfig.from_dims):

Param	Default	Purpose
`use_super_ordinal_goal_anchors`	`False`	master switch
`super_ordinal_n_slots`	`16`	cue-indexed anchor slots
`super_ordinal_salience_threshold`	`0.5`	WRITE gate (a)
`super_ordinal_complexity_mode`	`"novelty"`	WRITE gate (b) mode
`super_ordinal_complexity_threshold`	`0.3`	WRITE gate (b) floor
`super_ordinal_merge_similarity`	`0.8`	reinforce-vs-allocate
`super_ordinal_write_alpha`	`0.3`	EMA blend into a slot
`super_ordinal_seed_below_norm`	`0.4`	READ: seed only when z_goal below
`super_ordinal_seed_match_threshold`	`0.3`	READ: min retrieval match
`super_ordinal_seed_strength`	`0.1`	READ: cue-pull strength

Architecture Context

The store sits between GoalState (episodic z_goal attractor) and the cortical goal pipeline. It reuses GoalState.cue_pull (the SD-057 L6 directional nudge) for the adult seeding path and the existing update_z_goal salience/drive signals for the write gate. It does not modify GoalState, the IncentiveTokenBank, the hippocampal ghost-goal bank, or the E1 ContextMemory — it is a strictly additive agent-level store.

INV-037/INV-038 (stored vs active z_goal): the super-ordinal anchors are the stored representation; the live GoalState._z_goal is the active one. The write threshold is what makes childhood special (INV-041): constrained affordances + high-complexity contexts maximise super-ordinal formation, while adult routine contexts are low-complexity and do not trigger writes (selective neoteny, INV-056).

What This SD Enables

The infant_substrate:GAP-11 MECH-189 retest (a V3-EXQ-588 successor with a NEW letter), measuring adult z_goal seeding from childhood-formed anchors.
DEV-NEED-006 / DEV-NEED-024 gating.

Validation

V3-EXQ-588c (experiments/v3_exq_588c_mech189_super_ordinal_seeding.py, experiment_purpose=diagnostic, claim_ids=["MECH-189"], supersedes the 588 ISEF-002 chain). Forced-feed child phase (decoupled from the foraging-contact ceiling that blocked 588) forms cross-episode anchors → freeze → adult episodes seed a fresh sub-floor z_goal from the anchors with no benefit pulse. ARM_ON vs ARM_OFF × 3 seeds.

Load-bearing C1 (discrimination): ARM_ON adult median z_goal exceeds the per-seed ARM_OFF baseline by a clear margin (the MECH-189 write+read path moves adult z_goal where the no-store baseline gives ~0). This is the substrate question.
Advisory (reported, not gating): fraction of ARM_ON seeds crossing the DEV-NEED-006 / 588 governance gate (0.4). On the untrained-encoder readiness harness the matured-z_goal anchor norm ceilings at ~0.37 (the forced-feed z_world-EMA asymptote ≈ 0.9·‖z_world‖), so 0.4 is regime-bound — a near-miss with strong discrimination still validates the substrate and routes the absolute gate to a trained-encoder evidence successor, not a substrate failure.
Readiness / non-vacuity (else substrate_not_ready_requeue): ARM_ON anchors form, seeding fires, and a positive-control adult z_goal is non-zero.

Dry-run PASS (C1 discrimination 1.0; ARM_ON 0.28 vs ARM_OFF 0.0). Full-scale probe post gate-fix: ARM_ON adult median 0.364, anchor norm 0.373, vs ARM_OFF ~0.

MECH-094

Waking-only. update_z_goal is the waking seeding path; write(simulation_mode=True) is a no-op, so replay/DMN content cannot form super-ordinal anchors.

Phased training

Not applicable — pure stateful store + cosine arithmetic; no learned parameters.

INV-041, INV-037, INV-038, INV-055, INV-056, SD-016, MECH-112, MECH-116, MECH-117, MECH-329 (wanting-before-liking ordering; child of MECH-189), MECH-292/293 (ghost-goal bank; distinct hippocampal-anchor store), SD-057 (IncentiveTokenBank; distinct per-object per-episode store), MECH-094, DEV-NEED-006, DEV-NEED-024.