Goal / Wanting Signal Chain — Provenance Map

Scope: z_goal seeding, wanting pathway, and approach motivation in V3
Status: Diagnostic / architectural note
Claim IDs: SD-012, SD-015, SD-018, MECH-112, MECH-216, ARC-030, INV-065
Created: 2026-04-12
Why this file exists: EXQ-322, EXQ-328, EXQ-331 produced a cascade of non_contributory FAILs. Root cause is not amplification failure (SD-012 drive modulation is correctly implemented) but a structural bootstrap gap in the signal chain. This document maps the chain to show where the gap is and what each queued experiment targets.

1. The Chain: Source to Behavior 

world_obs [250 or 300 dims]
  |
  +--[ResourceEncoder (SD-015)]---> z_resource [32]
  |       supervised by: max(resource_field_view)         [TRAINED; wired into seeding]
  |       resource_prox_pred_r [0,1]                     [TRAINED; not wired as drive signal]
  |
  +--[SplitEncoder z_world path]---> z_world [32]
  |       supervised by: event_CE (SD-009) + proximity_head (SD-018)
  |
  +--[E1 LSTM hidden state]---> schema_salience [0,1]    (MECH-216)
          supervised by: resource_proximity_target (SD-018 target)
          gated by: schema_wanting_enabled flag (default False)

obs_body[3]  (agent_energy, [0,1])
  |--> drive_level = 1.0 - energy                        [SD-012, WIRED]

obs_body[11] (benefit_exposure, [0,1])
  |--> raw benefit contact signal                        [WIRED; zero except at contact]

GoalState.update() comparator (SD-012):

effective_benefit = benefit_exposure * z_goal_seeding_gain * (1 + drive_weight * drive_level)
if effective_benefit > benefit_threshold (0.1):
    z_goal <- (1 - alpha_goal) * z_goal + alpha_goal * z_source
              where z_source = z_resource (if SD-015 enabled) else z_world

update_schema_wanting() comparator (MECH-216):

if schema_salience >= schema_wanting_threshold (0.3):
    wanting_value = schema_salience * schema_wanting_gain * max(drive_level, 0.1)
    ResidueField.update_valence(z_world, VALENCE_WANTING, wanting_value)

Downstream:

z_goal (GoalState) -> E3.score_trajectory() via goal_weight
VALENCE_WANTING (ResidueField) -> hippocampal replay prioritization, trajectory scoring

2. Signal Provenance Table

Signal Source channel Encoder/module Supervision target Discrimination Wired into behavior? Known failure mode
z_resource world_obs ResourceEncoder max(resource_field_view) resource-type cells vs background YES — seeded into GoalState when use_resource_encoder=True Sparse training signal in random-walk warmup; position-correlated signal
resource_prox_pred_r z_resource ResourceEncoder.resource_prox_head max(resource_field_view) proximity regression NO — computed but not used as drive signal Output consumed only by training loss; never reads downstream as motivation
drive_level obs_body[3] compute_drive_level() none (derived directly) energy depletion vs satiation YES — scales effective_benefit in GoalState.update() Amplifies seeding at contact only; does not cause approach
benefit_exposure obs_body[11] none (env scalar) none (ground truth) contact vs non-contact (binary at threshold) YES — input to GoalState.update() Contact-gated: zero except at resource cell entry; bootstrap deadlock
z_world world_obs SplitEncoder event-CE + proximity_head scene encoding (full position + content) YES — fallback z_goal seed; E3 trajectory scoring Position-confounded; after respawn z_world at new location has no goal signal
schema_salience E1 LSTM hidden[0][-1] schema_readout_head resource_proximity_target resource cue presence from E1 sequence model Conditional — only when schema_wanting_enabled=True (default False) Requires trained E1 with resource encounter history; also sparse in warmup
z_goal GoalState attractor GoalState.update() benefit_exposure (above threshold) goal maintained vs zero YES — E3 trajectory scoring via goal_proximity() Never activated if contact never occurs (EXQ-322: n_cosine_samples=0)
VALENCE_WANTING ResidueField update_schema_wanting() / update_benefit_salience() benefit_salience (serotonin) wanting terrain in residue field YES — replay prioritization + hippocampal navigation Zero unless schema_wanting writes occur

3. The Bootstrap Gap

The chain has a circular dependency that prevents cold-start seeding:

z_goal needed for approach motivation
    -> approach motivation needed to reach resources
       -> resource contact needed to produce benefit_exposure > 0
          -> benefit_exposure needed to cross benefit_threshold
             -> threshold crossing needed to seed z_goal

SD-012’s drive modulation (effective_benefit = benefit_exposure * (1 + 2 * drive_level)) correctly amplifies the signal when contact happens, but does not cause contact to happen. With drive_level=1.0 (fully depleted), a single contact produces effective_benefit = 0.04 * 1 * 3 = 0.12 > 0.1 threshold — so SD-012 works correctly. The problem is that random-walk warmup produces near-zero contact rate, so the threshold is rarely reached at all, not that it is too high when reached.

This is the root cause of the non_contributory cascade:

  • EXQ-322 (SD-015): n_cosine_samples=0 — z_goal never seeded, so the comparison between RESOURCE_SEED and WORLD_SEED conditions could not be measured.
  • EXQ-328 (MECH-112): z_goal_norm=0.0 across all 6 seeds/conditions — same root.
  • EXQ-331 (ARC-030): “Go sub-channel making zero difference” — with z_goal=0, the approach drive is zero in both JOINT and HARM_ONLY conditions.

The failure is at the benefit_exposure stage: the signal source is contact-gated (binary: 0 except at resource cell entry) and contact is rare under random walk.

4. Wanting vs Liking — The Architectural Distinction

The gap corresponds to the Berridge wanting/liking dissociation:

Component Signal type Trigger Biology REE module
Liking hedonic impact contact (resource consumed) opioid system, NAc shell benefit_exposure / obs_body[11]
Wanting incentive salience cue detection (resource visible) dopamine, NAc core, VTA missing until MECH-216 is enabled

The current contact-gated pathway implements liking-triggered goal seeding, not wanting-triggered approach motivation. Without a cue-based wanting signal, the agent has no motivation to approach resources before first contact — which is the behavioral equivalent of having no appetite until food is already in your mouth.

MECH-112 (wanting/liking dissociation) cannot be tested until both signals exist independently. The claim requires wanting to fire on cue detection and liking to fire on contact, and for these to dissociate under resource relocation. That test is vacuous if the wanting pathway is a zero vector throughout.

5. What EXQ-326 / EXQ-332 Target in This Chain

EXQ-326 (wanting_gradient_nav_fix, in queue):

  • Enables schema_wanting_enabled=True — wires schema_salience from E1 LSTM into VALENCE_WANTING writes on each tick when salience > threshold.
  • Enables use_resource_proximity_head=True — SD-018 supervision forces z_world to represent resource proximity, giving E1 something discriminative to read from.
  • Sets drive_weight=2.0 (SD-012) — correct amplification when contact does happen.
  • Three-phase training: P0 encoder warmup (proximity supervision active, no goal seeding), P1 wanting active (schema_wanting writes VALENCE_WANTING into residue field), P2 evaluation (benefit_ratio lifted vs ablation baseline).

The VALENCE_WANTING terrain in the residue field acts as a cue-driven gradient field for trajectory scoring — the hippocampal navigator prefers trajectories through high-VALENCE_WANTING regions even before z_goal is seeded from contact. This breaks the bootstrap: the agent can develop approach tendency from field-view signals, encounter resources more often, and then seed z_goal from those encounters.

EXQ-332 (MECH-216 predictive wanting, in queue):

  • Redesigns the supervision target for the schema_readout_head: instead of current resource_proximity (reactive detection), the target is future_resource_encounter (binary: did the agent collect a resource in the next 10 steps, computed from episode-level replay of cached E1 LSTM states).
  • Attenuates direct resource gradient in observation (resource_obs_scale=0.3) so Landmark B (a predictive environmental cue co-located near resources) becomes the dominant discriminative input.
  • This makes schema_salience genuinely predictive: it fires on landmark detection before the resource is in immediate field of view, enabling earlier approach initiation. Compare corr(schema_salience, future_encounter) vs corr(schema_salience, current_prox) to distinguish predictive from reactive.

6. Open Gap: resource_prox_pred_r is Computed but Not Consumed

ResourceEncoder.resource_prox_pred_r is computed on every call to LatentStack.encode() when use_resource_encoder=True. Its training loss (MSE against max(resource_field_view)) backpropagates through ResourceEncoder weights. But the output value of resource_prox_pred_r is never read downstream by any behavioral module.

Compare this to MECH-216’s schema_salience, which IS wired: when schema_wanting_enabled=True, schema_salience writes into VALENCE_WANTING each tick.

The resource_prox_pred_r output could serve as a simpler, more direct wanting signal than schema_salience — it is already calibrated to [0,1] (Sigmoid) and supervised by resource proximity. However, directly wiring it would create a reactive (non-predictive) wanting signal. EXQ-332’s future-target redesign of schema_salience is the preferred solution because it captures anticipatory wanting, not just reactive detection.

This gap is not a bug — it is the correct design choice to prefer the predictive pathway — but it should be documented so that ablation experiments can isolate the two pathways: resource_prox_pred_r as reactive wanting vs schema_salience as predictive wanting.

7. Verification Checklist for New Wanting Experiments

Before claiming z_goal seeding is functional in an experiment:

  1. n_seeding_events > 0 — at least one z_goal update must have occurred (check: goal_norm_peak > 1e-6 across all seeds)
  2. n_cosine_samples > 0 — seeding must have occurred in enough episodes to measure the z_resource vs z_world cosine similarity comparison
  3. benefit_ratio > 0 — agent must be collecting resources in at least some episodes (otherwise the behavioral test is vacuous)
  4. schema_salience distribution — if MECH-216 is enabled, verify schema_salience is not always < threshold (check histogram, not just mean)
  5. VALENCE_WANTING writes — check mech216_write_count or equivalent diagnostic that wanting terrain is actually being written

If any of (1)-(3) are zero, the experiment is non_contributory for the same reason as EXQ-322/328/331. Report as non_contributory, do not reassign outcome.

Claim Role in chain Gate status
SD-012 Drive-scaled benefit threshold (correctly implemented) VALIDATED conceptually; EXQ-233 non_contributory due to gap
SD-015 ResourceEncoder object-type seeding IMPLEMENTED; EXQ-322 non_contributory (gap at benefit_exposure)
SD-018 z_world resource proximity supervision (prerequisite for MECH-216) IMPLEMENTED; required for schema readout training
MECH-112 Wanting/liking dissociation BLOCKED — requires both wanting and liking signals to exist independently
MECH-216 E1 predictive wanting via schema_salience IMPLEMENTED; enabled by EXQ-326
ARC-030 D1/D2 approach-avoidance balance BLOCKED — approach side zero until wanting seeded
INV-065 Proxy goal necessity (FIRST-PAPER GATE) Blocked on benefit_ratio > 0 evidence
EXQ-326 wanting_gradient_nav_fix — wires MECH-216 as approach signal In queue, pending
EXQ-327 MECH-163 goal-conditioned navigation Blocked on EXQ-326 PASS
EXQ-332 MECH-216 predictive wanting (future-target) In queue, lower priority than 326

Notes on Supervision Signal Quality

max(resource_field_view) as a supervision target for ResourceEncoder is a proximity scalar, not a categorical identity signal. The claim that it “forces z_resource to represent resource-type features, not spatial position” (SD-015 doc) requires a qualification: it forces z_resource to detect resource-type cells in world_obs, which IS identity-based (resource cells have a distinct type encoding in CausalGridWorldV2), but the training signal is sparse at any given position unless the agent is near a resource.

The identity-generalization property (works after respawn because resource cell type is the same across locations) depends on:

  • The resource cell encoding in world_obs being type-consistent (not position-encoded).
  • The encoder seeing sufficient examples of resource cells during training.

In practice, the second condition is violated in random-walk warmup before EXQ-326 enables the wanting gradient. Post-EXQ-326, the approach tendency generated by VALENCE_WANTING terrain increases resource encounter rate, which in turn provides denser training signal for ResourceEncoder. The two pathways are therefore mutually reinforcing once either achieves minimal activation — the problem is only in cold-start.

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