Anticipatory affect: conjunction architecture (MECH-307) vs SD-014 6-channel amendment

Created: 2026-05-08 Type: architecture proposal + comparative analysis Status: MECH-307 SUBSTRATE READY (IGW-20260521-023 closed 2026-05-21). Landed 2026-05-08/11; canonical validation V3-EXQ-540g PASS 2026-05-15. substrate_queue implementation_status=implemented. Claim candidate_substrate_landed; v3_pending retained until GAP-2/GAP-4 behavioural cohorts. SD-014 6-channel amendment retained as registered fallback.

2026-05-11 Option-b note (Gap 1): the design-doc proposed Option (a) signed single channel as default; the user-override directed Option (b) split into VALENCE_POSITIVE_SURPRISE / VALENCE_NEGATIVE_SURPRISE as separate channels. Implementation extends VALENCE_DIM 4 -> 6 in ree_core/residue/field.py with indices 4 (positive) and 5 (negative). MECH-205 write site at the PE-generating tick routes by harm-signal sign and ALSO writes magnitude to the legacy VALENCE_SURPRISE channel for backward-compat with existing consumers (MECH-205 replay-priority, SD-014 evaluate_valence readers). Master convenience flag use_mech307_conjunction propagates to the three substrate-side sub-flags via REEConfig.post_init. 309/309 contracts + 7/7 preflight PASS with master OFF (bit-identical OFF guarantee). Plan-of-record goal_pipeline_plan.md GAP-1 status open -> done 2026-05-11. Behavioural validation (4-arm discriminative pair) queued in a separate /queue-experiment session per the user “substrate only first” directive. Anchors: evidence/literature/targeted_review_excitement_5th_valence_channel/ (9 entries, lit_conf 0.77); evidence/literature/wanting_liking_sleep_consolidation_synthesis.md; mech188_vs_mech295_dual_path.md.

The architectural question

The 2026-05-08 lit-pull on excitement-as-distinct-affect established that anticipatory positive arousal at reward cues (NAcc-anticipation in Knutson 2001a et seq.) is neurally dissociable from VALENCE_WANTING (directional motivation), VALENCE_LIKING (consummatory hedonic), and z_beta arousal (timing modulation). The lit-pull’s naive verdict was REGISTER VALENCE_EXCITEMENT as a 5th channel in SD-014’s residue valence vector.

On code-level inspection, a more parsimonious answer emerged: excitement should fall out of the existing 4 channels as a derived conjunction-state, but currently does not, because of four specific wiring gaps. Fixing the wiring is more biologically faithful than adding a new channel — biology does not have a “VALENCE_EXCITEMENT neuron type”; the excitement signal measured in NAcc fMRI is the anatomical convergence of DA RPE + hippocampal preplay + ANS arousal at one structure.

This document records both proposals, their relationship, and the empirical test that adjudicates between them.

The four wiring gaps (MECH-307 first-line proposal)

Gap 1: VALENCE_SURPRISE is unsigned magnitude

Code: agent.py:3075-3077:

pe_mag = float(pe_val.detach())                            # magnitude only
self._pe_ema = (1 - self._pe_ema_alpha) * self._pe_ema + self._pe_ema_alpha * pe_mag
surprise = max(0.0, pe_mag - self._pe_ema)                 # max(0, ...) -- non-negative

A positive PE (good thing happened unexpectedly — the dopamine RPE that underlies excitement in biology) and a negative PE (bad thing happened unexpectedly — the habenula signature of dread) are written identically to VALENCE_SURPRISE. The sign is discarded at the write site. There is no way for any downstream consumer to distinguish “joy at unexpected reward” from “panic at unexpected harm” by reading the residue field.

Fix: Either (a) store signed PE in VALENCE_SURPRISE (negative for negative PE, positive for positive PE), with consumers reading sign or magnitude as needed; or (b) split into VALENCE_POSITIVE_SURPRISE and VALENCE_NEGATIVE_SURPRISE as separate channels.

Option (a) is cheaper (~5 lines, one field semantics change behind a config flag config.surprise_signed) and preserves backward compatibility (consumers reading magnitude get the same behaviour as today via abs()). Option (b) is cleaner but more disruptive. Recommend (a) as default; (b) only if a downstream consumer needs the channel-level separation.

Gap 2: MECH-216 schema readout writes only to VALENCE_WANTING

Code: agent.py:3753-3757:

self.residue_field.update_valence(
    z_world, component=VALENCE_WANTING,
    value=wanting_value, hypothesis_tag=False,
)

When schema_salience fires above threshold, the schema readout produces a single write — to VALENCE_WANTING only. Biology’s NAcc-anticipation signal at reward cues drives, jointly: (i) wanting amplification (already done), (ii) sympathetic activation through ANS pathways (z_beta arousal should rise — currently absent), and (iii) partial anticipatory hedonic activation (a VALENCE_LIKING write before contact — currently absent). MECH-216 currently implements only (i).

Fix: When schema_salience >= threshold AND use_mech307_conjunction=True, write all three jointly:

# proposed
gain = self.config.schema_wanting_gain
sal_val = float(self._schema_salience.squeeze())
drive_lvl = drive_level
predicted_z_world = self._z_world_predicted  # from gap 4 fix; fallback to current z_world
target_z = predicted_z_world if predicted_z_world is not None else z_world

# Gap 2 (i): VALENCE_WANTING (existing)
self.residue_field.update_valence(target_z, VALENCE_WANTING,
                                   sal_val * gain * max(drive_lvl, 0.1))
# Gap 2 (ii): partial anticipatory VALENCE_LIKING
self.residue_field.update_valence(target_z, VALENCE_LIKING,
                                   sal_val * self.config.mech307_anticipatory_liking_gain * drive_lvl)
# Gap 2 (iii) -- subsumes Gap 3: z_beta arousal pulse
self.latent_stack.tick_z_beta_pulse(sal_val * self.config.mech307_z_beta_schema_gain)

The tick_z_beta_pulse method does not yet exist; would need to be added to LatentStack.

Gap 3: z_beta has no input path from cue prediction

Code analysis: z_beta is encoded from sensory observation in LatentStack. It is read by MECH-093 to modulate E3 heartbeat rate. There is no path from MECH-216 schema readout to z_beta amplitude. So when the schema head says “imminent positive event” (the upstream signature of excitement), the arousal axis does not elevate.

Fix: Subsumed into Gap 2. The tick_z_beta_pulse method increments z_beta arousal amplitude by a small amount proportional to schema_salience. This wires the cue → ANS path that biology has anatomically.

Gap 4: MECH-216 writes at the agent’s current z_world, not at the predicted-imminent location

Code: agent.py:3754 writes to self._current_latent.z_world. The residue field is marked at “wherever the agent is now,” not at “where the agent predicts the goal will be.” Biology’s hippocampal place-cell preplay + NAcc-anticipation marks the predicted goal location.

Fix: Use E1’s forward prediction (already computed in _e1_tick) to identify the predicted z_world, write at that location instead. Falls back to current z_world when prediction is not available.

What “excitement” looks like under the conjunction reading

Under all four fixes, the residue field at a predicted-imminent-reward location carries:

VALENCE_WANTING amplitude high (anticipatory wanting via MECH-216)
VALENCE_LIKING amplitude high (anticipatory hedonic via Gap 2 fix)
VALENCE_POSITIVE_SURPRISE amplitude high (signed PE from a cue that exceeded prediction; Gap 1 fix)
z_beta arousal elevated globally (Gap 3 fix)

A consumer that needs to detect excitement reads the joint state at a location:

def is_excitement_state_at(z_loc, residue_field, z_beta_arousal, threshold=0.6):
    v = residue_field.evaluate_valence(z_loc)
    return (
        v[VALENCE_WANTING] > threshold
        and v[VALENCE_LIKING] > threshold * 0.5  # anticipatory liking is partial
        and v[VALENCE_SURPRISE] > 0  # signed -- positive PE only
        and z_beta_arousal > threshold
    )

Functional consumers that benefit:

MECH-205 surprise-gated replay write: priority weighted by conjunction detection. Adcock 2006 prediction: locations with high anticipatory-conjunction at encoding receive

= 1.5x replay frequency vs uniform-priority baseline.
MECH-292 ranked ghost-goal bank: ghost_priority += w_conj * is_excitement_state(r).
MECH-279 PAG freeze gate: VALENCE_NEGATIVE_SURPRISE + VALENCE_HARM_DISCRIMINATIVE + z_beta high → dread state → stronger freeze probability.

Why MECH-307 is first-line and SD-014 6-channel is fallback

MECH-307 advantages:

Biological fidelity: excitement is composed of multiple signals at convergence, not a primitive neuron type. The conjunction architecture matches biology’s actual organisation.
Cost: ~40 lines of code, all backward-compatible behind use_mech307_conjunction=False.
Diagnostic value: fixing Gap 1 (signed PE) likely also fixes the EXQ-141 curiosity drive failure (MECH-111), because that test depended on an upstream signed-novelty signal that the substrate cannot currently produce. One fix unblocks two tests.
Parsimony: avoids growing the residue field’s valence_vecs buffer from [K, 4] to [K, 6].

SD-014 6-channel amendment advantages:

Direct write site: a consumer that needs excitement reads v[VALENCE_EXCITEMENT] once, not a four-channel joint condition. Less dispersed, less error-prone if many consumers need the signal.
Easier to ablate: setting valence_excitement_enabled=False cleanly disables the anticipatory-positive-arousal writes; the conjunction architecture is harder to ablate cleanly because each gap-fix is a separate flag.
Phenomenologically transparent: the channel name maps to the construct name. Easier to communicate in claim text.

SD-014 amendment proposal text (FALLBACK – not yet applied to registry)

Should the MECH-307 conjunction-fix experiment fail or produce unreliable conjunction detection at consumer sites, this is the proposed registry amendment to SD-014:

SD-014 (proposed amendment 2026-05-XX, FALLBACK pending MECH-307 experimental outcome)

Status change: provisional (current) -> provisional (amended; quadrant unchanged
pending experimental validation of the 6-channel architecture).

V is extended from 4-component to 6-component:
  V = [VALENCE_WANTING, VALENCE_LIKING, VALENCE_HARM_DISCRIMINATIVE,
       VALENCE_SURPRISE, VALENCE_EXCITEMENT, VALENCE_DREAD]

VALENCE_EXCITEMENT (index 4, new): high-arousal-positive anticipatory affect.
  Write site: cue-stage anticipatory write by MECH-216 schema readout when
    schema_salience >= threshold AND drive_level >= threshold AND prediction_certainty
    >= threshold. Write magnitude proportional to (schema_salience x drive x certainty).
  Update rule: e <- (1-alpha_e) * e + alpha_e * (excitement_value)
  Decay: very slow (alpha_e ~ 0.005 to match VALENCE_WANTING decay timescale).
  Hypothesis_tag gating: writes only when hypothesis_tag=False (waking).

VALENCE_DREAD (index 5, new): high-arousal-negative anticipatory affect.
  Symmetric to VALENCE_EXCITEMENT. Write site: when MECH-216 schema readout predicts
    imminent harm (predicted z_harm_a > threshold) with low controllability prediction.
    Cross-link to MECH-305 controllability-weighted affective load.
  Drives MECH-279 PAG freeze gate as one input among several.

Functional consumers:
  - MECH-205 surprise-gated replay write: priority bonus on locations with high
    VALENCE_EXCITEMENT (per Adcock 2006 prediction).
  - MECH-292 ghost_priority: ghost_priority += w_e * VALENCE_EXCITEMENT(r) -
    w_d * VALENCE_DREAD(r).
  - MECH-279 PAG freeze gate: VALENCE_DREAD as input alongside z_harm_a, override_signal.

Backward compatibility: valence_excitement_enabled / valence_dread_enabled config flags,
  default True (the channels exist but are written only when the gating conditions fire).
  Bit-identical when both flags are False (channels exist as zero-valued buffer slots).

Lit anchors: Knutson 2001a [DOI](https://doi.org/10.1523/JNEUROSCI.21-16-j0002.2001),
  Knutson 2001b [DOI](https://doi.org/10.1097/00001756-200112040-00016),
  Adcock 2006 [DOI](https://doi.org/10.1016/j.neuron.2006.03.036),
  Berns 2006 [DOI](https://doi.org/10.1126/science.1123721),
  Posner 2009 [DOI](https://doi.org/10.1002/hbm.20553),
  Bromberg-Martin 2010 [DOI](https://doi.org/10.1016/j.neuron.2010.11.022),
  Bromberg-Martin 2010b [DOI](https://doi.org/10.1016/j.neuron.2010.06.016),
  Knutson 2008 [DOI](https://doi.org/10.1016/j.biopsych.2007.07.023),
  Johnson 2019 [DOI](https://doi.org/10.1016/j.nicl.2019.102018).

Discriminative experiment (queued but not registered)

v3_exq_XXX_anticipatory_affect_arch_ablation: 4 arms × 3 seeds, ~3-4h on Mac.

ARM_0_off: current substrate (all 4 wiring gaps unfixed, baseline)
ARM_1_signed_pe: Gap 1 fix only (Gaps 2-4 unfixed)
ARM_2_full_conjunction: All 4 fixes (MECH-307 first-line architecture)
ARM_3_5channel: SD-014 6-channel amendment (VALENCE_EXCITEMENT + VALENCE_DREAD added as discrete channels written by MECH-216, all 4 wiring gaps unfixed)

Acceptance:

C1: conjunction-state detection rate at predicted-reward locations: ARM_2 >= 0.7, ARM_0 < 0.2 (test that the fixes work).
C2: Adcock 2006 prediction: sleep-replay frequency at conjunction-state / uniform-priority baseline >= 1.5x in ARM_2.
C3: ARM_2 vs ARM_3 head-to-head: probe accuracy on subsequent identity discrimination (analog of EXQ-538 SD-049 Phase 2) is comparable, but ARM_2 uses fewer net residue-field writes (parsimony measure).

PASS = C1 AND C2 AND C3 → MECH-307 conjunction-fix is the right architecture; SD-014 amendment can be retired as fallback.

PARTIAL PASS (C1 + C2 only, ARM_3 outperforms on C3) → SD-014 amendment is the better architecture; apply registry amendment.

FAIL on C1 → wiring fixes don’t compose into a detectable conjunction state; investigate whether the issue is at the read site (consumer detection) or the write site (gap-fix implementation incorrect). Diagnostic before any architectural commitment.

Sequencing

EXQ-537 (SD-029 single-pass comparator) lands first to clear the currently-running commit-chain diagnostic queue.
EXQ-141b (MECH-111 retest on corrected commit-chain substrate) lands to clear the curiosity-drive question with the same fixes applied.
EXQ-XXX (this experiment, MECH-307 4-arm + SD-014-amendment alternative arm) lands to adjudicate between the two architectures.
Apply the winning architecture to claims.yaml + ree-v3 implementation.

Open considerations

The MECH-307 conjunction architecture moves complexity to the consumer side (each consumer does a 4-channel joint read). If the V3 registry grows to many consumers of excitement / dread states (MECH-205, MECH-292, MECH-279, future MECH-3XX), the bookkeeping cost grows linearly with consumer count. SD-014 amendment writes once, reads once at each consumer. The right balance depends on how many consumers eventually need the signal.
The signed VALENCE_SURPRISE fix (Gap 1) has implications beyond excitement: any consumer that currently reads VALENCE_SURPRISE assumes magnitude semantics. Auditing existing consumers is part of the implementation cost. Identified consumers as of 2026-05-08: MECH-205 surprise-gated replay (reads magnitude OK), MECH-094 hypothesis tag write gate (does not consume VALENCE_SURPRISE directly). No urgent breakage expected, but this audit should land before flipping surprise_signed=True as default.
The Gap 4 fix (write at predicted z_world) requires E1 to expose its forward prediction via a stable accessor. Currently E1’s prediction is internal; would need a new method e1.get_z_world_prediction() returning a [batch, world_dim] tensor or None.