MECH-295: Drive -> Liking-Stream -> Approach Cue Bridge

Claim ID: MECH-295 Subject: drive_to_approach.liking_stream_bridge Status: IMPLEMENTED (V3 substrate landed 2026-04-26) Reading commitment: WEAK necessity (provisional) Registered: 2026-04-26 Implemented: 2026-04-26 Depends on: SD-012, SD-014, SD-015, SD-016, MECH-117, ARC-036 Blocks: EXQ-483 catatonic-lock recovery; downstream behavioural validation of the V3 motivational stack

Functional restatement 

drive_level (SD-012)
   ->
liking_gain on goal-congruent outcome predictions
   ->
approach_cue_signal at action selection (E3 / BG)

The bridge is the missing wiring between SD-012 drive amplification, the SD-014 / SD-015 liking-stream substrate, and E3 / BG action selection. Without this bridge, drive amplification produces a passive z_goal latent without behavioural consequence – the EXQ-483 catatonic-lock signature: override_signal climbs to mean 0.563, PAG release ratio ON_ON / ON_OFF = 1.69, but approach_commit = 0.0 across all four arms.

Strong vs weak reading

The lit-pull (evidence/literature/targeted_review_mech295_liking_approach_bridge/SYNTHESIS.md) cleanly motivates the architecture but does not arbitrate two readings of necessity:

  • Strong necessity: elevated approach_commit requires elevated liking-stream activation (level-coupled).
  • Weak necessity: baseline liking-stream activation is sufficient, but the bridge wiring must be intact – if the link is severed, drive amplification produces no approach regardless of drive magnitude.

The DAT-knockdown finding (Pecina 2003: more wanting, unchanged liking) is incompatible with the strong reading but compatible with the weak reading. MECH-295 commits to the WEAK reading provisionally.

This V3 substrate implements the weak reading. The cue-side gain is a function of drive * goal_proximity – the “is the bridge intact?” surface, not drive * residue.liking which would be the level-coupled strong reading. Setting mech295_liking_to_approach_cue_gain=0.0 is the severed-bridge arm of the falsifiable test.

Architecture

Module

ree-v3/ree_core/regulators/mech295_liking_bridge.py:

  • MECH295LikingBridgeConfig – four sub-knob dataclass.
  • MECH295LikingBridge – pure arithmetic, no trainable parameters. Two compute methods:
    • compute_anticipatory_liking_write(drive_level, z_goal_norm, simulation_mode) -> scalar magnitude for the anticipatory liking write at the goal location.
    • compute_approach_cue_score_bias(drive_level, candidate_proximities, simulation_mode) -> per-candidate negative score_bias (E3 lower-is-better, so liking favours approach by reducing the score).
  • MECH295LikingBridgeOutput – diagnostic dataclass.
  • tick(drive_level, z_goal_norm, candidate_proximities, simulation_mode) – convenience orchestration helper that calls both compute methods and caches diagnostics. The agent normally calls the compute methods separately at their respective integration sites; tick exists for the contract tests and direct callers.

Integration sites

Two integration sites in ree-v3/ree_core/agent.py:

(a) Anticipatory liking write – update_z_goal()

After the existing SD-012 / SD-037 effective_drive computation and GoalState.update() call:

if (
    self.mech295_bridge is not None
    and self.goal_state is not None
    and self.goal_state.is_active()
    and hasattr(self.residue_field, "update_valence")
):
    z_goal_norm = self.goal_state.goal_norm()
    write_value = self.mech295_bridge.compute_anticipatory_liking_write(
        drive_level=effective_drive,
        z_goal_norm=z_goal_norm,
        simulation_mode=False,
    )
    if write_value > 0.0:
        self.residue_field.update_valence(
            self.goal_state.z_goal,         # write at GOAL location
            component=VALENCE_LIKING,
            value=write_value,
            hypothesis_tag=False,
        )

This is the anticipatory cue-side pulse, distinct from the existing consummatory-contact write in REEAgent.update_liking(benefit_exposure) (which writes at the agent’s current z_world, not at the goal). Both write paths target VALENCE_LIKING in the SD-014 valence vector but at different locations and from different mechanisms.

(b) Approach cue at action selection – select_action()

After the lateral_pfc + ofc score_bias composition, before e3.select():

if (
    self.mech295_bridge is not None
    and self.goal_state is not None
    and self.goal_state.is_active()
    and self._current_latent is not None
):
    # Build per-candidate first-step z_world summaries (reuse from
    # lateral_pfc / ofc blocks if available, else build fresh).
    m295_summaries = ...   # [K, world_dim]
    with torch.no_grad():
        cand_proximities = self.goal_state.goal_proximity(m295_summaries).detach()
        base_drive = float(getattr(self.goal_state, "_last_drive_level", 0.0))
        eff_drive_m295 = (
            self.pacc.effective_drive(base_drive)
            if self.pacc is not None else base_drive
        )
        m295_bias = self.mech295_bridge.compute_approach_cue_score_bias(
            drive_level=eff_drive_m295,
            candidate_proximities=cand_proximities,
            simulation_mode=False,
        )
    if dacc_score_bias is None:
        dacc_score_bias = m295_bias
    else:
        dacc_score_bias = dacc_score_bias + m295_bias.to(...)

The pACC sensitisation is applied so the cue side sees the same effective drive that the write side used (no double counting – the write side already received effective_drive post-pACC).

Computation

Write side (anticipatory liking) 

if simulation_mode or drive_to_liking_gain == 0.0:
    return 0.0
if drive_level < min_drive_to_fire:                  return 0.0
if z_goal_norm < min_z_goal_norm_to_fire:            return 0.0
return drive_to_liking_gain * drive_level * z_goal_norm

Cue side (per-candidate score_bias) 

if simulation_mode or liking_to_approach_cue_gain == 0.0:
    return zeros([K])
if drive_level < min_drive_to_fire:
    return zeros([K])
liking_signal = drive_level * candidate_proximities.clamp(0, 1)
return -liking_to_approach_cue_gain * liking_signal      # NEGATIVE

Note that the cue side fires on drive * candidate_proximities directly, NOT on the residue-field VALENCE_LIKING readout. This is the weak-reading commitment: baseline liking-stream activation is sufficient; the bridge does not require the level-coupled write to produce the cue. The drive_to_liking_gain only controls the anticipatory write magnitude, which is observed downstream (residue field, replay prioritisation via SD-014 drive-weighted priority).

Configuration

REEConfig.use_mech295_liking_bridge (master, default False). Sub-knobs:

Sub-knob Default Purpose
mech295_drive_to_liking_gain 1.0 Multiplier on drive * z_goal_norm for the anticipatory write. Set 0 to disable the write side without touching the cue side.
mech295_liking_to_approach_cue_gain 0.5 Multiplier on per-candidate liking signal for the score_bias. Set 0 to sever the cue side without touching the write side.
mech295_min_drive_to_fire 0.1 Drive floor. Below this, both sides are silent.
mech295_min_z_goal_norm_to_fire 0.05 Goal-norm floor for the write side.

All wired through REEConfig.from_dims().

Backward compatibility

Master flag default False; agent.mech295_bridge is None; both integration sites are no-ops. 154/154 contracts + 7/7 preflight PASS with flag OFF (bit-identical to pre-MECH-295 HEAD, 2026-04-26).

With master flag ON but both gain knobs at zero (mech295_drive_to_liking_gain=0.0, mech295_liking_to_approach_cue_gain=0.0), the bridge runs but produces zero write and zero score_bias – still bit-identical to OFF for action selection.

MECH-094

simulation_mode argument is honoured at both compute methods. When True:

  • write returns 0.0,
  • cue returns zero score_bias,
  • internal counters do not advance.

The agent calls the compute methods only on waking paths (update_z_goal from the experiment loop, select_action from the agent loop), passing simulation_mode=False. Replay / DMN paths must not route through these call sites; this is the same call-site-scoping pattern used by MECH-269 / MECH-287.

Falsifiability

Two falsifiable predictions:

Primary (weak reading)

A V3 factorial with the drive -> liking link intact vs severed (under matched drive_level) should show approach_commit recovers when the bridge is intact and collapses when severed. Operationally:

  • Arm A (severed, control): use_mech295_liking_bridge=True, mech295_drive_to_liking_gain=1.0, mech295_liking_to_approach_cue_gain=0.0. Bridge wired but cue side cannot reach E3.
  • Arm B (intact): use_mech295_liking_bridge=True, mech295_drive_to_liking_gain=1.0, mech295_liking_to_approach_cue_gain=0.5.

Prediction (weak reading): under matched drive elevation, Arm B approach_commit_count > 0 while Arm A approach_commit_count = 0 (or substantially lower).

Secondary (strong-reading test)

Pharmacologically-inspired manipulation: block the hedonic-hotspot mu-opioid analog in the V3 substrate while drive is elevated. Under the strong reading, approach_commit collapses. Under the weak reading with intact bridge, approach_commit recovers via baseline liking. Substrate not yet built – this is a future EXQ.

Acceptance for V3-EXQ-493

V3-EXQ-493 is a six-part substrate-landing diagnostic (queued 2026-04-26). All sub-tests must PASS:

  • UC1 module-importable – regulators.mech295_liking_bridge importable.
  • UC2 master-OFF no-op – agent built with default config has mech295_bridge is None.
  • UC3 write-side fires – 30-tick env loop with forced drive=0.8 + benefit=0.4 -> n_write_fires > 0.
  • UC4 cue-side negative bias – compute_approach_cue_score_bias(drive=0.6, prox=[0.1, 0.5, 0.9]) returns negative tensor with monotone proximity dependence.
  • UC5 severed-bridge collapse – with liking_to_approach_cue_gain=0.0 and drive_to_liking_gain=1.0, cue-side bias is exactly zero even at drive=0.9; write side STILL fires. This is the falsifiable signature for the weak-necessity reading.
  • UC6 MECH-094 simulation gate – simulation_mode=True -> zero on both sides, counters do not advance.

PASS = all 6 UCs PASS. Smoke verified all 6 PASS via --dry-run on 2026-04-26.

Behavioural validation (deferred)

Behavioural EXQ-483-style approach_commit recovery (4-arm with the orexin substrate ON plus this bridge in arms 2-3, measuring approach_commit recovery against the EXQ-483a all-zero baseline) is deferred to a successor EXQ once V3-EXQ-490 (MECH-269b) lands. The combined V_s rollout gating + drive -> liking -> approach bridge stack will be tested together; the Q-040 factorial structure already cross-points to MECH-295 as the dominant blocker if MECH-269b alone fails to recover approach.

Biological grounding

Substrate evidence and architectural articulation, anchored in the lit-pull (evidence/literature/targeted_review_mech295_liking_approach_bridge/):

  • Pecina & Berridge 2005 (J Neurosci, doi:10.1523/JNEUROSCI.3331-05.2005) – anatomical hedonic hotspot in NAc shell.
  • Castro & Berridge 2014 (Phys Behav, doi:10.1016/j.physbeh.2014.02.063) – decade of hotspot mapping consolidated.
  • Smith Berridge & Aldridge 2011 (PNAS, doi:10.1073/pnas.1101920108) – VP single-unit recording: drive change recodes palatability before cue firing. The strongest direct mechanistic anchor.
  • Berridge & Kringelbach 2015 (Neuron, doi:10.1016/j.neuron.2015.02.018) – architectural articulation of liking as a necessary input to motivated approach, not a passive readout.
  • Dickinson & Balleine 1994 (Animal Learn Behav, doi:10.3758/BF03209866) – foundational behavioural demonstration: drive-state shifts do not directly rewire action selection; they must route through experienced hedonic value of the outcome.
  • Pecina et al. 2003 (J Neurosci, doi:10.1523/JNEUROSCI.23-28-09395.2003) – DAT-knockdown: more wanting, unchanged liking. The structural caution that constrains MECH-295 to the WEAK reading.

The cleanest direct test of MECH-295’s necessity claim (mu-opioid antagonist in NAc shell hotspot + hunger + cue-approach) does not exist in the literature in a single paper. This is an experimental gap that the V3 substrate is designed to probe directly.

  • SD-012 – homeostatic drive (the input).
  • SD-014 – valence vector substrate (VALENCE_LIKING component is the write target).
  • SD-015 – z_resource encoder (upstream of GoalState.update; goal-congruent surface depends on this seeding).
  • SD-016 – frontal cue-indexed integration (cue_action_proj reads z_world; complementary to MECH-295 at the cue side but not redundant).
  • MECH-117 – existing wanting/liking dissociation in REE (benefit_eval_head vs z_goal_latent). MECH-117 specifies the dissociation; MECH-295 specifies the necessity-chain that must hold for behavioural output.
  • ARC-036 – hedonic hotspot anatomical substrate (prerequisite for the architectural claim).
  • MECH-094 – call-site scoping + simulation_mode argument honour.
  • SD-037 – broadcast override regulator. The bridge consumes effective_drive after SD-037’s override_goal_seeding_gain amplification has already been applied in update_z_goal().
  • MECH-269b – complementary candidate cause for EXQ-483 wired-but-inert. Q-040 factorial points evidence at MECH-295 if MECH-269b alone fails to recover approach_commit.

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