SD-035: Amygdala Analogue — BLA + CeA Salience Stage

Claim ID: SD-035 (substrate) + MECH-046 (CeA mode prior) + MECH-074 parent + MECH-074a-d (split children) Subject: amygdala.analog_substrate Status: candidate, v3_pending — NOT YET IMPLEMENTED Registered: 2026-04-21 Depends on: SD-011, SD-032a, ARC-005, ARC-007, MECH-039 Paired with: SD-032 (cingulate integration) — SD-035 is the missing upstream stage that feeds SD-032a

Problem

ree-v3 produces z_harm_a via AffectiveHarmEncoder (ree-v3/ree_core/latent/stack.py:168) and consumes it downstream in the cingulate cluster (SD-032a–e). The partial urgency interrupt at ree-v3/ree_core/agent.py:869-886 gives us a CeA-like fast pulse, and AICAnalog (SD-032c) gives us cortical urgency. But the explicit amygdala claims — MECH-046 (fast salience classification writes a mode prior) and MECH-074 (amygdala as read/write head for valenced hippocampal map) — have no dedicated substrate.

Before SD-035, MECH-046/074 were promised-but-unimplemented. The cingulate is downstream; it consumes what the amygdala writes. Without an upstream amygdala stage:

  • There is no BLA-style encoding-gain multiplier on hippocampal writes under arousal.
  • There is no content-selective retrieval bias on hippocampal readout — only a scalar-gain approximation would fit downstream, which is the named failure signature LaBar & Cabeza 2006 rule out.
  • There is no remap signal emitted on harm-PE spikes; MECH-073’s valenced map cannot reorganise under threat prediction violation.
  • The fast subcortical priming channel (MECH-074c) is conflated with the cortical urgency interrupt, obscuring both.
  • The CeA mode-prior write (MECH-046) has nothing that produces it; SalienceCoordinator has an unfilled peer slot.

The lit-pull (evidence/literature/targeted_review_amygdala_analog/synthesis.md, 8 entries across targeted_review_connectome_mech_046 and targeted_review_connectome_mech_074) resolves the architecture:

  • BLA and CeA are biologically distinct subdivisions with distinct jobs. Collapsing them is the same error as collapsing the cingulate into a single hub. The correct form is two peer modules.
  • Retrieval bias must be content-selective (per-trace weight vector), not scalar. Scalar is a named failure signature — it cannot reproduce the central/peripheral dissociation that is the hallmark of amygdala-lesion studies (LaBar & Cabeza 2006).
  • Remap requires predictor attribution, not broadcast. The contextual-vs-auditory dissociation in Moita 2004 (Z = -1.36 vs -0.34, p = 0.02) rules out a broadcast architecture.
  • Fast subcortical route is real (~75 ms) but bounded (Méndez-Bértolo 2016 + Pessoa & Adolphs 2010). CeA biases a coordinator that also takes cortical input — never unilateral, always overridable.

BLA / CeA split rationale

Module Biological analogue Inputs Outputs Owns
BLAAnalog Basolateral amygdala z_harm_a, z_harm_a_pred, hippocampal_context encoding_gain (scalar), retrieval_bias (per-trace weight vector), remap_signal (per-code binary + strength) MECH-074a, MECH-074b, MECH-074d
CeAAnalog Central amygdala z_harm_a, cue_features=None, escapability_hint=None mode_prior (vector over operating modes; pre-softmax additive log-odds), fast_prime (scalar action/candidate prior) MECH-046, MECH-074c

Why two modules, not one:

  1. Distinct latencies. CeA fires within 1–2 sim steps (~75 ms biological); BLA operates on slower encoding/retrieval windows (minutes to days). A single tick function would have to gate both behaviours on different time constants — clearer as two modules.
  2. Distinct targets. CeA writes to SalienceCoordinator (and indirectly to candidate generation via fast_prime). BLA writes to HippocampalModule (encoding strength, retrieval weights, remap). Collapsing them forces a crossbar that does not exist biologically and invites contamination.
  3. Distinct failure modes. Falsification signatures for MECH-074a (encoding gain) and MECH-074d (remap) test BLA arithmetic. Falsification for MECH-046 and MECH-074c test CeA arithmetic. Splitting preserves the ability to ablate each subdivision independently in EXQs.
  4. Mirrors SD-032’s pattern. SD-032 split cingulate into five subdivisions (a–e) each as its own module. SD-035 follows the same discipline at smaller scale: two subdivisions, two modules.

Subdivisions

BLAAnalog — MECH-074a/b/d

Non-trainable arithmetic. File: ree-v3/ree_core/amygdala/bla.py (NEW).

Interface:

BLAAnalog.tick(
    z_harm_a: Tensor,
    z_harm_a_pred: Tensor,
    hippocampal_context: HippocampalContext,
) -> BLAOutput
# BLAOutput fields:
#   encoding_gain: float              (MECH-074a)
#   retrieval_bias: Tensor[num_traces]   (MECH-074b; per-trace weights, NOT scalar)
#   remap_signal: Tensor[num_codes]   (MECH-074d; per-code binary with strength)

Encoding gain (MECH-074a). Inverted-U over arousal magnitude with exponential decay post-event. Defaults from Roozendaal & McGaugh 2011:

  • encoding_gain_max = 2.5 at ||z_harm_a|| ~ 0.7
  • encoding_gain_floor = 1.0 below ||z_harm_a|| = 0.4
  • post-event window: 18000 sim steps (~30 min biological)
  • half-life: 3600 sim steps (~6 min biological)

Multiplies HippocampalModule write strength at encoding time.

Retrieval bias (MECH-074b). Per-trace weight vector w_i = 1 + alpha * arousal_tag_i, with alpha in [0.3, 1.0]. Requires HippocampalModule to carry a per-trace arousal_tag field populated at encoding time. Non-transient: BLA contribution grows with trace age (amygdala-MTL connectivity increases from 20 min to 1 week, LaBar & Cabeza 2006).

Remap signal (MECH-074d). Fires when ||z_harm_a - z_harm_a_pred|| exceeds ~1 SD of the running harm-PE distribution AND a predictor-attribution head flags at least one candidate latent code. Both conditions required. Binary per-code shape; partial (~1/3) remap amplitude (Moita 2004).

CeAAnalog — MECH-046, MECH-074c

Non-trainable arithmetic. File: ree-v3/ree_core/amygdala/cea.py (NEW).

Interface:

CeAAnalog.tick(
    z_harm_a: Tensor,
    cue_features: Optional[Tensor] = None,
    escapability_hint: Optional[Tensor] = None,  # Q-036 placeholder (no-op)
) -> CeAOutput
# CeAOutput fields:
#   mode_prior: Tensor[num_modes]   (MECH-046; pre-softmax additive log-odds)
#   fast_prime: float               (MECH-074c; scalar candidate-prior pulse)

Mode prior (MECH-046). Pre-softmax additive log-odds bias on SalienceCoordinator’s harm-related mode channels. Gated on ||LowFreq(z_harm_a)||_1 > theta_cea_fast. Emits within 1–2 sim steps (~75 ms, Méndez-Bértolo 2016) of threshold crossing. Cortical AIC/dACC discriminate at ~5–10 sim steps (~400 ms); fast:slow ratio ~5:1.

Fast prime (MECH-074c). Distinct from mode_prior: a scalar pulse biasing candidate generation directly (short subcortical route, not a cortical mode write). Magnitude ceiling |fast_prime| <= max cortical AIC/dACC log-odds adjustment. If cortical confirmation is absent within 5–10 sim steps, fast_prime decays with tau_decay in [3, 5] sim steps toward baseline.

Selectivity constraint. CeA must fire on harm-affective valence, not generic arousal. Named failure: if CeA fires on arousing non-threat stimuli (e.g., arousing scenes in Méndez-Bértolo’s calibration), fast_prime is a generic salience detector, not a CeA analogue.

Config surface

Lives in ree-v3/ree_core/utils/config.py. Master switch REEConfig.use_amygdala_analog=False gates both modules (defaults make them no-ops — backward-compat requirement).

@dataclass
class BLAConfig:
    enabled: bool = False                          # master switch
    encoding_gain_max: float = 2.5                 # Roozendaal 2011
    encoding_gain_arousal_threshold: float = 0.4
    encoding_gain_arousal_peak: float = 0.7
    encoding_gain_window_steps: int = 18000
    encoding_gain_half_life_steps: int = 3600
    retrieval_bias_alpha: float = 0.6              # midpoint of 0.3-1.0
    retrieval_bias_compensation: float = 0.0       # start at 0; enable 0.1-0.3 later
    retrieval_tag_at_encoding: bool = True         # required per LaBar 2006
    remap_pe_sigma_threshold: float = 1.0          # units of running PE stdev
    remap_code_fraction: float = 0.33              # Moita 2004
    remap_requires_attribution: bool = True        # Moita 2004 attribution gate

@dataclass
class CeAConfig:
    enabled: bool = False                          # master switch
    fast_route_threshold: float = 0.5              # theta_cea_fast on LowFreq(z_harm_a)
    fast_route_input_is_lowfreq: bool = True       # Mendez-Bertolo 2016
    mode_prior_log_odds_max: float = 0.8           # bounded below AIC/dACC ceiling
    fast_prime_decay_tau_steps: int = 4            # 3-5 range
    fast_prime_override_window_steps: int = 8      # 5-10 range; after, decays
    pre_softmax_additive: bool = True              # NOT post-softmax multiplicative

Data flow 

  AffectiveHarmEncoder ──► z_harm_a ──┬──► BLAAnalog.tick ──► BLAOutput
  (ree_core/latent/stack.py)          │                        │
                                      │                        ├─► encoding_gain  ───► HippocampalModule write
                                      │                        ├─► retrieval_bias ───► E3.select candidate reweighting
                                      │                        └─► remap_signal   ───► HippocampalModule update hook
                                      │
                                      └──► CeAAnalog.tick ──► CeAOutput
                                                               │
                                                               ├─► mode_prior     ───► SalienceCoordinator (pre-softmax additive)
                                                               └─► fast_prime     ───► candidate generation bias
  • agent.sense() ticks BLAAnalog and CeAAnalog after z_harm_a is produced; outputs are attached to LatentState.
  • SalienceCoordinator.tick accepts mode_prior from CeA and adds to pre-softmax logits (reuses existing API; mirror SD-032c AICAnalog wiring).
  • HippocampalModule write path multiplies write strength by BLA encoding_gain; accepts remap_signal on its existing update hook.
  • E3.select accepts BLA retrieval_bias to reweight hippocampal candidate proposals.
  • Urgency interrupt at agent.py:869-886 stays inline this pass (deliberate; migration into CeAAnalog.fast_prime is a second pass once MECH-074c is validated).

Biological grounding

Full evidence records at evidence/literature/targeted_review_amygdala_analog/synthesis.md. Load-bearing papers:

Role Paper Confidence Claim
BLA encoding gain Roozendaal & McGaugh 2011 0.82 MECH-074a
BLA retrieval bias LaBar & Cabeza 2006 0.72 MECH-074b
BLA remap / reconsolidation Nader, Schafe & LeDoux 2000 0.80 MECH-074d
BLA remap / place-cell dissociation Moita et al 2004 0.78 MECH-074d (attribution gate)
CeA fast route (latency) Méndez-Bértolo et al 2016 0.88 MECH-046, MECH-074c
CeA fast route (bounds) Pessoa & Adolphs 2010 0.72 MECH-046, MECH-074c
BLA canonical (pre-existing) McGaugh 2004 0.82 MECH-074 parent
BLA retrieval canonical (pre-existing) Dolcos et al 2004 — MECH-074b

The biology-before-formal-definitions rule (MEMORY feedback_biology_before_formal_definitions.md) was applied: SD-035 and MECH-074a-d were registered only after the lit-pull extracted quantitative defaults. Config values are seeded from that synthesis, not chosen a priori.

Falsification signatures

Per sub-claim

MECH-074a (encoding gain): Falsified if EXQ-B shows threat-context recall does not improve under BLA-modulated gain relative to gain=1, OR if neutral recall is harmed by gain>1. The inverted-U and 30-min window are the specific parametric commitments; a flat or monotone gain surface also falsifies.

MECH-074b (retrieval bias): Falsified if central/gist threat-associated items are NOT retrieved preferentially over peripheral/neutral items under BLA retrieval-bias ON, or if a uniform retrieval boost across all traces reproduces the same behaviour (collapses to scalar form).

MECH-074c (CeA fast prime): Falsified if fast_prime emits later than ~2 sim steps, OR if it fails to decay under cortical non-confirmation within the override window, OR if it fires on arousing-but-neutral stimuli (selectivity fail — indicates a generic salience detector, not a CeA analogue).

MECH-074d (remap with attribution): Falsified if remap_signal fires on sub-threshold PE, OR if it perturbs untagged codes uniformly (attribution gate broken), OR if remap amplitude is wholesale (closer to 100% than 33%).

MECH-046 (CeA mode prior): Falsified by EXQ-A if CeA-ON does not beat CeA-OFF on mode-switch latency, OR if the zeroed-prior control does not match CeA-OFF (which would show the effect runs through some pathway other than the mode prior).

Cross-cutting (SD-035 substrate)

Backward-compat signature: With use_amygdala_analog=False, one existing EXQ (e.g. baseline CausalGridWorld run) MUST produce bit-identical metrics to the pre-SD-035 codebase. Any drift falsifies the no-op guarantee.

Dominance signature: CeA fast_prime and mode_prior magnitudes must stay ≤ cortical AIC/dACC ceilings. If any EXQ shows CeA outputs dominating cortical inputs (i.e., SalienceCoordinator switching modes regardless of cortical state when CeA fires), the magnitude ceilings are miscalibrated — Pessoa & Adolphs 2010 “many roads” framing is violated.

Selectivity signature: A calibration run on arousing-but-neutral stimuli (analogue of Méndez-Bértolo’s non-face arousing scenes) must show CeA NOT firing. If it does, CeA is a generic arousal detector.

Implementation plan

  1. Configs — BLAConfig, CeAConfig, use_amygdala_analog master switch in ree-v3/ree_core/utils/config.py.
  2. Modules — ree-v3/ree_core/amygdala/bla.py and ree-v3/ree_core/amygdala/cea.py.
  3. LatentState extension — add optional bla_out, cea_out fields in ree_core/latent/stack.py; tick both in agent.sense() after z_harm_a.
  4. Wiring — SalienceCoordinator accepts mode_prior; HippocampalModule accepts encoding_gain and remap_signal; E3.select accepts retrieval_bias; HippocampalModule grows a per-trace arousal_tag field (MECH-074b requirement — flagged as a broader change).
  5. Backward-compat smoke — replay one existing EXQ with flag OFF; confirm bit-identical metrics.
  6. Activation smoke — with flag ON and lit-pull defaults, confirm CeA mode_prior shifts under threat cue, BLA encoding_gain > 1 during threat, BLA remap_signal fires on synthetic PE spike.
  7. CLAUDE.md entry in ree-v3/CLAUDE.md (mirror SD-032c format).
  8. Experiments — queue EXQ-A (CeA mode-prior ablation, 3 arms) and EXQ-B (BLA encoding + remap, 2 arms). MECH-074b retrieval-bias and MECH-074c fast-prime get follow-up EXQs once A/B pass.

Q-036 escapability follow-up

CeAAnalog carries an escapability_hint: Optional[Tensor] = None parameter as a no-op placeholder. The parameter is accepted by tick() but the initial non-trainable arithmetic does not consume it. This is deliberate: when MECH-219 / Q-036 (escapability signal) lands, it can be wired in without a module-interface refactor.

Rationale for placeholder rather than full integration:

  • Q-036 is not a blocker for MECH-046 / MECH-074a-d. Blocking SD-035 on it would stall the amygdala work indefinitely.
  • The parametric form of escapability’s contribution to CeA priming is not yet resolved (would it gate fast_prime amplitude? shift mode_prior away from defensive modes? both?). Better to defer than to guess.
  • Adding the parameter now, as a no-op, costs nothing and prevents a breaking change later.

Follow-up trigger: when Q-036 is promoted to a mechanism claim with a proposed parametric form, a follow-up EXQ (TBD) will add the escapability pathway and measure behavioural differences under controllable vs uncontrollable threat.

Open risks / follow-ups

  1. Per-trace arousal_tag in HippocampalModule. MECH-074b requires HippocampalModule to carry a per-trace arousal_tag field that BLAAnalog writes at encoding and reads at retrieval. This is a broader wiring change than a simple multiplier. If HippocampalModule does not currently carry such a field, adding it is a sub-task of Step 4.

  2. Predictor-attribution head. MECH-074d requires a predictor-identification step before emitting remap_signal. Initial pass approximates by picking codes whose contribution to the harm-PE exceeds a threshold; a learnable attribution head is a deliberate second pass.

  3. Cortical-ceiling calibration. CeA fast_prime magnitude ceiling must stay ≤ max AIC/dACC log-odds adjustment. Read the current AIC ceiling from SalienceCoordinatorConfig and set mode_prior_log_odds_max accordingly; do not guess.

  4. SD-032b dACC FAIL. v3_exq_445 shows the affective PE path is currently underperforming. EXQ-A/B acceptance criteria must be calibrated against this — don’t expect the mode-prior ablation to produce clean signals if upstream PE is broken. Flagged as a prerequisite risk, not a blocker.

  5. Learnable fear-conditioning head. The initial pass is non-trainable arithmetic. A learnable fear-conditioning head (threat-cue → z_harm_a amplification with experience) is explicitly a second pass.

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