Serotonergic Cross-State Substrate Requirements

Created: 2026-04-06 Claims: MECH-203 (replay salience tagging), MECH-204 (REM gate zero-point) Status: design — pre-implementation

Why This Blocks Hippocampal and Sleep Functioning

MECH-203 and MECH-204 are currently V4 scope. But SD-017 (minimal sleep-phase infrastructure) is V3. This creates a dependency problem:

Without MECH-203, SWS replay (MECH-121) has no benefit-salience signal. Replay will be entirely harm-salience-driven (from the residue field). Consequence:

  • Hippocampal map consolidation is biased toward threat content
  • Benefit-relevant experiences are under-consolidated
  • z_goal representations degrade overnight even if they were seeded during waking
  • This is the depressive consolidation asymmetry — but it will occur in ALL agents, not just those with 5-HT deficit, because the tagging mechanism is absent entirely

Without MECH-204, REM precision recalibration (MECH-123) has no principled zero-point. The system recalibrates against… what? Without 5-HT withdrawal defining “neutral”, the precision system has no reference. Consequence:

  • Precision recalibration is undefined or arbitrary
  • The precision system cannot distinguish “benefit-gradient-active” from “no-gradient”
  • INV-045 phase ordering constraint has no implementation mechanism

The blocking dependency chain:

SD-017 (V3: minimal sleep phases)
  |
  +-> MECH-121 (V4: SWS replay) needs MECH-203 for balanced replay content
  |     Without it: harm-biased consolidation in ALL agents
  |
  +-> MECH-123 (V4: REM recalibration) needs MECH-204 for calibration target
  |     Without it: undefined recalibration reference
  |
  +-> INV-045 (V4: phase ordering) needs MECH-204 for phase boundary mechanism
        Without it: no implementation of SWS-to-REM transition trigger

Recommendation: Pull at least the substrate primitives for MECH-203/204 into V3 scope, even if the full mechanisms remain V4. The V3 sleep system needs these hooks to function correctly, even in minimal form.

Substrate Requirements

SR-1: Tonic 5-HT State Variable

A slow-accumulating scalar tonic_5ht that tracks the agent’s serotonergic tone.

Waking dynamics:

  • Increases slowly when benefit_exposure > 0 (resources encountered)
  • Decays slowly toward a baseline when no benefits occur
  • Suppressed by chronic high z_harm_a (MECH-186 interaction)
  • Range: [0.0, 1.0] where 1.0 = fully supported benefit gradient

Sleep dynamics:

  • During SWS: held at waking level (5-HT is active during SWS)
  • During REM: drops to 0.0 (dorsal raphe quiescence, MECH-204)
  • On wake: restored to pre-sleep level

Implementation sketch:

# In agent state or a new SerotoninModule
tonic_5ht: float = 0.5  # baseline
5HT_RISE_RATE = 0.01    # per benefit event
5HT_DECAY_RATE = 0.001  # per step without benefit
5HT_HARM_SUPPRESS = 0.1 # suppression rate when z_harm_a elevated

This is architecturally analogous to drive_level (SD-012) but for the benefit gradient rather than the energy/need axis.

SR-2: Experience Benefit-Salience Tagging

When experiences are stored (for potential replay), tag each with:

benefit_salience = tonic_5ht * benefit_exposure_at_time

This creates a replay priority signal complementary to the existing harm salience from the residue field. The hippocampal replay system (MECH-121) should then prioritize by:

replay_priority = max(harm_salience, benefit_salience)
# Or: weighted combination allowing balanced consolidation
replay_priority = alpha * harm_salience + (1 - alpha) * benefit_salience

Where to store the tag:

  • Option A: In the ThetaBuffer experience entries (alongside z_world, z_harm, action)
  • Option B: In the ResidueField valence vector (SD-014 already has VALENCE_WANTING component)
  • Option C: New field in HippocampalModule’s trajectory proposals

Option B is most natural — SD-014’s valence_vecs already record [wanting, liking, harm_discriminative, surprise] per RBF center. benefit_salience could modulate the VALENCE_WANTING component or be a 5th dimension.

SR-3: REM Zero-Point Mechanism

During the REM phase (when offline_mode=True and phase=REM):

  1. Set tonic_5ht = 0.0 (withdrawal)
  2. Record the current precision state as precision_at_zero
  3. Recalibrate precision priors against this neutral reference

Implementation requires:

  • A precision tracking mechanism (currently E3-derived but not explicitly accumulated)
  • A way to compare “precision under benefit gradient” vs “precision without gradient”
  • The difference defines what to recalibrate

This is more complex than SR-1/SR-2 and may genuinely be V4. But SR-1 and SR-2 can be implemented in V3 as hooks that the sleep system uses.

SR-4: Phase Boundary Trigger

The SWS-to-REM transition could be triggered by:

if sws_consolidation_complete():  # e.g., replay_buffer exhausted or convergence metric
    tonic_5ht = 0.0  # begin withdrawal
    enter_rem_phase()

The serotonergic state transition IS the phase boundary. This connects the pharmacological claims (SSRIs affect tonic_5ht -> affect tagging -> affect consolidation) to the sleep architecture claims (phase ordering requires neuromodulatory state transitions).

V3 vs V4 Scope Decision

Substrate Complexity V3 Necessity Recommendation
SR-1: tonic_5ht variable Low High (without it, all replay is harm-biased) V3
SR-2: benefit_salience tag Medium High (replay prioritization needs both signals) V3
SR-3: REM zero-point High Medium (REM can use simpler proxy in V3) V4
SR-4: Phase boundary trigger Medium Medium (V3 uses fixed alternation) V4

Minimum V3 pull: SR-1 + SR-2. Add tonic_5ht as a state variable and tag experiences with benefit_salience. SWS replay uses both harm and benefit salience for prioritization. REM and phase boundary remain V4.

Connection to Existing Substrate

Existing Component Connection Point
GoalConfig.valence_wanting_floor (MECH-186) tonic_5ht modulates this floor
GoalConfig.z_goal_seeding_gain (MECH-187) tonic_5ht IS the gain signal source
ResidueField.valence_vecs (SD-014) benefit_salience stored in VALENCE_WANTING
ThetaBuffer.consolidation_summary() benefit_salience weights replay selection
Agent.enter_offline_mode() triggers tonic_5ht dynamics switch
E1.shy_normalise() (MECH-120) operates during SWS (tonic_5ht active)
Agent.compute_drive_level() (SD-012) drive_level and tonic_5ht are independent axes

The key insight: valence_wanting_floor, z_goal_seeding_gain, and z_goal_inject are already in the substrate as config parameters. But they are currently STATIC — set once at experiment start. MECH-203/204 require them to be DYNAMIC, modulated by tonic_5ht which itself evolves over time and across wake/sleep states.

The minimum change: make z_goal_seeding_gain a function of tonic_5ht rather than a static config parameter. This single change connects the waking serotonergic system to the sleep consolidation system.

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