Psychiatric Failure Modes as Architectural States

Emergent from EXQ-237a LONG_HORIZON condition. Seed=42. First observed 2026-04-06.

Overview

The REE architecture produces psychiatric phenomenology as emergent failure modes, not by design. When specific architectural components fail or are absent, the resulting behavioural patterns correspond structurally to recognised clinical syndromes. This document catalogues the 2x2 motivational state taxonomy, the three missing serotonergic mechanisms that define the V4 treatment model, and the predictions that follow.

The key insight from EXQ-237a is that the LONG_HORIZON condition (3 hazards, 1 resource, 150 steps) produced goal_norm=0.004 in the PLANNED condition, with resource_rate and harm_rate identical between HABIT and PLANNED modes. The goal-directed planning system was structurally intact, but behaviourally silent – the environment had collapsed the motivational terrain that planning requires.

2x2 Psychiatric State Taxonomy

The motivational state of the agent at any time can be characterised by two independent variables: the level of z_harm_a (aversive arousal) and the status of z_goal (wanting signal).

  z_goal INTACT z_goal ABSENT
z_harm_a ELEVATED GAD-like state Anxious depression (INV-053 attractor)
z_harm_a LOW Healthy goal-directed behaviour Melancholic depression (burnt-out)

Cell descriptions

Healthy (z_harm_a low, z_goal intact) Normal goal-directed behaviour. Planning and habitual modes diverge: PLANNED achieves higher resource_rate, lower harm_rate. The architecture functions as designed.

GAD-like (z_harm_a elevated, z_goal intact) Goals exist but are repeatedly interrupted by threat responses. z_harm_a fires during goal pursuit, triggering avoidance that displaces goal-directed sequences before completion. The agent wants but cannot sustain action sequences long enough to succeed. Anxiety without motivational collapse: the planning system is engaged but constantly derailed.

Anxious depression (z_harm_a elevated, z_goal absent) – INV-053 attractor This is the state observed in EXQ-237a LONG_HORIZON. Chronic harm exposure has suppressed benefit_exposure below the SD-012 seeding threshold. z_goal cannot form. The VALENCE_WANTING terrain in the residue field has collapsed to near-zero. HABIT and PLANNED modes are behaviourally identical because the planning system has no motivational signal to act on. Aversive arousal is elevated (harm_rate=0.14) but wanting is absent (goal_norm=0.004). This corresponds structurally to anxious depression: the patient is activated (aroused, vigilant, distressed) but cannot initiate or sustain goal pursuit.

Melancholic depression / burnt-out attractor (z_harm_a low, z_goal absent) A further evolution of the anxious depression attractor: z_harm_a has habituated to the chronic aversive environment, leaving flat affect (low aversive arousal) combined with absent wanting. The agent is neither distressed nor motivated. This corresponds to melancholic depression: psychomotor retardation, anhedonia, flat affect, absent initiative. Both depressive cells share the z_goal=0 foundation; they differ in whether aversive arousal remains active.

The Three-Stage Motivational Pipeline

Goal-directed behaviour in aversive environments requires tonic regulation at three stages (INV-052):

Stage 1: TERRAIN         Stage 2: TRANSDUCTION    Stage 3: MAINTENANCE
VALENCE_WANTING floor    benefit -> z_goal gain    z_goal persistence
(MECH-186)               (MECH-187)                (MECH-188)

Stage 1 – Terrain maintenance (MECH-186)

A slow-accumulating tonic signal (serotonergic analog) maintains a non-zero floor on VALENCE_WANTING entries in the residue field even when phasic benefit events are sparse. Without this, a chronic harm-dominated environment suppresses the benefit gradient until it is flat. The agent has no terrain to navigate toward resources.

Failure mode: benefit terrain collapse. Agent cannot orient toward goals because the residue field contains no WANTING gradient. HABIT and PLANNED modes become equivalent.

Stage 2 – Gain regulation (MECH-187)

The serotonergic system modulates the gain on the benefit->z_goal transduction function. Currently SD-012 implements a hard threshold: benefit_exposure must exceed a floor to produce any z_goal seeding. MECH-187 converts this to an adaptive graded function.

Failure mode: anhedonia as wanting failure. Benefit events occur (liking is intact), but z_goal is not seeded (gain is suppressed). This is the Berridge dissociation (MECH-117): liking preserved, wanting abolished. The agent can enjoy resources when it encounters them reactively but does not seek them.

Stage 3 – Goal persistence (MECH-188)

Once z_goal has been seeded, a PFC-mediated working memory mechanism (modulated by 5-HT projections from DRN to dlPFC) maintains the representation through periods of elevated harm or temporarily reduced benefit exposure.

Failure mode: goals form but cannot be held. z_goal collapses at the first harm event. This is the “cannot hold a plan” phenomenology: the patient articulates goals but fails to sustain them against distraction, threat, or delay. Architecturally distinct from the GAD-like state (where goals are present but interrupted by z_harm_a) because here z_goal itself collapses, not just execution sequences.

The Depressive Maintenance Loop (INV-054)

Once the INV-053 attractor is established, it is self-maintaining:

VALENCE_WANTING terrain collapses
         |
         v
Benefit-oriented exploration reduces (no gradient to follow)
         |
         v
Benefit encounters reduce
         |
         v
benefit_exposure stays below SD-012 threshold
         |
         v
z_goal seeding remains blocked
         |
         v
VALENCE_WANTING terrain cannot recover  <-- closed loop

This closed negative feedback loop predicts:

  1. Chronicity: the depressive attractor is self-maintaining once established; environmental improvement alone may be insufficient.

  2. Phase transition recovery: the transition from depressive attractor to normal function requires a threshold crossing (sufficient benefit_exposure to cross the SD-012 seeding threshold), not a gradual linear improvement. This maps to the clinical observation that antidepressant response is often non-linear.

  3. Behavioural activation as first-line treatment: externally scaffolding benefit encounters (behavioural activation therapy) directly targets the maintenance loop by restoring benefit_exposure above threshold, rebuilding the terrain gradient, allowing z_goal to re-form.

Q-034: The Hazard/Resource Ratio Prediction

EXQ-237a revealed the depressive attractor by varying environment structure (3 hazards, 1 resource). Q-034 asks whether the relevant parameter is the ratio of harm exposure to benefit exposure opportunity, not absolute stress intensity.

Prediction: a fixed hazard_harm level with sufficient resource density never produces the INV-053 attractor; the same hazard_harm with insufficient resources does.

Testable: CausalGridWorldV2, hold hazard_harm constant, vary num_resources (or hold num_hazards constant, vary num_resources). Sweep the ratio.

Implications if confirmed:

  • Depression is a resource-availability disorder as much as a stress disorder
  • The same stressor produces depression or not depending on available positive affordances
  • Interventions targeting benefit availability (not just stress reduction) are therapeutic
  • Environmental enrichment (adding positive affordances) has mechanistically distinct effects from anxiolytic treatment

Frame-Tag Failure Modes: Developmental Etiology (INV-061, MECH-200, MECH-201, MECH-202)

Registered 2026-04-06. Extends MECH-094 (hypothesis tag loss = psychosis mechanism) with a developmental pathway via MECH-198 (pretend play as frame-distinction training).

Common Architectural Substrate (INV-061)

The REE architecture distinguishes real from synthetic/hypothetical frames via two co-operating tag systems: the hypothesis tag (MECH-094, agent-internal, marks simulation/replay as non-real) and the play frame tag (ARC-049, bilateral, marks play episodes as authorized-synthetic). Pretend play (MECH-198) is the developmental phase where these two systems first co-operate.

When pretend play is absent or insufficient, the frame-distinction mechanism is undertrained. The resulting adult failure modes are not independent disorders but different failure signatures of the same undertrained mechanism, varying by:

  1. Direction of confusion: real→synthetic (derealization cluster) vs synthetic→real (delusion cluster)
  2. Affected subsystem: frame tag, commitment gate, or threat evaluation

This provides a unified developmental etiology for conditions that are clinically distinct but architecturally related.

MECH-200: Real→Synthetic Confusion (Derealization Direction)

The agent treats real consequences as if they were synthetic. The frame tag reads “play” when the environment is real.

Mechanism: the real/synthetic distinction was never properly calibrated during pretend play. The agent defaulted to tagging ambiguous frames as synthetic (the safer error during play, where real consequences are absent). In adulthood, this default persists: when frame signals are ambiguous, the agent discounts real harm signals as if they were play-harm.

Behavioral signatures:

  • Actions under real-harm conditions with play-appropriate (i.e. absent) defensive responses
  • Commitment without adequate threat evaluation
  • Harm signals discounted or experienced as distant/unreal

Psychiatric mappings:

Signature Clinical analogue REE mechanism
Real harm experienced as unreal Derealization Frame tag stuck on “synthetic”; z_harm_s fires but is discounted
Elevated commitment, reduced consequence-sensitivity Manic episodes Play frame leaks into real operation; commitment gate fires without harm-gate check
Actions feel consequence-free Dissociative states Frame tag intermittently flips to “synthetic” under stress

MECH-201: Synthetic→Real Confusion (Delusion Direction)

The agent treats internally generated or simulated representations as real. The hypothesis tag (MECH-094) fails to suppress real-consequence responses during simulation, replay, or imagination.

Mechanism: during pretend play, the hypothesis tag must be active (the stick is not really a sword) while the play tag is also active (the fight is not really dangerous). If pretend play was insufficient, the hypothesis tag was never properly co-exercised with the play tag. In adulthood, internally generated representations — simulation, replay, imagination — lack the hypothesis tag marker and are processed through the real-consequence pipeline.

Behavioral signatures:

  • Internally generated scenarios trigger full emergency responses
  • Replay episodes experienced as happening now, not as memory
  • Pre-commit simulation generates real threat responses

Psychiatric mappings:

Signature Clinical analogue REE mechanism
Internal representations experienced as external Delusion formation Hypothesis tag absent; E3 treats simulated trajectories as real percepts
Replay triggers full emergency response PTSD flashback SWR replay (MECH-121) lacks hypothesis tag; experienced as real re-occurrence
Simulation generates real threat Anxiety disorders Pre-commit simulation (normally hypothesis-tagged) triggers z_harm_a as if real
Imagined social scenarios feel real Paranoid ideation Simulated other-agent threat (MECH-127 counterfactual) processed without hypothesis tag

MECH-202: Commitment Gate Developmental Failure

The commitment gate (E3 commit boundary) is undertrained because it was never exercised in synthetic mode during pretend play. Two failure directions:

Direction A — Insufficient inhibition (impulsivity): The gate fires without completing E3 evaluation. During pretend play, the gate would have been exercised in a low-stakes context where premature commitment has no real consequences, allowing calibration. Without this calibration, the gate threshold is set too low.

Signature Clinical analogue REE mechanism
Commits without evaluation Impulsivity Commitment gate fires before E3 harm/benefit evaluation completes
Elevated commitment + reduced harm sensitivity Mania (commitment component) Gate threshold too low + MECH-200 frame confusion compound
Cannot delay gratification Impulse control disorders Gate does not wait for E2 rollout to complete

Direction B — Excessive inhibition (paralysis): The gate never fires even when evaluation is complete. Without pretend play calibration, the gate threshold was never lowered from the infant default (very high commit threshold, INV-055). The agent remains in perpetual pre-commit evaluation.

Signature Clinical analogue REE mechanism
Endless re-evaluation without commitment OCD (checking/rumination) Gate threshold never calibrated down from infant default; E3 loops without committing
Complete inability to initiate action Catatonia Gate frozen at maximum threshold
Decision paralysis despite completed evaluation Severe indecision Commitment gate signal never reaches threshold

Developmental Evidence

The frame-confusion etiology predicts that pretend play deficits in early childhood should correlate with later frame-distinction failures. Existing evidence is consistent:

  • Reduced pretend play predicts later psychotic symptoms (Cannon et al. 2002)
  • Children with autism show reduced pretend play AND elevated psychosis risk
  • Pretend play → theory of mind → reality testing are developmentally linked
  • Imaginative play correlates with later impulse control (Diamond & Lee 2011)

Relationship to MECH-094

MECH-094 identifies hypothesis tag loss as a psychosis mechanism. The frame-confusion cluster (MECH-200/201/202) extends this by explaining WHY the tag system might be fragile: because pretend play (MECH-198) is where the real/synthetic frame distinction is first calibrated. MECH-094 describes the proximal mechanism (tag loss); MECH-198/200/201 describe the developmental etiology (insufficient calibration during childhood).

Claims Covered

ID Label Stage
MECH-186 5HT_tonic.benefit_gradient_maintenance Terrain (Stage 1)
MECH-187 5HT.incentive_salience_gain_regulation Transduction (Stage 2)
MECH-188 5HT_PFC.goal_persistence_under_adversity Maintenance (Stage 3)
INV-052 tonic_regulatory_system_for_goal_persistence All three stages required
INV-053 computational_depression_attractor Anxious depression cell
INV-054 depressive_maintenance_loop Chronicity mechanism
Q-034 hazard_resource_ratio_depression_threshold Open testable question
INV-061 frame_confusion_unified_developmental_etiology All frame-tag failures share common substrate
MECH-200 frame_confusion_real_as_synthetic Derealization, mania, dissociation
MECH-201 frame_confusion_synthetic_as_real Delusions, PTSD flashbacks, anxiety
MECH-202 commitment_gate_developmental_failure Impulsivity, OCD, catatonia (commit-gate paralysis subtype)
SD-036 gabaergic.cross_stream_decay_regulator Catatonia (harm-stream lock-in subtype) – see EXQ-471 exemplar below
MECH-279 pag.gabaergic_freeze_gate Freeze response duration; failure -> sustained immobility

Updated claims:

  • INV-034: added clinical inverse note (goal maintenance failure = depression, EXQ-237a)
  • MECH-117: added clinical note (Berridge dissociation as architectural consequence of MECH-187 failure)
  • MECH-094: developmental etiology added via MECH-198/200/201 (pretend play calibrates the tag system)

Implementation Status

All three serotonergic mechanisms (MECH-186, 187, 188) are V4 scope. They are not present in the current V3 substrate. The depressive attractor state (INV-053) is therefore an emergent property of the V3/current architecture under adversarial environment conditions.

V4 design will need to specify:

  • Tonic accumulator separate from phasic DA signals in the residue field update rule
  • Gain parameter on the SD-012 seeding function, modulated by tonic 5-HT level
  • PFC working memory module for z_goal maintenance between benefit events

Serotonergic Cross-State Architecture (MECH-203, MECH-204)

Registered 2026-04-06. The serotonergic system operates across waking and sleep states. The three waking mechanisms (MECH-186/187/188) are incomplete without their sleep-state counterparts.

The Incomplete Substrate Problem

MECH-186/187/188 model tonic regulatory systems during waking behavior. But serotonin’s architectural role spans the wake-sleep boundary:

  1. During waking: tonic 5-HT maintains the benefit gradient (MECH-186), regulates incentive salience gain (MECH-187), and supports goal persistence (MECH-188).

  2. During SWS: the tonic 5-HT level tags which experiences get prioritized for replay. High-5-HT experiences (benefit-salient) should be replayed alongside harm-salient experiences from the residue field. (MECH-203)

  3. During REM: 5-HT withdrawal (dorsal raphe quiescence) defines the zero-point for precision recalibration. Without this, MECH-123 has no reference state to recalibrate against. (MECH-204)

MECH-203: Serotonergic Replay Salience Tagging

The residue field marks harm-dense experiences for preferential NREM replay (MECH-099, MECH-121). But benefit-salient experiences are equally important for consolidation – particularly missed opportunities (high benefit_exposure, low z_goal_norm) where the agent could have formed a goal but failed to.

Tonic 5-HT level at time of experience serves as the benefit-salience tag:

  • High tonic 5-HT -> experience tagged for benefit-replay
  • Low tonic 5-HT (depressive state) -> benefit-experiences under-tagged

Psychiatric prediction: Depressed agents show reduced benefit-replay density during SWS relative to harm-replay density. This creates a consolidation asymmetry: harm knowledge is maintained (the world remains threatening in memory) while benefit knowledge degrades (opportunities are forgotten). This contributes to the depressive maintenance loop (INV-054) via a sleep-mediated pathway.

MECH-204: Serotonergic REM-Gate Zero-Point

MECH-123 (REM precision recalibration) currently has no specification for how the recalibration target is established. This claim proposes: 5-HT withdrawal during REM actively sets the zero-point by removing the benefit gradient that biases waking precision. The system recalibrates against a neutral reference.

This makes REM recalibration not merely “allowed by” 5-HT withdrawal but “calibrated by” it. Connects to:

  • INV-045 (phase ordering): SWS needs active benefit gradient (for MECH-203 tagging); REM needs gradient absent (for MECH-204 zero-point). The serotonergic state transition implements the SWS-to-REM computational boundary.
  • MECH-178: noradrenergic REM suppression blocks 5-HT withdrawal, disrupting both replay salience (MECH-203) and recalibration zero-point (MECH-204) – a double hit explaining rigidity rather than just reduced recalibration.

Pharmacological prediction: 5-HT agonists administered during REM should impair precision recalibration quality specifically (not SWS consolidation), because they prevent the zero-point from being established.

Architecture note: EXQ-237a LONG_HORIZON condition (3 hazards, 1 resource, 150 steps), seed=42. Full replication pending seeds 7 and 13.

Dream Phenomenology as Diagnostic and Treatment-Response Marker (INV-062, MECH-205–210)

The four-type dream taxonomy (INV-062) follows directly from the offline pipeline architecture and provides a clinically accessible window into which process is running, whether it converged, and whether the pipeline is intact.

The four types and their computational signatures

Dream type Process Phase Key phenomenological marker
Type 1: Stress/escape frustration Harm-contrastive replay not converging (MECH-205) SWS / early REM Agent trying to escape; everything fails systematically; no resolution
Type 2: Joy/goal frustration Goal-contrastive replay not converging (MECH-205) SWS / early REM Agent trying to reach goal; everything fails systematically; no resolution
Type 3: Procedural superposition NREM SWR semantic consolidation (MECH-121, MECH-209) NREM2 / SWS No agent; one thing as all instances simultaneously; all angles, scales, times at once; peaceful
Type 4: Detailed story, close to reality REM E1 world-model free-running (MECH-210, MECH-123) REM Agent in coherent novel world; narrative arc; close to plausible; super interesting

The frustration in Types 1 and 2 is not incidental affect — it is the convergence failure signal. The dream continues because the termination criterion (causally relevant features extracted) has not been met. Non-resolution is a direct phenomenological report of the replay loop still running. The self-dissolution in Type 3 is the signature of episodic context being stripped during semantic abstraction. The narrative coherence and novelty in Type 4 is E1’s world model generating in-distribution but never-experienced sequences — interesting because E1 is surfacing latent structure from its prior.

Psychiatric predictions

PTSD (MECH-205, MECH-094, MECH-208): Recurring non-resolving Type 1 dreams are a direct marker of convergence failure on harm-contrastive replay. The causal features of the threatening episode have not been extracted; the residue field update is incomplete. The intrusive quality of PTSD nightmares follows from MECH-094 hypothesis tag degradation under high-affect replay load — the replay loses its simulation marker and is processed as real re-occurrence rather than a search procedure.

Treatment-response prediction: Successful trauma therapy should produce a shift from non-resolving to resolving Type 1 dreams before the patient reports conscious relief. The dream convergence precedes the waking resolution because the offline process completes before the waking narrative catches up. This is testable with prospective dream diaries alongside standard clinical outcome measures.

Anxiety disorders (MECH-208): Elevated harm-weighting (z_beta / residue field) allocates excess replay budget to Type 1 at the expense of Types 3 and 4. Anxious individuals are predicted to report:

  • Predominantly Type 1 dreams, longer and more intense
  • Reduced frequency of Type 3 (insufficient budget for semantic consolidation)
  • Reduced frequency or degraded quality of Type 4 (insufficient budget for world model exploration)

The self-reinforcing loop (MECH-208): elevated harm-weighting → more harm-path replay → more harm-predictive feature extraction → more avoidance → less corrective evidence → maintained harm-weighting. Dream diary composition (ratio of Type 1 to Types 3+4) is a potential proxy for the state of this loop.

Depression (MECH-203, INV-054): Reduced benefit-replay density during SWS relative to harm-replay density (MECH-203 disruption under low tonic 5-HT) predicts a specific Type 2 deficit: the joy-frustration goal dreams should be reduced or absent, while Type 1 may be maintained. This is the replay-mediated maintenance of the depressive maintenance loop — the system has insufficient approach-side contrastive replay to extract goal-predictive features, while harm-side replay continues normally.

Dementia (MECH-210, MECH-178, INV-046): As E1’s world model degrades through the noradrenergic rigidity pathway (MECH-178), Type 4 dreams should degrade in a predictable sequence:

  1. Type 4 becomes less interesting (prior no longer surfaces novel structure — repetitive familiar patterns)
  2. Type 4 collapses toward habit-like sequences (the world model can only generate well-worn paths)
  3. Type 4 disappears entirely (REM integrity too disrupted to sustain free-running)

This gradient may precede other detectable symptoms. Prospective dream diary tracking in cognitively healthy individuals over-65 may identify early E1 degradation before standard neuropsychological tests reach threshold.

REM-suppressing medications (MECH-173, MECH-210): Any medication that suppresses REM (anticholinergics, MAOIs, most antidepressants, benzodiazepines) should selectively eliminate Type 4 while leaving Types 1–3 relatively intact. Type 3 (NREM SWS process) may actually increase if the replay budget freed from REM is redistributed, but Types 1/2 (early REM contrastive replay) may also be affected by REM suppression. A testable cross-over design: measure dream-type composition before and after REM-suppressing medication, with and without sleep-stage verification.

Epistemic status

The four-type taxonomy is first-person phenomenological data from one observer (DG), noted as definite types but not claimed to be exhaustive. The mapping to computational processes is an interpretation requiring corroboration from other observers and eventual connection to polysomnographic sleep-stage measurement. Dream diaries from multiple individuals, classified by structural type (agent/no-agent, resolving/non-resolving, sequential/simultaneous), would allow initial corroboration without sleep-lab equipment.

Self-Model Failure Modes: E1 Schema Poverty and E2 Capacity Degradation (INV-064, MECH-214, MECH-215)

Registered 2026-04-07. A class of failure modes distinct from motivational/reward pathology: the preconditions for coherent goal-directed agency are absent, not the motivational system itself.

The Architectural Distinction

The motivational failure modes above (INV-052–054, 2x2 taxonomy) assume that the agent has a functioning self-model and an accurate model of its own capabilities. These failures are located in the goal-seeding and terrain-maintenance pipeline — the agent could form goals but the signal is absent or degraded.

A separate class of failure exists upstream of that pipeline: the self-model required to be the subject of goal-directed action is itself degraded. This produces a superficially similar phenomenology (absence of goal-directed behaviour, motivational flatness, inability to plan) but through a different architectural pathway with different treatment implications.

The distinction maps onto three claims:

  • INV-064: E1 schema differentiation → E2 transition differentiation → E3 evaluative capacity is a strict computational dependency. E3 cannot productively train until its inputs (z_world from E1, action_objects from E2) carry sufficient information. The biological fact that the prefrontal cortex myelinates last is an architectural necessity, not a developmental accident.

  • MECH-214: Goal content is constituted by E1 schema space. A goal aimed at an E1-unrepresented state is a random vector in E3’s operating space — it cannot generate coherent viability terrain or trajectory proposals. Goal coherence co-develops with E1 schema richness.

  • MECH-215: Goal pursuit requires E2 self-transition accuracy. The “I” that makes a goal first-personal — I can reach that — is constituted by E1’s self-schemas (who I am, what I feel, what my body is) and E2’s self-transition model (what I can do, how my actions transform my state). Without both, E3’s viability estimates have no coherent subject.

Two Failure Modes

Mode A — Goal incoherence (MECH-214 failure):

E1 schema differentiation is insufficient. The agent cannot form stable representations of the states it nominally wants. Goal-directed behaviour appears absent or directionless — not because the wanting system is suppressed (MECH-187 failure) but because there is no representationally coherent object for wanting to point at.

Signature Clinical mapping REE mechanism
Wanting without object — diffuse, undirected motivation Certain anhedonic/dysthymic presentations z_goal seeded but referent E1-unrepresented; z_goal is effectively noise
Goals stated but behaviorally inert Alexithymia, emotional-cognitive disconnection Goal object not grounded in affective E1 schema; E3 evaluates an abstract token
Difficulty identifying what one wants at all Schizoid features, depersonalisation E1 self-schema and world-schema both poorly differentiated; goal space is near-empty
Repeated pursuit of goals that produce no satisfaction Compulsive achievement without hedonic gain Goal object represented abstractly (status, money) but E1 has no schema for the hedonic state they were proxies for

Crucially, this failure is not dopaminergic. The wanting system (MECH-187 gain) may be intact; there is simply no coherent target for it to amplify. Standard antidepressants and dopamine agonists do not address E1 schema poverty.

Mode B — Capacity estimate failure (MECH-215 failure):

E2 self-transition accuracy is degraded. The agent’s model of what it can do — how its actions transform its own z_self state — is systematically wrong. E3 receives a corrupted capacity model and produces viability estimates that are either catastrophically pessimistic (the world looks unnavigable) or unrealistically optimistic (the agent overestimates its reach).

Direction Signature Clinical mapping REE mechanism
Pessimistic Cannot start — world appears unnavigable Learned helplessness, agoraphobia, avoidance E2 self-transitions predict action → no change or harm; E3 viability estimates uniformly low
Pessimistic Knows the goal, cannot move toward it Depression (action-initiation subtype) E2 capacity model: I cannot do what this goal requires
Optimistic Overestimates capability, repeatedly fails Manic planning, grandiosity E2 self-transitions inflated; E3 produces unreachable viability trajectories
Noisy Variable, unpredictable capacity Chronic fatigue, chronic pain, episodic illness E2 self-transition model trained on inconsistent data; capacity estimates unreliable across days

The pessimistic direction is particularly important: avoidance, agoraphobia, and learned helplessness may be E2 failures rather than E3 over-valuations of harm. The standard cognitive framing — the patient overestimates threat — aims treatment at E3’s harm evaluation. But if E2’s capacity model is degraded (I cannot do what is required to be safe), E3’s output is not biased — it is responding correctly to a broken capacity input. Correcting E3’s output directly (cognitive restructuring) will fail if E2 remains inaccurate.

Developmental Etiology (INV-064)

Both failure modes have a developmental pathway: early-life disruption of E1 and E2 differentiation constrains E3 training quality permanently. E3 trains on E1/E2 outputs. If those outputs carry low information during the critical developmental window, E3 learns to evaluate degraded inputs — and its harm/goal representations are noise-fitted to the actual structure of the world.

Early-life adversity that disrupts sensory schema formation (sensory deprivation, trauma that narrows the experienced world, chronic unpredictability that prevents schema stabilisation) therefore does not merely cause emotional injury. It degrades the representational substrate from which coherent goals and accurate capacity estimates can ever be drawn. The downstream effects on executive function and goal-directed behaviour persist into adulthood not because of stored trauma but because E3 was trained on impoverished inputs.

This is consistent with the empirical finding that early adversity produces deficits in goal-directed behaviour and executive function that are disproportionate to, and dissociable from, the emotional sequelae of the same adversity.

Treatment Implications

The treatment target for Mode A (MECH-214 failure) is E1 schema enrichment: expanding the experiential vocabulary, building representational differentiation. Novel high-entropy environments, arts, movement, practices developing fine perceptual discrimination, and (carefully) psychedelic experiences that temporarily dissolve existing schema boundaries can expand the space of coherent goals available to E3. This is not symptom management — it is enlarging the space from which goals can form.

The treatment target for Mode B pessimistic direction (MECH-215 failure) is E2 self-transition recalibration via action in the world. This is why exposure-based therapies show consistent superiority over purely cognitive approaches in phobias and PTSD. Cognitive restructuring addresses E3’s harm estimates directly; exposure-based therapy rebuilds E2 self-transitions through direct experience of what the agent can actually do. The capacity model is updated from below, not argued into shape from above.

Somatic and interoception-focused therapies (body-based trauma approaches, movement practices, breathwork) operate at the correct architectural level: they rebuild E1 self-schemas (who I am, what my body is) and E2 self-transitions (what I can do) directly, restoring the preconditions for coherent agency rather than targeting E3 output.

Claims Covered

ID Label Failure type
INV-064 e1_e2_e3_maturational_sequence_necessity Developmental: E3 cannot train on degraded E1/E2 inputs
MECH-214 goal_referent_e1_representability Mode A: goal content requires E1 schema differentiation
MECH-215 self_model_prerequisite_for_agentive_prediction Mode B: goal pursuit requires E1 self-schema + E2 capacity model

Catatonia, Subtype II: Harm-Stream Lock-In (SD-036, MECH-279)

Registered 2026-04-22 from V3-EXQ-471 exemplar. Distinct from MECH-202 Direction B (“commit-gate paralysis”). The two subtypes share the surface phenomenology — sustained motor immobility despite preserved consciousness — but differ in proximal mechanism.

The exemplar

V3-EXQ-471 (full-stack agent fishtank showcase, seed 0, episode 0) produced a clean catatonia trace. The agent began the episode with apparently competent goal-directed behaviour (ate food, navigated to resources). After encountering a hazard (harm_signal = -0.084 at t=0), behaviour collapsed:

  • Mode locked to avoid for all 200 steps. No flip back to a goal-seeking mode.
  • Action locked to motor 2 for all 200 steps — single repeated motor command.
  • Position stuck at [2,1] for 194 consecutive steps (world_change_norm = 0).
  • Energy fell from 0.99 to 0.00 by t=110, then remained at 0 for 90 more steps with no mode flip and no homeostatic rescue.
  • harm_signal returned to 0 at t~10, but z_harm_norm stayed pinned at ~0.7 for the remaining 190 steps. The harm latent was self-sustaining absent input.
  • z_beta_val flat at ~0.0019 for all 200 steps. The bistable beta gate (MECH-090) was not oscillating.
  • n_cands = 32 every step. Candidacy gating was not pruning.

Not the MECH-202B failure: the commit gate was not “frozen at maximum threshold” — the agent was committing, repeatedly, to the same evade action. The lock was upstream of the commit gate, in the harm-stream representation feeding it.

Mechanism

Three architectural absences compound:

  1. No harm-stream decay regulator. SD-010/SD-011 integrate harm input but contain no mechanism by which the absence of harm over time drives z_harm back toward baseline. A single harm event therefore produces a permanently elevated harm latent, which biases mode arbitration toward avoid indefinitely.

  2. No homeostatic priority override. SD-012 (drive-modulated goal seeding) requires benefit_exposure > 0 to seed z_goal. In avoid mode the agent does not approach resources, so benefit_exposure stays at zero. Drive saturates as energy depletes, but has no authority to override mode lock and force approach.

  3. No anchor reset criterion. The hippocampal proposer (MECH-269) has no signal that says “the situation has changed; re-anchor on a new world state.” The proposer keeps producing trajectories anchored on the t=0 harm event, even as 190 quiet steps accumulate. Each proposal is a variation on “evade the threat from the original frame.”

The single-action repetition is the policy reading a frozen state vector and emitting the deterministic best response to it. From the agent’s internal point of view, t=200 still is t=0.

Distinction from MECH-202 Direction B

Feature MECH-202B (commit-gate paralysis) SD-036/MECH-279 (harm-stream lock)
Commitment behaviour Gate never fires; endless re-evaluation Gate fires repeatedly; same action
z_harm trajectory Variable; not the locus of failure Pinned high without input
Action diversity Variable (no commit reached) Single action repeated
Mode behaviour Variable Locked in avoid
Treatment target Lower commit threshold; pretend-play calibration Restore decay; restore homeostatic override; anchor reset
Clinical analogue Severe indecision; rumination-loop catatonia Stupor/freeze catatonia; PTSD-adjacent immobility

These are different routes to the same surface. Treating one as the other will fail. The MECH-202B treatment (lowering commit threshold) would, on an SD-036/MECH-279 agent, accelerate the lock — more rapid commitment to the same evade action.

Clinical mapping

The clinical picture this matches is the stuporous / akinetic catatonia seen in:

  • Severe melancholic depression with psychomotor freeze
  • Post-traumatic immobility (the “frozen” pole of PTSD)
  • Some neuroleptic malignant / catatonic spectrum presentations
  • Anti-NMDA-receptor encephalitis catatonia (where GABAergic regulation is disrupted)
  • Lethal catatonia of the Stauder type (sustained autonomic activation + immobility)

The shared signature is sustained autonomic activation (z_harm-equivalent elevated) without behavioural exit — distinct from the resistive catatonia of MECH-202B (rumination, indecision, gate paralysis).

The clinical observation that benzodiazepines (GABAergic agonists) are first-line for this catatonia subtype is the empirical hook for SD-036: the missing decay regulator is GABAergic. Restoring GABAergic tone allows z_harm decay to proceed, which unlocks mode arbitration. By the same logic, ECT — which produces broad GABAergic and homeostatic recalibration effects — is second-line and works on the same architectural target from a different angle.

What SD-036 and MECH-279 commit to

SD-036 (gabaergic.cross_stream_decay_regulator): A regulatory layer that applies exponential decay to specified latent streams in the absence of input, with the decay rate itself modulated by a tonic GABAergic level. Stream coverage at minimum: z_harm (both sensory and affective per SD-010/SD-011), z_beta (precision-weighting), and possibly the drive accumulator. Architectural commitment: decay is a property of a regulatory layer that touches multiple streams, not a property of each stream’s update rule. This predicts that GABAergic disturbance produces simultaneous failures across streams — matching the clinical observation that benzodiazepine withdrawal, GABA-A receptor antibodies, and other GABAergic insults produce clusters (catatonia + autonomic dysregulation + perceptual changes + sleep disruption) rather than isolated single-stream failures.

MECH-279 (pag.gabaergic_freeze_gate): PAG-analog gating of the freeze response specifically. When z_harm × duration crosses threshold, the system commits to freeze (motor immobility + sustained autonomic activation). Exit from freeze requires GABAergic decay to complete via SD-036. Failure of this exit pathway produces sustained freeze even after the precipitating harm has passed. The fact that the freeze response itself is GABA-mediated (calming -> immobilization is one mechanism, not two) is the clinical hook — the same neurotransmitter system that initiates freeze is responsible for terminating it, and disturbance produces stuck-in-freeze.

Predictions

  1. Adding SD-036 to the V3 substrate should resolve the EXQ-471 lock pattern without any other change. Specifically, with a z_harm decay rate tuned so the latent returns to baseline within 30–50 quiet steps, the seed-0 ep-0 trace should show mode flip from avoid back to a goal-seeking mode by ~t=50.

  2. Adding SD-036 alone should not rescue seeds 1 and 2 (the more severely degraded seeds), because their failure has additional contributors. This is a substrate-readiness test, not a single-cause test.

  3. A homeostatic-override mechanism (separate, see homeostatic_override_litpull session prompt) is required to fully prevent the SD-036-rescued agent from re-locking under sustained threat with depleting energy. SD-036 provides decay; the override provides authority for hunger to outweigh threat when survival demands it.

  4. MECH-269 anchor-reset criteria (see hippocampal_anchor_selection.md) must specify conditions under which the proposer re-anchors despite a sustained mode. Without this, even a decayed z_harm may leave the proposer producing stale trajectories from the original harm anchor.

Claims Covered

ID Label Role
SD-036 gabaergic.cross_stream_decay_regulator Substrate: cross-stream decay layer with GABAergic gain
MECH-279 pag.gabaergic_freeze_gate Freeze response gating with same GABAergic substrate

Updated claims:

  • MECH-202: noted as a distinct catatonia subtype (commit-gate paralysis) from the harm-stream lock subtype documented here. Both are real; treatment differs.

Hyperarousal Insomnia and Schema-Repair Starvation (MECH-286, MECH-285, MECH-094)

The exemplar

A trauma-exposed patient (combat veteran, assault survivor, ICU survivor with post-intensive-care syndrome) presents with chronic insomnia: cannot initiate sleep despite exhaustion, frequent nocturnal awakening with full alertness, intrusive ruminative imagery on attempted sleep onset, no recoverable dream-narrative content in the morning. Daytime computation appears intact in the consulting room — language, working memory, planning all measure within normal limits. But the patient describes deteriorating performance on novel tasks at work, increasing rigidity in familiar routines, and a sense that “nothing new sticks.” Standard hypnotics (z-drugs, short-acting benzodiazepines) produce sedation but not restoration: the patient sleeps and still wakes feeling unrested. Standard SSRI augmentation does not durably help.

This is not depression-with-insomnia (the depressive-maintenance loop INV-054). It is not the harm-stream catatonic subtype (SD-036 / MECH-279) — the patient is hyper-active not hypo-active. It is not MECH-202 frame-tag failure — reality testing is intact. The architectural account REE owes is: the agent’s offline schema-repair pipeline is being starved of the offline phase it requires to run.

Mechanism

The cluster involves three interacting failures, each separately documented:

  1. Saturated MECH-286 override (sleep-onset gate stays high): under chronic threat re-experiencing, SD-037 override_signal does not relax. The wake-stability arm of orexin keeps the agent above theta_sleep_permit. Sleep onset is suppressed.

  2. MECH-284 staleness accumulator continues to fill: waking experience continues to deposit V_s residuals (the schema’s running record of “this prediction did not match”). Under chronic novel-and-unprocessed input, this accumulator grows monotonically. There is no clearance pathway in waking — clearance requires the offline replay readout (MECH-285).

  3. MECH-285 priority readout never fires (or fires degraded): because MECH-286 never permits sleep onset, the priority readout that would consume the staleness tags never runs. Schema regions with the most accumulated “this is wrong” evidence are never replayed, never re-aggregated by the sleep-half of MECH-273, and the schema does not update.

The compounding result is schema-repair starvation: novel input continues to deposit residuals that never clear, the schema becomes progressively more mis-fitted to the world the agent is currently in, and waking computation becomes progressively more rigid (defaulting to old-schema responses because the new-schema update never lands). This is a slow-motion failure: a single missed night does not produce it; weeks-to-months of pinned override do.

The PTSD-chronicity architectural account

Acute PTSD has a recoverable trajectory in many patients. Chronic PTSD does not. REE’s account of the difference is a co-failure pattern across this cluster:

  • MECH-285 priority loss alone → degraded but not absent schema repair; emotionally salient events still get dopaminergically tagged for replay; recoverable.
  • MECH-094 hypothesis-tag loss alone → simulated content can no longer be distinguished from real content; replay events can intrude as flashbacks; but the schema can still update if MECH-285 is intact, eventually integrating the trauma narrative; recoverable.
  • MECH-286 saturation alone → acute insomnia, daytime exhaustion, but on a reasonable timescale (days) the override relaxes and sleep is restored; recoverable.
  • All three jointly (MECH-286 saturated + MECH-285 priority lost + MECH-094 tag dysfunctional) → the staleness tag accumulates from the traumatic episode but never gets cleared, because (i) the priority signal that should pull it into replay is broken, (ii) the override signal that should permit the offline phase is pinned, and (iii) the replay that does manage to occur cannot update the schema because simulated and real are conflated.

The chronic phenotype is the joint failure. This is consistent with the clinical observation that PTSD chronicity correlates with persistent sleep architecture disruption and that interventions which restore sleep architecture (prazosin for nightmares, sleep-targeted CBT) move some patients out of chronicity even when the explicit trauma-processing arm has stalled.

Distinction from depressive-maintenance loop (INV-054)

INV-054 is a goal-pipeline failure (terrain-maintenance MECH-186, gain MECH-187, goal-persistence MECH-188 collectively underactive). Its phenotype is anhedonic withdrawal, low arousal, motivational poverty. The hyperarousal-insomnia cluster is the opposite arousal sign: SD-037 override is over-active, drive-coupled arousal is over-recruited via MECH-281, and the failure is in the offline recovery pipeline rather than the waking goal pipeline.

Both can co-occur (depression with hyperarousal-insomnia is common) but they are architecturally distinct and should respond to distinct interventions:

  • INV-054 → serotonergic gain restoration, behavioural activation
  • This cluster → sleep architecture restoration; in V3, predicted to require damping of the override signal under conditions where threat context has attenuated but the override has not relaxed (orexin-receptor antagonist literature; suvorexant/lemborexant clinical signal in chronic insomnia)

Distinction from acute insomnia

A single bad night, jet lag, exam stress: the override is transiently elevated but relaxes within days; MECH-284 staleness has not yet accumulated to clinically significant levels. The architectural cluster only produces the schema-repair- starvation phenotype under sustained override saturation, on the order of weeks-to-months.

This timeline matches the clinical timeline: acute insomnia is common and self-limiting; the cognitive deterioration profile only emerges in chronic hyperarousal states.

Predictions

  1. Override-relaxation interventions (orexin-receptor antagonist analogs) should restore schema-repair function even without targeting the trauma narrative directly. Specifically: in a V3 agent with pinned override and accumulated MECH-284 staleness, pharmacological-analog damping of the override signal should permit sleep-mode entry, MECH-285 priority firing, and observable clearance of the staleness map across cycles.

  2. MECH-286 saturation should be diagnostically distinguishable from MECH-281 cataplexy despite sharing an upstream substrate (orexin-analog). Both predicted to co-occur clinically (narcolepsy with cataplexy) but pharmacological dissociation possible: agonising the motor-coupling axis without affecting the sleep-state axis, or vice versa, should produce selective rescue of one phenotype.

  3. Chronic schema-repair starvation should produce a characteristic cognitive profile: preserved working-memory and crystallised performance, degraded novel-stimulus learning, degraded schema flexibility on tasks requiring updating of prior beliefs. This is dissociable from the depressive cognitive profile (which shows generalised slowing) and from substance-induced cognitive impairment (which shows acute attention deficits). The signature is learning failure under preserved performance.

  4. Time-course prediction: in a V3 agent with chronically pinned override, the deterioration in schema fit on novel inputs should follow a predictable accumulation curve (roughly linear in number of cycles with high MECH-284 deposition and no MECH-285 firing), distinguishable from the step-function degradation produced by acute substrate damage.

Claims Covered

ID Label Role
MECH-286 sleep.override_gated_state_transition Sleep-onset recruitment authority, the failing component when over-saturated
MECH-285 hippocampal.sleep_consolidation_priority_from_v_s_residuals The offline-phase readout that is never recruited
MECH-284 hippocampal.v_s_residual_schema_staleness_accumulator The waking-phase accumulator that fills without clearance
MECH-094 hypothesis_tag_routing Co-failure that converts the cluster into chronic-PTSD architecture
MECH-281 orexin.drive_arousal_coupling Sibling axis (motor-coupling); co-lesioned in narcolepsy/cataplexy but dissociable in this insomnia phenotype
SD-037 broadcast.override_regulator Substrate: the override signal whose pinning produces the cluster

Narcolepsy and Cataplexy: Bilateral Orexin-Loss Failure (MECH-281, MECH-286)

A note on scope: REE does not partition disorders into “psychiatric” and “neurological” in the way clinical specialty boundaries do. The split is institutional, not architectural — narcolepsy and cataplexy involve the same regulatory substrate (SD-037 override layer) as the insomnia, depressive, and PTSD-chronicity failure modes already documented here. We document them in this file because they share the substrate and because the architectural account unifies what specialty boundaries fragment.

The exemplar

A patient (typically adolescent or young adult) presents with two symptom clusters that conventional nosology treats as a single syndrome (narcolepsy with cataplexy) but which REE’s architecture treats as two dissociable failures of the same upstream substrate:

  • Sleep-state instability: irresistible sleep attacks during the day even in alerting contexts (driving, conversation, eating); fragmented nocturnal sleep with frequent awakenings; sleep-onset REM (collapse of the normal NREM-first architecture); hypnagogic and hypnopompic hallucinations (REM-content intrusion into wake transitions); sleep paralysis (REM motor atonia outlasting the sleep state).
  • Cataplexy: sudden bilateral loss of muscle tone triggered by emotional input (laughter, surprise, anger), preserved consciousness throughout, lasting seconds to minutes. The emotional state is computed normally — the patient laughs, then loses motor recruitment.

Standard treatment recognises both clusters but treats them with separate agents (modafinil/sodium oxybate for sleep stability; venlafaxine/clomipramine for cataplexy via REM-suppression). The architectural account should explain why one substrate lesion produces two clinically separable phenotypes treated with different drugs.

Mechanism

Narcolepsy with cataplexy is, architecturally, the bilateral failure of the SD-037 orexin-analog override at its two principal downstream targets:

  1. MECH-286 axis (wake-stability gate) fails → loss of theta_sleep_permit gating. Sleep-mode entry can occur without staleness demand, without permissive threat context — the agent transitions to offline mode at inappropriate times. Sleep architecture also degrades because the same override signal stabilises the wake state against premature REM intrusion; when override is absent, REM can intrude on wake (hypnagogic hallucinations, sleep paralysis).

  2. MECH-281 axis (drive-arousal coupling) fails → drive_level and emotional/affective input still compute (the patient feels the surprise, feels the joke is funny) but they fail to recruit the motor-pipeline gain needed to maintain postural tone. The architectural signature is recruitment failure under preserved upstream computation — exactly what distinguishes cataplexy from a primary motor disorder.

The two axes are dissociable in principle but co-lesioned in practice when the substrate (SD-037) itself is damaged, because both downstream gates are fed by the same override signal. This is the architectural prediction of the clinical co-occurrence pattern: the substrate lesion explains why both phenotypes appear, and the dissociable downstream targets explain why each phenotype responds to a different pharmacological intervention.

Dissociation predictions

The architecture is falsifiable through the dissociation pattern:

  1. Selective MECH-286 lesion (sleep-state instability without cataplexy): should be observable in early/prodromal narcolepsy patients before cataplexy onset, in narcolepsy type 2 (without cataplexy), and in some idiopathic hypersomnia presentations. The architectural prediction: a partial or selective failure of the sleep-state gating axis without the motor-coupling axis.

  2. Selective MECH-281 lesion (cataplexy without narcolepsy): rarer but reported as “isolated cataplexy” in case-series. Architectural prediction: selective failure of the drive-arousal coupling axis without the sleep- state axis. Should respond to MECH-281-targeted intervention without sleep- architecture restoration.

  3. Substrate-level lesion (full narcolepsy with cataplexy): orexin neuron loss as in narcolepsy type 1 / Hcrt-deficient cases. Both axes co-lesioned. This is the most common clinical presentation precisely because substrate damage is more common than selective downstream-target damage.

  4. V3 substrate prediction: in a V3 agent with SD-037 implemented, lesioning the override_signal at source should produce both sleep-state instability and drive-arousal decoupling. Lesioning at the MECH-286 readout while preserving the MECH-281 readout should produce sleep-state instability only. Lesioning at the MECH-281 readout while preserving MECH-286 should produce cataplexy-analog motor decoupling only. This three-arm dissociation is the architectural test.

  • vs MECH-202 catatonia (commit-gate paralysis): catatonia is a commit failure (cannot initiate or terminate motor sequences); cataplexy is a recruitment failure (cannot maintain motor tone despite intact emotional computation). Catatonia patients do not show emotional-trigger pattern; cataplexy patients do. Different substrate, different intervention.

  • vs SD-036 / MECH-279 catatonic harm-stream lock: that is a stream-level pinning failure under intact regulatory machinery; this is a regulatory machinery failure under intact streams. The harm-stream-lock patient is immobile and silent; the cataplexy patient is hypotonic but alert and communicating.

  • vs hyperarousal-insomnia (preceding section): the opposite failure mode of MECH-286 — saturated rather than absent. Both phenotypes share a substrate (SD-037 override) and they sit at opposite ends of its operating range, predicting that interventions which restore mid-range function (rather than purely up-regulating or down-regulating) should help both. This is consistent with the modafinil-and-sodium-oxybate combination in narcolepsy (selectively targeting different parts of the operating curve).

  • vs primary REM-sleep behaviour disorder (RBD): RBD is REM motor atonia failure during sleep (acting out dreams); cataplexy is REM motor atonia intrusion into wake. Architecturally both involve atonia-gate dysregulation but at different state transitions. RBD is plausibly a separable claim cluster not yet registered in REE.

What this places in REE

This failure mode adds the positive lesion clinical evidence anchor for SD-037: narcolepsy with cataplexy is the natural lesion experiment that the architecture predicts should produce bilateral phenotype, and the clinical literature confirms this pattern. The dissociation predictions above sharpen the SD-037 claim from “there exists an override regulator” to “the override regulator has at least two dissociable downstream readouts (MECH-281, MECH-286) which can fail jointly under substrate damage and selectively under downstream- target damage”.

The point about psychiatric/neurological boundaries: a regulator-layer failure that produces depression-like phenomena (hyperarousal-insomnia, schema-repair starvation) when saturated and produces narcolepsy/cataplexy when absent is, mechanistically, a single dysregulation surface viewed from two clinical specialties. The REE architecture commits to that unification — the same SD-037 → {MECH-281, MECH-286} substrate failure space generates both clinical clusters depending on whether the failure is over- or under-recruitment, and whether the downstream targets fail jointly or selectively.

Predictions

  1. Pharmacological dissociation tests: agents acting selectively on the MECH-286 axis (sleep-state stability) should rescue sleep architecture without rescuing cataplexy, and vice versa. The current clinical landscape already shows this dissociation (modafinil/sodium oxybate vs anti-cataplectic antidepressants); the architecture predicts the dissociation rather than merely accommodating it.

  2. Prodromal MECH-286 signal in prodromal narcolepsy: sleep-state instability should be detectable architecturally before cataplexy onset, consistent with clinical observation that hypersomnia often precedes cataplexy by months to years.

  3. Co-lesioning prediction in iatrogenic / immune-mediated cases: autoimmune-mediated orexin neuron loss (the established etiology in narcolepsy type 1) should produce roughly synchronous emergence of both axes’ phenotypes, distinguishable from selective downstream-target damage which should produce one phenotype and not the other.

  4. V3 substrate experiment: a three-arm lesion in a V3 agent with SD-037 implemented (substrate-source lesion vs MECH-286-only lesion vs MECH-281-only lesion) should produce the three predicted phenotype patterns above. This is the architectural validation experiment for the downstream-readout dissociability claim.

Claims Covered

ID Label Role
SD-037 broadcast.override_regulator Substrate: the override layer whose source-level damage produces full narcolepsy-with-cataplexy
MECH-286 sleep.override_gated_state_transition Wake-stability axis: selective failure produces sleep-state instability without cataplexy
MECH-281 orexin.drive_arousal_coupling Motor-coupling axis: selective failure produces cataplexy without sleep-state instability

OCD: A Three-Layer Architectural Failure (SD-034, MECH-260, MECH-266, SD-045, SD-046)

Consolidated 2026-04-28 from the four-pull architectural-question chain. Earlier framing in docs/thoughts/2026-04-20_ocd1.md through ocd4.md (goal-side reading); this section adds the chunk-side and multi-goal-arbitration readings that emerged from the action-policy decomposition lit-pull (evidence/literature/targeted_review_action_policy_decomposition/) and the type-prototype follow-on (docs/thoughts/2026-04-28_action_policy_and_multi_goal.md).

The exemplar

Compulsive hand-washing, ordering rituals, checking behaviours, intrusive doubts that cannot be released even when the patient agrees they are unfounded. Two phenomenological hallmarks: (a) an active goal that the patient cannot release (“I have to be sure my hands are clean” won’t yield to “I have washed three times, I will be late for work”); (b) a stereotyped action sequence that, once entered, plays out to a fixed completion that exceeds what the goal would justify (a single thorough washing becomes a 20-minute ritual). Egodystonia is common: the patient knows the goal is over-weighted, knows the ritual is excessive, and still cannot interrupt either.

The 2026-04-20 ocd1-ocd4 thought files framed OCD as goal-side over-binding – inability to release the active goal slot. This is correct but incomplete. Graybiel 2008 (Annu Rev Neurosci, DOI 10.1146/annurev.neuro.29.051605.112851; indexed at evidence/literature/targeted_review_action_policy_decomposition/) frames OCD compulsions and rituals as runaway action chunking – the basal-ganglia chunk substrate gets stuck in a self-reinforcing loop that the goal-release machinery cannot terminate because the failure mode is on the action-chunk side, not the goal side. Both readings are true; they describe two architectural layers that fail in characteristic combination.

Mechanism

Three interlocking failure layers, each anchored to a distinct REE substrate:

Layer 1 – Goal-side over-binding (SD-034 closure operator failure). The current goal slot persists past the point where the closure operator would normally fire. SD-034’s five-part closure signal – beta_gate.release, dacc.inject_nogo, residue domain discharge, salience signal_event, dacc.reset_episode_pe – depends on detecting that the agent’s rule-state is stable, beta gate is elevated, and the closure-domain predicate is satisfied. In OCD, the closure-domain predicate is set so tightly (per ocd4 over-bound mode-exit thresholds) that no realistic completion state crosses the closure threshold. The goal does not release because the closure operator cannot fire.

Layer 2 – Chunk-side runaway (SD-045 action-chunk cache failure mode). Independent of the goal-side problem, the action-chunk substrate (when added) can enter a self-reinforcing loop where successful execution of a cached chunk strengthens the chunk’s reinforcement signal via MECH-290 backward credit sweep, which raises its priority for retrieval next tick, which favours re-invoking it. Without an explicit chunk-level No-Go injection (the basal-ganglia task-bracketing-cell analogue), there is nothing to attenuate the runaway. SD-034’s MECH-260 No-Go injection currently fires on action-class indexing at the dACC level; a chunk-side analogue is required to attenuate runaway chunking specifically.

Layer 3 – Multi-goal-arbitration failure (SD-046 multi-slot GoalState absent). The patient cannot maintain a competing goal (“get to work”) alongside the active goal (“be sure my hands are clean”) because the GoalState is single-slot. There is no substrate where the competing goal can sit in parallel and accumulate priority through drive amplification, dACC arbitration, or MECH-266 asymmetric mode hysteresis. The phenomenology of “I know I should stop but I can’t switch” is, architecturally, “the GoalState slot is occupied and there is no second slot for the alternative to compete from.”

In combination, the three layers explain the egodystonic quality of OCD: the patient’s knowledge that the ritual is excessive and the alternative goal is more important is not represented in the goal-arbitration system at all. The cognitive evaluation is present in higher-level frontal substrates (SD-033b OFC analogue, SD-033c vmPFC analogue) but cannot be translated into goal-arbitration action because there is no multi-slot GoalState to receive the alternative. Goal-side closure cannot fire because the rule-state never stabilises in the over-bound mode-exit-threshold regime. Chunk- side runaway proceeds because the chunk substrate has no goal-independent No-Go.

vs GAD-like state (z_harm_a elevated, z_goal intact): GAD has goals that get interrupted by threat responses; OCD has goals that cannot be released. GAD is engagement-with-derailment; OCD is engagement-with-non-termination. The architectural difference is in which mechanism fails: GAD fails the AIC urgency- interrupt threshold (SD-032c) or the harm-driven mode switch (MECH-091); OCD fails the closure-domain predicate (SD-034) and the chunk No-Go (SD-045 missing).

vs catatonia subtype II (SD-036 / MECH-279 harm-stream lock-in): Catatonia has freeze-state persistence with elevated z_harm_a sustaining the freeze gate; OCD has goal-state persistence with the closure operator failing to fire. Catatonia is motor-suppression locked open; OCD is motor-execution locked into a specific cached sequence. Both show “stuck” phenomenology but at different substrate layers – one is PAG-analogue motor gating, the other is goal-arbitration plus chunk cache.

vs psychosis / delusion (MECH-094 hypothesis-tag failure, MECH-201 synthetic-real confusion): Delusion is about what the agent believes; OCD is about what the agent does despite its beliefs. The MECH-094 / MECH-201 framing is at the latent-stream level (frame-tag errors); OCD is at the action-arbitration level (goal release plus chunk cache). The patient with OCD has correct frame tags – they know the threat is unfounded – and the failure is downstream of belief.

Clinical mapping

OCD phenomenology Architectural locus
Inability to release current goal SD-034 closure-operator threshold over-tightened (Layer 1)
Stereotyped ritual sequence SD-045 action-chunk cache runaway (Layer 2)
Inability to switch to competing task SD-046 multi-slot GoalState absent (Layer 3)
Egodystonia (knowing the ritual is excessive) Higher cognitive evaluation (SD-033b/c) intact but cannot reach goal-arbitration without Layer 3
Resistance escalates ritual length Layer 2 chunk-cache runaway: blocked execution increases reinforcement on completion
Comorbid OCD + tic disorder Layer 2 chunk-cache failure shared substrate: tics are sub-second motor chunks; OCD compulsions are second-to-minute behavioural chunks; same runaway primitive at different timescales (Graybiel 2008)
Symptom heterogeneity (washers vs checkers vs orderers) Different chunk-cache content with shared Layer 1 + Layer 3 vulnerability
Familial / heritable component Vulnerability is in substrate-level parameters (closure threshold, chunk-cache update rule, goal-slot capacity) more than in any specific symptom-content trace

What the substrates commit to

  • SD-034 closure operator (V3 implemented): the closure-domain predicate threshold is the parameter that distinguishes healthy goal-release from over-binding. EXP-0156 (V3-EXQ-460) and EXP-0162 (V3-EXQ-466) validate the substrate; the OCD failure mode is what happens when the predicate is set too tight relative to the rule-state variance.
  • MECH-260 No-Go injection (V3 implemented): the goal-side No-Go that currently fires on dACC arbitration. The OCD-specific extension is a chunk-side No-Go (Layer 2) that fires when a chunk has been re-executed beyond a context-dependent count threshold. Currently absent in V3.
  • MECH-266 asymmetric mode hysteresis (V3 implemented): per-mode enter / exit thresholds. OCD over-binding is the case where exit threshold is set extremely low (the current mode must collapse to near-zero probability before leaving). Already implemented; V3-EXQ-464 (smoke-PASSed all sub-tests) validates substrate; OCD behavioural validation deferred to V3-EXQ-464b when the dual-resource environment variant lands.
  • SD-045 action-chunk cache (V4 default, V3 PULL-FORWARD CONDITION): highest-priority candidate for V3 pull-forward if EXQ-495 successors surface monostrategy persistence. The chunk-cache failure mode (Layer 2) cannot be expressed without this substrate. Without it, REE can model the goal-side reading of OCD but not the chunk-side reading.
  • SD-046 multi-slot GoalState (V4 default, V3 PULL-FORWARD CONDITION): second pull-forward candidate. The multi-goal-arbitration failure mode (Layer 3) cannot be expressed without this substrate. Without it, REE can model an agent that fails to release a goal but cannot model an agent that fails to switch to a competing goal – the two are subtly distinct, and the second is the more clinically faithful reading.

Predictions

  1. Closure-threshold sweep predicts symptom severity. SD-034 closure-domain predicate threshold parameter, swept at fixed rule-state variance, should produce a continuous gradient from healthy goal-release to compulsive non-release. The sweep is the falsifiable test that the over-binding reading is parametric rather than categorical.
  2. Chunk-cache No-Go ablation produces comorbid tic-OCD signature. With SD-045 landed and a chunk-side No-Go added, ablating the No-Go selectively (leaving goal-side No-Go intact) should produce stereotyped ritual repetition with intact goal-arbitration – pure compulsion without obsession. With ablation extended to multiple timescales of chunk cache, sub-second-chunk repetition surfaces as tic-like behaviour.
  3. Multi-slot vs single-slot ablation distinguishes “won’t release” from “can’t switch”. With SD-046 multi-slot GoalState landed, single-slot ablation in an environment with two equally-rewarded competing goals should produce pure non-switching behaviour even when closure operates correctly on the active slot. This dissociates Layer 1 (closure failure) from Layer 3 (multi-slot failure).
  4. SSRI augmentation pattern follows Layer 3 dependency. SSRIs in clinical OCD produce gradual symptom improvement over 8-12 weeks. The architectural reading: tonic 5-HT (MECH-203 / MECH-187) modulates GoalState seeding gain; with chronically elevated 5-HT, the threshold for the alternative goal to enter a competing slot (when the slot exists, Layer 3) drops, allowing arbitration to start working again. This predicts that SSRI response in OCD should correlate with measures of competing-goal availability (was there an alternative the patient was ignoring?) more than with measures of compulsion intensity. Falsifiable in a naturalistic-data study or via an ecological-momentary-assessment trial.
  5. CBT exposure-response prevention works on Layer 1. ERP forces the patient to experience the chunk-cache invocation followed by failure to complete the chunk to its cached endpoint. The architectural reading: ERP attenuates the chunk-cache reinforcement signal (Layer 2) and forces the closure operator to fire from a non-canonical termination state (Layer 1). This predicts CBT-ERP should be partially effective in single-slot architectures but maximally effective when combined with interventions that recruit competing goals (Layer 3) – e.g. CBT-ERP plus behavioural activation. Existing meta-analytic evidence supports this combined approach; the architectural account makes the dissociation testable in component analyses.

Claims Covered

  • SD-034 (governance.closure_operator) – Layer 1 substrate
  • MECH-260 (cingulate.dacc_bias_suppression) – existing goal-side No-Go
  • MECH-266 (cingulate.asymmetric_per_mode_hysteresis) – mode-stickiness substrate
  • SD-045 (bg.action_chunk_cache_dorsolateral_loop) – Layer 2 substrate, V4 default with V3 pull-forward
  • SD-046 (goal.multi_slot_state_per_goal_workstream) – Layer 3 substrate, V4 default with V3 pull-forward
  • MECH-187 / MECH-188 / MECH-203 (5-HT goal-pipeline gain) – SSRI prediction substrate
  • MECH-290 (hippocampal.backward_trajectory_credit_sweep) – chunk reinforcement source
  • SD-033b / SD-033c (PFC subdivisions) – egodystonic-cognition substrate (intact in OCD)

Claims NOT covered

  • The vulnerability factor (heritability, prefrontal cortical thickness, striatal volume, etc.) is not modelled. The architectural account locates the failure parametrically (in closure threshold, chunk-cache update rule, multi-slot capacity) but does not commit to which biological mechanism sets those parameters developmentally.
  • Pharmacological prediction (4) above is registered tentatively; entry into the PHARM-NNN registry pending review of clinical-trial literature on SSRI dose-response curves in OCD subtypes (washers vs checkers etc.).
  • The relationship between OCD and OCPD (obsessive-compulsive personality disorder) is not modelled. OCPD’s perfectionism / inflexibility may share Layer 1 / Layer 3 failure modes at sub-clinical thresholds without the Layer 2 chunk-cache runaway. This is testable but not registered.

Pharmacological Predictions Registry

The pharmacological / clinical-intervention predictions surfaced in the failure-mode sections above are registered structurally in evidence/planning/pharmacological_predictions.v1.json. Each PHARM-NNN entry is tagged with the claim IDs it depends on, the phenotype it targets, the dissociation pattern the architecture predicts, the falsification condition that would weaken the cited claims, and the literature anchors (existing trials, naturalistic data, or studies the architecture indicates are needed).

Relevant claims in claims.yaml carry a pharmacological_predictions field listing the PHARM-IDs that depend on them. This is the first cycle of pharmacological-prediction registry; governance review of these predictions against clinical literature is a separate cycle from substrate-evidence governance and has not yet been instituted.

PHARM ID Intervention Axis Phenotype Polarity
PHARM-001 DORA class (suvorexant, lemborexant, daridorexant) substrate (override damping) hyperarousal-insomnia / schema-repair starvation positive
PHARM-002 modafinil, sodium oxybate downstream sleep-state (MECH-286) narcolepsy with cataplexy (sleep-attack endpoint) positive
PHARM-003 venlafaxine, clomipramine, anti-cataplectic SNRI/TCA downstream motor-coupling (MECH-281) narcolepsy with cataplexy (cataplexy endpoint) positive
PHARM-004 lorazepam, benzodiazepines regulator-layer GABA (SD-036) catatonia subtype II (harm-stream lock) positive
PHARM-005 prazosin (alpha-1 antagonist) indirect override damping PTSD nightmares + hyperarousal-insomnia positive
PHARM-006 SSRIs (sertraline, paroxetine) regulator-layer 5-HT (orthogonal) hyperarousal-insomnia / schema-repair starvation negative
PHARM-007 anticholinergics (oxybutynin, diphenhydramine, TCAs, paroxetine) dual pathway (REM-suppression + diurnal cholinergic mimicry) dementia conversion in MCI/older adults negative
PHARM-008 mirabegron (beta-3 agonist; non-CNS antimuscarinic alternative) indication substitution (urinary urgency without ACh load) dementia conversion in older adults on bladder anticholinergics positive
PHARM-009 benzodiazepines / Z-drugs (zolpidem, zopiclone) regulator-layer GABA (architecture-disrupting at hypnotic dose) dementia conversion via Phase 1/2 SWS suppression negative (cf. PHARM-004 positive in catatonia – same drug class, different phenotype)
PHARM-010 DORA class for MCI/dementia substrate (architecture-preserving sleep) MCI/early-AD episodic decline rate positive (sibling of PHARM-001)
PHARM-011 trazodone 25-100mg (5-HT2A antagonism, sub-antidepressant dose) downstream Phase 1/2 SWS enhancement MCI/early-AD episodic decline rate positive
PHARM-012 mirtazapine, sertraline, escitalopram (low-REM-burden antidepressants) vs paroxetine/TCA/MAOI antidepressant-class dissociation dementia risk in long-term antidepressant users positive (low-REM-burden) / negative (REM-suppressing)
PHARM-013 ramelteon, melatonin-ER upstream Phase 0 circadian restoration MCI/early-AD episodic decline rate positive
PHARM-014 composite pipeline-rebuild dissociation (cross-class predictor) falsifiable architecture test reverse phase-dependency ordering of cognitive recovery architectural

The dementia/medications cluster (PHARM-007 to PHARM-014) is not a “psychiatric” failure mode in the standard nosological sense — but the same scope note applied to narcolepsy and cataplexy applies here: REE does not partition disorders into psychiatric and neurological. Dementia is a sleep-pipeline failure mode under INV-046/047/048 with a substantial pharmacological prediction surface; full mechanistic discussion lives in sleep/medications_dementia.md.

Difficulty-Gated Proposal Entropy: failure modes of stuck-state cognition (MECH-343, Q-056 — working hypothesis)

Status: WORKING HYPOTHESIS (registered 2026-06-03). Not an established mechanism; no experimental evidence yet. Hypothesis note: docs/thoughts/2026-06-03_difficulty_gated_proposal_entropy.md. Experiment proposal EXP-0176 (evidence/planning/manual_proposals.v1.json).

MECH-343 proposes that when a goal is blocked, a healthy agent should transiently widen its internal proposal distribution (more diverse hippocampal/world-model rollouts, higher within-class sampling temperature) under preserved goal/harm constraints, defer commitment until a candidate clears the E3 score-margin threshold, then narrow once a workable candidate is found. Controlled entropy, not noise: hard problem → wider internal proposals → more candidate diversity → longer arbitration → selective committed action — never hard problem → random behaviour.

Read as a coordinate system for cognitive pathology, each step of that loop has a distinct failure mode — and several map onto states already in this document:

Failure signature Architectural locus Psychiatric analogue
entropy rises but behaviour unchanged selection-authority failure at E3.select (the current modulatory-bias-selection-authority bottleneck; MECH-314/MECH-320 fire but never change argmax — V3-EXQ-604a/624a) inert effortful striving without behavioural change
proposals widen but commitment never occurs action-release / commitment-threshold failure (flat score margins never clear admission — cf. MECH-342 / V3-EXQ-629; MECH-090 floor) catatonia / abulia, action-release subtype — distinct from the SD-036/MECH-279 harm-stream lock-in subtype above: there the gate is jammed by harm, here the candidate distribution never resolves to a committable winner
entropy fails to rise despite a stuck state proposal-generation rigidity (ARC-018 rollout set does not widen) perseverative / cognitively rigid states; failure to “step back and think differently”
entropy rises but never narrows arbitration that cannot converge (no decay, no commitment threshold met) manic / disorganised exploration — flight of ideas without selection
entropy rises and behaviour becomes random/harmful arbitration failure (goal/harm scoring not preserved during widening) disorganised/agitated states; the diagnostic arm-4 ablation in EXP-0176

The clinically important point this hypothesis surfaces is that catatonia/abulia is not a single architectural state. SD-036/MECH-279 (above) is the harm-stream lock-in subtype — the commitment gate is held shut by an over-weighted harm stream. The MECH-343 action-release subtype is different: goal salience and harm scoring are intact, but the proposal→arbitration→commitment loop never resolves a committable candidate (the score-margin distribution stays too flat to clear threshold). Same surface phenomenology (inert, non-committing), different locus, potentially different intervention. This hypothesis is one candidate route from catatonic-like inert cognition toward controlled exploratory cognition, and it predicts that the direction of repair differs by subtype: lower harm gain (lock-in subtype) vs. restore score-margin contrast / proposal-entropy gating (action-release subtype).

Epistemic status: forward-reference only. None of these mappings are validated; they are predictions the MECH-343 / Q-056 experiment program (EXP-0176) would test. The loop is currently blocked on two upstream substrates (modulatory-bias-selection-authority and a difficulty-gated proposal-entropy regulator), so MECH-343 is registered epistemic_category: substrate_conditional and is not promotable until they land.

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