Psychiatric Failure Modes as Latent Vulnerability Axes
Registered 2026-06-09 from the 2026-06-06_psychiatric_genetic_overlap_latent_failure_axes thought intake. Home doc for RA-002, ARC-086, Q-064, MECH-371. This doc adds a second, cross-cutting index on top of the per-syndrome failure-mode maps in psychiatric_failure_modes.md; it does not replace them.
Motivation
A large-scale genomic study (RA-002) finds that 14 clinically distinct psychiatric disorders share a small set of latent genomic factors that cut across diagnostic boundaries. The REE-relevant extraction is not genetic determinism. It is the surface-cluster principle: psychiatric categories are emergent clusters over overlapping latent vulnerability dimensions, not genetically discrete natural kinds.
REE already holds this stance — see IMPL-005 (docs/REE_failure_modes.md), the FAILURE-2026-02-12_COORDINATE-SYSTEM-FOR-COGNITIVE-PATHOLOGY thought (“orthogonal taxonomy organised by dynamical failure geometry rather than historical narrative”), INV-061 (“different failure signatures of the same undertrained mechanism”), and MECH-343/Q-056 (one loop, many syndrome-analogue failure modes). What this cluster adds is an explicit, finite, named cross-cutting vulnerability-axis vocabulary as a second formal index over the named failure modes, a representation of syndromes as weighted axis-vectors, and a p-factor analogue.
Guardrail (carried into every claim here). The axis layer is an indexing layer, never a merging layer. Two failure modes sharing an axis does not make them the same failure. Psychosis (MECH-094 / MECH-201), confabulation (MECH-094 recall/completion sense), derealization (MECH-200), OCD (SD-034/045/046), the depressive attractor (INV-053/054), and the catatonia subtypes (SD-036/MECH-279 vs MECH-343) remain mechanistically distinct. See
memory/feedback_psychosis_confabulation_distinction.md.
RA-002 — Psychiatric genomic-factor overlap (external anchor, out-of-domain)
Grotzinger A. D. et al., “Genetic overlap across 14 psychiatric disorders,” Nature Genetics (2025), DOI 10.1038/s41588-025-02494-7. Genomic SEM across 14 disorders in >1 million individuals; the majority of case-vs-control genetic variance is captured by five shared genomic factors — Compulsive; Schizophrenia-bipolar (psychotic-mood); Neurodevelopmental; Internalizing; Substance-use — spanning 238 variants, with factor-specific biology (oligodendrocyte genes enriched in the Internalizing factor) acting early in brain development.
epistemic_category: out_of_domain. The test domain is a human-genetics cohort; no REE substrate distinguishes genomic factors, so this is never cited as REE mechanism evidence — it is an external corroboration anchor for the surface-cluster principle and motivates ARC-086 / Q-064 / MECH-371. Anti-determinism guardrail: genetic overlap is not proof diagnoses are meaningless, and individual presentations cannot be read from polygenic factors.
ARC-086 — Two-axis failure-mode index (named mode × latent vulnerability axis)
REE’s failure-mode taxonomy is indexed on two axes at once:
- Named failure mode — the existing mechanistically-distinct entries (psychosis, confabulation, derealization, OCD, depressive attractor, catatonia subtypes, …). What breaks.
- Latent vulnerability axis — a finite, named set of dimensions that cut across the named entries. Along which shared dimension the vulnerability sits.
Candidate axis set (descriptive; refine as the registry matures):
| Axis | Closest existing REE substrate |
|---|---|
| precision / confidence allocation | ARC-005 control plane, MECH-251 precision family |
| commitment threshold / action-release authority | E3 commit boundary, MECH-090/342, SD-034 |
| provenance / source tagging | MECH-094, MECH-037, INV-011 |
| agency attribution | (self-vs-other authorship tagging; cross-ref taxonomy repo) |
| residue / consequence persistence | ARC-013 residue geometry, SD-010/SD-011 |
| goal / salience / incentive coupling | INV-034, MECH-112/116, SD-012, MECH-186-188 |
| self-other coupling | ARC-010 social cluster, Steve cluster (MECH-182/193 …) |
| offline integration / replay | INV-045/049, MECH-121/273/285 |
| social-contagion / cross-agent coupling | (multi-agent; V4/V5) |
| global instability (p-factor analogue) | no existing claim → MECH-371 |
Named clinical syndromes are emergent clusters — weighted combinations — over these axes, not primitive kinds; comorbidity is overlapping axis activation; pleiotropy is one mechanism loading multiple axes.
claim_type: architectural_commitment, status: candidate, epistemic_category: substrate_conditional (explicit, mirroring ARC-085): a candidate organizing principle whose weighted-vector read-out (Q-064) has no V3 substrate. The documentation indexing can begin now; the vector substrate is V4/V5. depends_on IMPL-005, INV-061, MECH-343, RA-002.
Q-064 — Syndrome as a weighted axis-vector? (open question)
Should a clinical syndrome be represented as a weighted vector over the ARC-086 axes (comorbidity = axis-overlap; pleiotropy = one mechanism loading multiple axes) rather than as the current per-syndrome locus mappings? epistemic_category: answer_state. Two sharp sub-questions:
- Registry economy — comorbidity as overlapping axis activation must not require a combinatorial blow-up of claim entries.
- Predisposition vs expression — genetic-risk ≠ inevitability maps to axis-predisposition + environment + developmental timing + plasticity; the representation must separate “which axes are fragile” from “which failure mode is currently expressed.”
Resolvable in REE only once the ARC-086 vector substrate exists (V4/V5); off the V3 critical path.
Q-075 — Are the vulnerability axes grounded in dissociable neuromodulator-receptor systems? (open question)
Reaped 2026-06-12 from the receptor-subtype intervention layer lit-pull (strands 3A/3B/3C; evidence/literature/targeted_review_receptor_subtype_layer/). Where Q-064 asks how to represent syndromes as vectors over the ARC-086 axes, Q-075 asks the logically prior empirical question: are the axes grounded in dissociable neuromodulator-receptor systems at all? The receptor layer proposes a specific grounding —
- precision / terrain axis ← serotonin subtypes: 5-HT2A-cortical (narrative exclusivity) vs 5-HT1A-hippocampal (aversive-basin/terrain depth) [MECH-006/MECH-085; Ren 2019, Weber & Andrade 2010, Akimova 2009]
- goal/salience-coupling + commitment-release axes ← M4-striatal dopamine-selection damping, the non-D2 antipsychotic route [MECH-086/MECH-087; Foster 2016, Kaul 2024 EMERGENT-3]
- commitment-threshold axis ← the kappa/mu commitment-entropy overlay (MECH-048; kappa = destabiliser/aversive-state-lock, mu = stabiliser) [Land 2008, Krystal 2020, Bari 2025 — KOR antagonism restores optimal perseveration]
epistemic_category: out_of_domain — the test domain is clinical pharmacology (receptor-selective RCTs / cohorts), not the V3 substrate (which has no receptor subtypes; the receptor rung is out_of_domain per receptor_subtype_intervention_layer.md §4). What would answer it: a receptor-selective double-dissociation — each agent moves its predicted axis-endpoint without equivalently moving the partners’ — as set out concretely in the PHARM-015..018 falsification conditions (e.g. if kappa antagonism gives the same anhedonia-endpoint profile as an SSRI, the commitment-entropy axis collapses into the serotonin-terrain axis → answered NO for that pair). The V3-tractable sibling is the MECH-087 cross-plane non-rescue test — a plane-level, not receptor-level, dissociation. Promotes nothing; lit grounds the design only. depends_on ARC-086, MECH-048, MECH-085, MECH-086, MECH-087.
MECH-371 — Global-instability / control-plane vulnerability axis (p-factor analogue)
A single broad dimension of control-plane regulatory fragility — instability in precision/gain/threshold setpoint regulation (ARC-005) rather than any one comparator or stream — loads across all failure modes, raising joint susceptibility, distinct from and orthogonal to any single-axis failure. This is the general-psychopathology / p-factor analogue, and it is the one axis in the ARC-086 set with no existing REE home (the other axes map onto owned substrate; see the table above).
claim_type: mechanism_hypothesis, status: candidate, epistemic_category: substrate_conditional, V4/V5. RA-002’s shared-variance finding and the Caspi & Moffitt p-factor literature motivate the dimension; a brief general-psychopathology lit-pull would sharpen it before any promotion. Guardrail: a global-instability axis is an additional orthogonal loading — it does not claim all psychopathology is one thing, and it does not erase the distinctness of the specific failure modes. depends_on ARC-086, ARC-005, RA-002.
Cross-repo
This cluster also seeds a latent-vulnerability-axis layer in the separate Latent-Fields/ai-cognitive-failure-taxonomy repository (a docs/latent_vulnerability_axes.md layer beneath that repo’s named failure-mode entries). That work lives in that repo, not in claims.yaml.
Status
All four claims are status: candidate and promotion-suppressed by epistemic category (out_of_domain / substrate_conditional / answer_state). None promote or demote any existing claim. Off the V3 / GAP-7 critical path; no experiment is queued (a probe would be vacuous — there is no multi-axis-vulnerability substrate in V3, and RA-002 is out-of-domain).