Cognifold Signed Coupling (ARC-084 / MECH-363 / Q-058)

Status: candidate cluster, V4/V5, off the V3 critical path. Registered 2026-06-09 from the competitive-interactions thought intake.

Source intake: evidence/planning/thought_intake_2026-06-06_competitive_interactions_cognifold_stability.md Raw thought: docs/thoughts/2026-06-06_competitive_interactions_cognifold_stability.md Empirical anchor (verified): Luppi, A. I. et al., “Competitive interactions shape mammalian brain network dynamics and computation.” Nature Neuroscience 29(4):915-933 (2026). DOI 10.1038/s41593-026-02205-3.

This is a stub / compass anchor, not a build spec. It exists to give the registered claims a location: target and to record the design questions. No V3 substrate work is licensed by this document. Promotion to a V4 design spec requires an explicit version decision and (per the biology-before-formal-definitions rule) a biology lit-pull on signed coupling first.

The idea

A cognifold cannot be “everything positively coupled to everything”. An all-positive routed-field model is unstable: every salient stream reinforces every other, producing runaway resonance, over-synchronisation, and loss of modular segregation. Luppi et al. show empirically (human / macaque / mouse whole-brain models) that combining modular cooperative interactions with diffuse long-range competitive interactions reproduces mammalian dynamics far better than cooperative-only models, and that competition raises synergy, hierarchy, and neuromorphic-computing performance.

The REE extraction is a structural one: signed competitive coupling may be a necessary generative ingredient of stable cognifold dynamics — not damage, not punishment, not synaptic inhibition, and not merely an error-correction afterthought.

ARC-084 — Typed signed cognifold coupling

The cognifold represents inter-field coupling as a typed, signed edge:

cognifold_edge = {source, target, sign, gain, precision, gate, timescale, write_authority}

with three coupling modes:

  1. Cooperative (sign > 0) — binding, coherent trajectory generation, shared context.
  2. Competitive (sign < 0) — suppress incompatible trajectories, preserve boundaries, prevent runaway resonance, stop one field absorbing all others.
  3. Gated-decoupling — temporarily isolate simulation, offline integration, or unsafe action candidates from release authority.

The genuinely-new element is the sign as a first-class, explicit edge property at the cognifold-field level — competition represented structurally, not left to emerge from selection softmaxes.

V3 reality (containment-only — do NOT build a parallel signed-edge module): REE already runs competition, but locally and implicitly:

Coupling mode Existing V3 instance
Competitive (local) BG winner-take-all lateral inhibition + Cisek/Kalaska affordance-competition (MECH-090); symmetric Go/NoGo dMSN/iMSN competitive model (ARC-030); top-k competitive selection into E3 (MECH-254)
Gated-decoupling MECH-094 waking-only write gate; MECH-090 commitment latch; MECH-261 mode-conditioned write gating
Cooperative latent binding, trajectory generation across E1/E2

This is structurally the same situation as the [attention = distributed precision-selection] map (ARC-005, MECH-251/254/255/259/261/347, SD-032a): REE owns the pieces but lacks the explicit map. ARC-084 is a unifying MAP / commitment, not a new module.

MECH-363 — Diffuse long-range competition as a generative stability requirement

Stable, synergistic, hierarchical cognifold dynamics require signed competitive coupling — specifically diffuse, long-range competition between distant fields with opposite profiles — not merely local damping. Falsifiable in principle by a multi-field coupling ablation (cooperative-only vs cooperative+long-range-competitive on the same field set), but substrate-gated: V3 has no explicit multi-field signed-edge layer to ablate, so a probe today would be vacuous.

Two NON-equivalent failure axes (do not conflate):

  • (a) Runaway positive coupling / hypersync — MECH-363’s target. REE analogues: feedback entrapment, shared delusional coupling, over-stabilised attractor (MECH-076). Addressed by signed-coupling damping.
  • (b) Monostrategy / regime-collapse — the opposite pole (too little apprehended structure, not too much coupling). Owned by MECH-309 / ARC-062 / ARC-063. Addressed by the rule-apprehension layer, not by competitive damping.

Q-058 — Explicit vs emergent competition, and the safety taxonomy

Which subsystems genuinely need an explicit signed competitive edge rather than the existing local softmax/WTA? And what are the safety differences between inhibition, competition, decommitment, and residue-based veto across waking action, simulation, and offline integration? Open design questions:

  1. Which inter-field couplings need an explicit sign vs existing local softmax/WTA (MECH-090/ARC-030/MECH-254)?
  2. Safety differences between inhibition, competition, decommitment (MECH-090/MECH-105), and residue-based veto (harm stream, SD-010/SD-011)?
  3. How should competitive coupling differ across waking action vs simulation (MECH-094) vs offline integration (MECH-272/273 sleep cluster)?
  4. Should the AI-cognitive-failure taxonomy be re-indexed by missing/misweighted competitive coupling? (Cross-repo: Latent-Fields/ai-cognitive-failure-taxonomy, not REE_assembly.)

Cautions (carry forward verbatim)

  • Competition != synaptic inhibition (Luppi et al. say so explicitly).
  • A competitive/negative edge is not intrinsically harmful, pathological, or ethically negative.
  • Competition is not reward subtraction / punishment — it is a structural relation between active fields.
  • The paper does not prove REE’s architecture; it is corroborating compass, not evidence for a claim.
  • Do not “add inhibition everywhere.” The correct extraction is: preserve signed competitive coupling as a possible constructive cognifold primitive.
  • Do not make this a V3 implementation target absent a specific failure or explicit version decision.

Dependency wiring

  • ARC-084 depends_on ARC-005 (control plane routes precision and modes).
  • MECH-363 depends_on ARC-084, ARC-005.
  • Q-058 depends_on ARC-084, MECH-363.

All three: status: candidate, epistemic_category: substrate_conditional, implementation_phase: v4, version_relevance: v4_v5. Whether to take them into V4 design work is a later, explicit decision.

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