Sense-specific perceptual manifolds before shared world-model entry
Status: SEED / candidate cluster (V4/V5 perception). Off the V3 / GAP-7 critical path.
Home doc for the claim cluster reaped from docs/thoughts/2026-06-07_sight_specific_perceptual_manifolds.md (source trigger: the ScienceDaily report on completing Schrodinger’s colour theory).
Seed
Senses differ in how much transformation they need before entering the shared world model. A sense does not contribute raw data directly; it contributes a shaped geometry of possible differences.
- Smell is naturally gradient-like (stronger/weaker, nearer/farther, attractive/aversive, trail/no-trail, source-seeking/avoiding) and can enter cognition as a near-raw primitive orientation signal with a minimal adaptor – a local gradient that can be followed, escaped, strengthened, weakened, remembered, and associated with outcome. This makes a smell-like sense unusually compatible with early REE world-model construction.
- Sight is not like this. Even colour perception requires a structured perceptual geometry (hue/saturation/lightness emerge from the geometry of perceptual similarity, not as labels attached to raw stimulus) before becoming cognitively useful. Vision needs a deep adaptor.
The architectural pattern:
physical signal
-> receptor encoding
-> modality-specific adaptor
-> perceptual metric / manifold
-> stable perceptual primitives
-> shared world-model update
-> action arbitration
The adaptor is not an optional feature extractor: it is the stage where raw physical signal becomes a usable geometry of difference. The shared world model is built from sense-specific geometries made mutually negotiable, not from interchangeable raw sensory channels.
What REE already owns (cross-ref, not duplicated)
| Owned claim | What it covers | Why this cluster is distinct |
|---|---|---|
| ARC-017 sensory stream tags | Minimal stream-tag set with typed exteroception + reality-coherence lane | Tags the streams; says nothing about per-sense transformation depth or manifold geometry |
| MECH-103 E1 multimodal exteroceptive fusion | Each modality has its own encoder pathway into the shared latent; complementary structure; multi-source precision-weighted fusion | Treats the pathway as a feature extractor / fusion step; does not claim it is a metric/manifold constructor whose depth is sense-specific |
| ARC-027 harm stream parallel pathway | Nociceptive stream runs parallel to z_world | A specific stream-separation, not the general adaptor-depth heterogeneity principle |
| ARC-019 staged curriculum | Development is gated/staged | This cluster adds which sense is admitted when, indexed by adaptor depth (smell-first) |
Claims in this cluster
- ARC-087 (architectural_commitment): sense-specific perceptual-manifold adaptor – each modality requires its own adaptor transforming raw stimulus space into a perceptual-manifold (metric) space before shared-world-model entry; senses are not interchangeable raw streams. The adaptor is a metric/manifold constructor, not merely a feature extractor.
depends_on: ARC-017, MECH-103, ARC-004, ARC-005, ARC-019. - MECH-372 (mechanism_hypothesis): modality-heterogeneity / adaptor-depth gradient – senses differ in transformation depth (smell near-raw gradient; sight deep manifold: colour geometry, edges, figure-ground, depth, invariances, affordance/gaze salience). Developmental-ordering corollary: gradient-like (low-adaptor-depth) senses enter earliest.
depends_on: ARC-087, ARC-017, ARC-019, MECH-103. - Q-065 (open_question): (1) is each modality’s perceptual metric learned or partly defined? (2) how are heterogeneous sense-specific geometries (smell gradient; sight manifold; touch boundary/pressure/texture/contact; hearing temporal-source/rhythm/pitch/localisation; proprioception body-state transition) made mutually negotiable inside one shared world model?
depends_on: ARC-087, MECH-372.epistemic_category: substrate_conditionalset explicitly sonarrow_open_questiondoes not fire.
V3 vs V4/V5 boundary (DO NOT build in V3)
All three claims are status: candidate, epistemic_category: substrate_conditional, implementation_phase: v4, version_relevance: v4_v5. For REE-v3, smell/gradient-only sensing remains appropriate and a deep adaptor is unnecessary – this cluster is off the V3 / GAP-7 critical path. The benchmark dates in the source thought (optimistic 2026-06-28 / strict green-board 2026-07-19 / pessimistic 2026-08-16) are V3 milestones; this work must not distract from them. For V4+, multimodal sensing must not be treated as “add more input channels” – each sense needs its own preprocessing grammar, and the shared world model should receive already-shaped perceptual structure.
Because all three carry epistemic_category: substrate_conditional, they are suppressed from promotion/demotion and from the IGW /queue-experiment proposal lane during the V3 phase (same construction as the play-mode cluster) – no blocked_substrate STOPs needed.