JEPA as a sub-layer of CMVG
ADR-048 → ADR-058 · SIGReg isotropic-Gaussian manifold
↓ scroll to descend
One direction of flow. The sensor network is primary; the world model is one additive subscriber. ExoChain is the attestation spine that lets standby and peer services catch up from replay alone.
“The right move is not to replace ECC with LeWM — it is to put LeWM under ECC as a subsymbolic sensor substrate.”
— synthesis.md, Round 1
Two timescales, one loop: the causal graph learns what the right variables are (slow, symbolic); the predictor learns how they evolve (fast, differentiable). The sensor network is primary and self-sufficient. The world model is an additive consumer.
A lattice fed by observers. Sensors point inward at a shared world-model. Coupling everywhere.
A self-sufficient sensor pipeline with the world model as one optional subscriber. Data flows one-way.
Each sensor class runs its own three-step pipeline. All emit onto the same SIGReg-regularized isotropic-Gaussian latent so any consumer can subscribe uniformly.
depth + color stream — primary spatial sensor for grasp planning, free-space, and object identity.
fast inertial sensor — drives the 1 kHz servo loop and bootstraps every prediction.
12 joint angles + torques — DEMOCRITUS reads here every millisecond.
spatial truth — long-range geometry that anchors the visual world model.
event sensor — surprise spikes, voice commands, environment classification.
Three small models per sensor class. Each is RVF-hosted, separately versioned, separately attested, and hot-swappable at tick alignment with 30-second SIGReg rollback.
Three RVF segments · three independent rollback domains · one shared latent contract.
SIGReg + VICReg pull every encoder’s output toward an isotropic Gaussian. That is what makes the wire universal — any subscriber can read from any sensor class without bespoke decoding.
The contract is versioned, ExoChain-attested, and degrades — never crashes. If a new edge model drifts off the manifold, sigreg_health drops, and the four-condition rollback engages within 30 seconds.
SIGReg health · probing accuracy · VoE surprise differentiation · temporal-straightening score. Any single dip can be tolerated; the conjunction is what triggers an automatic rollback to the last attested checkpoint.
The world-model contract is identical. The deployment shape adapts to fleet size, mesh quality, and trust boundaries — every variant is observationally equivalent to local ECC consumers.
One WorldModelService for the entire fleet. Local ECC consumers fall back cleanly when it is unreachable. Simplest to operate, the right default for small teams.
Active primary with a hot standby trained off the same ExoChain replay. Failover is a leader election, not a re-train. Standby promotes within seconds.
Every node runs its own WorldModelService and gossips checkpoints over mesh. No central authority. Used when fleets cross trust boundaries or partition tolerance dominates.
RGB-D · IMU · Proprio · LiDAR · Audio stream into a three-step learned pipeline:
Collect (transmit-gate) → Aggregate (multi-sensor fusion) → Encode (SIGReg-manifold ViT-tiny).
Three trainable RVF-hosted small models per sensor class. Hot-swap at tick alignment. Auto-rollback on sigreg_health < 0.85 for 30 s.
Three topics under mesh.sensor.v1.*: encoded (primary, 10 Hz, burstable 100 Hz), consensus (conditional, world-model output), control (drift / surprise / plans).
Packets are observational-only. No gradient proxies. No consumer-specific framing.
Every frame Ed25519-signed, ExoChain-indexed. Replay is the single source of truth.
hnsw_eml (ruvector) · ECC Causal Graph (causal.rs, 3417 LOC) · CMVG forest with StructureTag::LatentWorldModel.
LatentHandle(Uuid) lives in CausalNode.metadata — no schema break.
DEMOCRITUS runs the 1 kHz servo off an ArcSwap<ActionPlan>. The 1 ms cognitive tick never runs the planner inline.
Fusion Transformer → Predictor pred_φ → LatentPlanner (CEM default · MPPI warm-start · gradient-shooting).
Exposes LatticeApi as a first-class ServiceApi. observe · predict · plan · recall · subscribe_surprise · subscribe_drift.
Three deployment topologies: single primary · Raft-elected primary + standby · peer-to-peer.
ServiceUnavailable is a contract value. Callers fall back to local-ECC-only paths.
Edge intelligence: offline, per-sensor-class. Weights bundled as RVF segments. Hot-swap safe because the SIGReg manifold is contract, not implementation.
World-model streaming merge: online, in-service. Importance-weighted replay per sensor class — rare-event sensors retain long residency.
Four-condition AND rollback: SIGReg health · probing accuracy · VoE surprise differentiation · temporal-straightening score.
Every sensor frame · every checkpoint · every RVF segment activation · every rollback · every governance topology change · every LatticeApi capability grant.
Standby and peer WorldModelServices catch up from replay. Training data comes from ExoChain, never from the live bus.
the architecture · now · the cycle it thinks through ↓
Three timescales run concurrently: 1 ms DEMOCRITUS servo tick · 10 Hz planner + predictor + HNSW insert · continuous streaming-merge checkpoints. Scroll to set the cycle speed.
Causal graph stays authoritative · latent world model publishes impulses
Framework: candle · Model: ours · EmlBackend trait for swap
Versioned · ExoChain-attested · degrade not crash
CEM default · MPPI feature-gated · per-node advertised
observe · predict · plan · recall · surprise · drift
mesh.sensor.v1.* · CBOR · Ed25519 · observational-only
Single · hot-standby · peer-to-peer · ServiceUnavailable
Streaming merge · importance-weighted replay · 4-condition AND rollback
In-process · separate-process · delegated-key attestation
RVF-hosted · hot-swap at tick · 30 s SIGReg rollback
Sensor network primary · world model additive · the constitution
The sensor network is a self-sufficient three-step learned pipeline whose emissions live on a SIGReg manifold so any consumer, including an optional world model, can subscribe uniformly.
ADR-058 — the constitutional invariant — is what makes every other decision on this page consistent.
weftos.weavelogic.ai · feature/lewm-worldmodel · 2026-04-22