Substrate · what's inside
The cascade is the architecture. Most useful work resolves at deterministic tiers for picojoules; a model tier loads only when the spine cannot place a request. Receipts are structural, not billed — you can read them on this device. Four pieces:
The cascade emits its trace as the chain of thought.
A small cascade walks each query through a ladder of tiers — closed-form, retrieved, identifiable, cited, model-class. Every hop records what was tried, what matched, and what it cost. The trace IS the reasoning; replaying it reproduces the answer.
Zero new compute. The transcript is the chain of thought.
open the exhibit →Four deterministic encoders project every modality into the same hypervector space.
Text, image, audio, and video each ride a modality-specific encoder — trigram bag, cell-level bind, FFT-band bag, and per-frame image-bind with temporal permute-and-bundle for video. All four produce a 10 000-dim bipolar hypervector; cross-modal cosine is meaningful because the algebra is shared.
~150 LOC per encoder. No model weights. Modality is a substrate property.
open the exhibit →Every cascade tier is split into two sub-experts.
A reasoning expert classifies the form of the query; a generation expert produces the answer in that form. Generation only fires when reasoning matches. Both sub-hops are individually accountable in the trace; the receipt prices each separately.
A refactor of the cascade, not new compute. Resolved tiers contribute two hops; unmatched tiers contribute one.
open the exhibit →The model tier loads only when the deterministic spine cannot resolve.
Most queries resolve at the spine for picojoules. When a query escapes, the page dynamic-imports a separate WASM bundle, pays its byte cost plus per-token inference, and surfaces both costs as a receipt. The leaf is small because the spine catches almost everything.
Two bundles visible in the browser's Network panel: the spine on first paint, the leaf only on escalation.
open the exhibit →The substrate's lane is energy-optimised work output. Hold the envelope small enough to live on a drone, a sensor camera, a microcontroller, a wearable, a medical implant. State the position; measure it; publish the receipt.
Every metered call returns its joules — computed at the call site against a silicon-specific cost model, the Landauer floor, and a TDP envelope. No abstract throughput metric. No tokens-per-second proxy. Joules.
The cascade resolves at the lowest tier whose grammar covers the request. Closed-form math, named-fact tables, and identifiable shape-matches catch almost everything. A model fires only when the spine could not place the request.
Every claim carries an explicit replayability class — L0-closed, L0-retrieved, L0-identifiable, L1, L1.5, L2. The substrate refuses to fabricate; unmarked claims cannot be constructed at the type level.
Two replicas serving the same command stream land on identical FNV-1a transcript digests. Drop, mutate, or reorder — any of those — and the digests diverge. Trust is the comparison, not a separate protocol.
Every exhibit reads its joule cost on your device. The session aggregator on /receipts sums them; /scale plots your last frame on a log axis from the Landauer floor to a kilometre in an EV. None of it is uploaded.