A memoizing cascade over Imagine, Semantic, and Math-Ground AIs. Every inference returns a signed receipt with the energy it cost and the datacenter that ran it.
Two ways in
A user asks for a real-estate match. The cascade routes through Imagine, Semantic, and Math-Ground — but a near-identical prior query has already paid the upstream cost. The bill drops to 8.4 J.
Pick a query class. See which tier serves it, what it costs in joules, what fraction is memoized, and the receipt it emits. Every tier annotated with the model it runs.
The receipt
This is what comes back from the cascade. Not just an answer — a signed JWP ReceiptPayload with the tier sequence, the memoization status, the joules consumed, and the datacenter that performed the work. Forging the audit means forging a signature.
Four surfaces, one substrate
The same cascade + receipt + joule-ledger exposes four discrete enterprise surfaces. Pick one or all; the receipts span them.
AI Gateway
One API in front of Imagine, Semantic, Math-Ground, and any external model you bring. The router picks the cheapest tier per query class, semantic cache absorbs near-duplicates, and every routed call returns the signed receipt naming the tier, the joules, and the datacenter that ran it.
Runtime Safety
Prompt-injection screening, PII redaction, output-policy checks, and hallucination guards run inline on every cascade call. Each guardrail decision — accept, redact, or refuse — emits its own receipt under the same signing key, so "why was this output blocked" has a verifiable answer.
Observability
Every inference, retrieval, tool call, and guardrail decision joins a single signed trace. Drill from a session to the per-token receipt chain, the cache hit-rate, the joule ledger, and the InformationOS citations the model consulted. The trace itself is a tamper-evident artifact.
Eval
Every prior answer is replayable from its receipt: the inputs, the tier sequence, the memoized fragments. A new model, a new policy, or a new prompt is graded against that frozen evidence set automatically — drift, regressions, and energy deltas all surface as signed diffs, not screenshots.
What this platform believes
A memoizing cascade keeps every prior answer addressable. Near-identical queries hit cache; the joules already spent stay spent. The bill goes down as the system learns.
Watch a memoized query →Every call returns a signed receipt with the energy consumed and the datacenter that performed the work. Tokens are downstream of joules; joules are the only universal billing unit.
Inspect the cascade →Imagine (perception), Semantic (language), Math-Ground (deterministic reasoning). All sit behind the same cascade. The router picks the cheapest tier that solves the query.
See the tiers →