Caching the operator tunes — with proof it ran.
Instead of a static cache you hand-configure and hope pays off, the operator learns each namespace's TTL and similarity threshold from your own traffic — and writes a verifiable receipt for every call it avoids, so the saved cost is attributable, not invisible.
The operator caches without creating a blind spot
The operator serves repeat and near-duplicate requests from cache. In simple mode it keys on the exact request; in semantic mode it matches by similarity above a threshold it tunes per namespace. Either way it records the hit — instead of taking a silent shortcut around your ledger, it leaves a verifiable event.
Every hit links to the cached response on the same chained record as a live dispatch. When finance asks why a call cost $0, or an auditor asks what was returned, the operator already wrote the answer.
- Simple + semantic modes. Exact-match, or similarity-match against a threshold the operator tunes.
- Per-namespace TTL. The operator scopes each cache to a use case and learns when to expire it.
- Every avoided call is a receipt. The saved cost stays attributable, never invisible.
Every avoided call is a verifiable receipt
Instead of letting savings vanish from your audit trail the way most caches do, the operator chains every hit — so a cached answer carries the same proof as a fresh one, and you can check the avoided cost yourself.
The receipt distinguishes a hit from a live call
The operator receipts and chains a cached dispatch like any other, with metering that reflects the saved cost. Recompute the leaf to confirm what was served — and that no one altered the record after the fact.
Same client-side check that runs on the Trust portal. SHA-256 runs in your browser via Web Crypto. No Agentics call.
- •Recompute leaf
- •Chain integrity
- •Fold Merkle proof
- •Anchor memo == root
SHA-256 runs in your browser via Web Crypto. No Agentics call.
You set the bounds; the operator tunes inside them
Instead of hand-tuning thresholds you'd have to revisit every week, name the namespace and the operator learns the TTL and similarity threshold from your traffic — within the ceiling you set.
{
"version": 2,
"strategy": { "mode": "single", "targets": [{ "provider": "anthropic" }] },
"cache": {
"mode": "semantic",
"ttl_s": 3600,
"namespace": "support-triage",
"threshold": 0.92
}
}
Watch what the operator avoided
The console shows hit rate, saved spend, and the receipt behind each call the operator chose not to make.
Response Cache console
Hit rate and saved spend per namespace, with a drill-down to the receipt behind every call the operator avoided.
Open Response Cache →