Live demo · Layer 1 · Reasoning LOD
Let the large model plan. Let bounded workers do the work.
One frontier call signs a plan. Each step then runs at the cheapest tier that can prove its answer — plain code, a small model, a medium one — and climbs one tier only when the evidence fails verification. Approval is bound to the plan. Revocation stops the next run cold. Every event is a receipt an auditor can re-verify.
Offline demo. The workers are fixtures and the token counts are assigned demo values, not measurements. What is real: the reconciliation arithmetic, the verification, the contract, the approval gate, the revocation, and the receipt chain.
-
1
Plan once, then route
Three supplier invoices. Reconcile → brief → handoff. Reconciliation is pure code; the brief tries the small tier first.
-
2
The same job, frontier only
Same records, same plan, same verifiers — every model-capable step at the frontier.
-
3
Compare — honestly
The comparator refuses mismatched inputs, stopped runs, incomplete usage or unequal evidence before it will print a percentage.
-
4
Break the small tier
Inject a one-cent error into the small model's brief. The verifier recomputes from source, fails it, and the runtime climbs exactly one tier.
-
5
Approve the handoff
Reuse the signed contract with the handoff approved. No re-planning. The handoff writes into the signed report and does nothing external.
-
6
Revoke the contract, run it again
The contract's bytes and signature do not change. Only the revocation list does.
-
7
Verify the receipts — then tamper with them
The chain verifies. Change one number, or swap two receipts, and it does not.
Simulated counts on one invoice workload. The mechanism is real; the percentage is a demo value, not a measured saving.
The signed run receipt
Rendered from the same report the verifier checks. Human-readable; the JSON is what is signed.
Public demo — no login, no tenant data, rate-limited per IP. Each visitor gets an isolated
in-memory session that expires after 15 minutes; nothing is written to disk. The same
runtime is the axiom_reasoning_lod
MCP tool, where a server-configured provider file can replace the fixtures with real endpoints.