Skip to content

CONCEPT Cited by 1 source

Routing receipt

Definition

A routing receipt is the audit-grade record produced for every AI-agent task-routing decision in an OmniNode-shaped system. The receipt captures four fields, all required:

  1. Which model was chosen to handle the task.
  2. How many tokens it took (input + output).
  3. What it cost (zero for on-prem-routed, dollar-amount for cloud-fallback).
  4. Whether the output passed its compliance checks.

The receipt is the evidence half of the OmniNode "decision-is-a-contract / receipt-is-the-evidence" pairing — it makes cheapest-capable routing trustable rather than hopeful.

"Every routing decision produces a receipt: which model was chosen, how many tokens it took, what it cost, and whether the output passed its compliance checks… The decision is a contract. The receipt is the evidence. Neither lives in someone's head." — Source: sources/2026-06-02-redpanda-how-omninode-uses-redpanda-to-scale-ai-agent-workflows

Why receipts are load-bearing for cheapest-capable routing

Cheapest-capable routing degenerates into hopeful cost reduction without a verification step. The receipt is the verification step. It serves three distinct functions:

  1. Cost attributionwhich tasks routed to which model produced what spend? Without per-decision receipts, the monthly bill is a mixed-fleet aggregate that's hard to reconcile against any unit of business value.
  2. Quality auditing — did the output pass the compliance check? "When a local model can't meet the bar (output is too short, missing citations, or hallucinated identifiers), the task automatically escalates to a stronger model." The receipt records which checks passed and which triggered escalation. OmniNode's disclosed week-of escalation rate: 1.3% of delegations.
  3. Routing-decision feedback — receipts can feed back into the classifier that decides which model class a task routes to. If a class consistently escalates above a threshold rate, the classifier should learn to send those tasks to the stronger model upfront.

What the receipt is not

The OmniNode post is careful about scope:

  • Not just cost data — without the compliance-check field, you'd know what you spent but not whether it was well-spent.
  • Not just compliance data — without the cost field, you can't do cost optimisation.
  • Not stored per task only in operator memory"neither lives in someone's head." The receipt is materialised somewhere durable.

Sibling framings on the wiki

Caveats

  • Storage substrate undisclosed: the OmniNode post says receipts are produced and surfaced on the OmniNode dashboard, but doesn't characterise where they live (event-bus topic? audit table? queryable warehouse?), retention, indexing, or query shape.
  • Schema undisclosed: the four named fields (model / tokens / cost / compliance-pass) are illustrative; whether the receipt records the input prompt, the output text, the classifier's reasoning, or the per-check breakdown of compliance is undisclosed.
  • Cost calculation for on-prem: the post says on-prem is "zero marginal cost" — but a serious cost model would amortise hardware, electricity, and operator time. The receipt's "cost" field appears to be marginal-cost, not amortised.
  • Compliance-check definition: "output is too short, missing citations, or hallucinated identifiers" are illustrative examples; the full check rubric, per-task-class threshold configuration, and false-pass / false-fail rates are undisclosed.

Seen in

  • sources/2026-06-02-redpanda-how-omninode-uses-redpanda-to-scale-ai-agent-workflows (2026-06-02, OmniNode founder Jonah Gray on Redpanda Blog) — canonical disclosure source. Provides the four-field receipt schema (model / tokens / cost / compliance-pass) and the decision/evidence/recovery triad framing: "cheap routing only works if you can trust it"; "the decision is a contract. The receipt is the evidence." Names three load-bearing functions: cost attribution, quality auditing, escalation evidence. Disclosed week-of metric: 1.3% of delegations escalated with receipt-recorded reason.
Last updated · 542 distilled / 1,571 read