Infer's local receipt inspector: token-charge arithmetic with explicit limits

[tarunspandit]

We build Infer. We made a small browser tool for a question that token counts alone do not answer: does a request's displayed charge follow its displayed rates?

Try the receipt inspector. No account is needed to load the example.

Choose Load fixture, then Inspect locally. The demo below uses a synthetic documentation fixture, not a customer bill or a new inference request. Its illustrative rates are not current offer prices.

34-second product walkthrough (MP4 download).

The inspector uses integer micro-dollar arithmetic to check token relationships and reconstruct the token charge. In this fixture, 12 fresh input tokens and 6 output tokens reproduce 0.000060 USD, with a difference of 0.000000 USD.

A standalone Infer usage receipt includes the rate table needed for that calculation. A normal Responses result can be checked for metadata consistency but lacks those rates. Receipt text stays in the tab; ordinary page and asset requests still reach Infer.

The limits matter: matching arithmetic does not prove who issued the JSON, which model deployment served it, or whether the wallet settled. The current receipt schema also omits a fixed per-request rate, so the inspector cannot independently explain every positive difference. This is a standalone tool for Infer receipts; it does not import LLM's logs or install an LLM plugin.

For people building cost reports, which receipt fields are most useful when reconciling a request with your own logs?