# Inference cost experiment ledger

- Workflow and accepted-result definition:
- Baseline/candidate configuration and fixed workload:
- Price sheet date, currency and unit:

| Mutually exclusive billing category | Baseline usage/rate/cost | Candidate usage/rate/cost | Source |
| --- | --- | --- | --- |
| Uncached input | | | |
| Cache creation if separately billed | | | |
| Cache reads | | | |
| Billed output including reasoning where applicable | | | |
| Tools/storage/other separate meters | | | |
| Additional review/operations | | | |

Include retries in the underlying usage rows; do not price them twice. Reconcile whether cached input is a subset of total input and whether reasoning is already billed as output.

Illustrative arithmetic: 40 + 16 + 80 + 14 = 150; reducing output cost from 80 to 64 gives 134 gross, but 25 extra operating cost gives 159 net. These are invented euros/rates, not provider prices or a benchmark.

- Per-slice quality and uncertainty:
- Truncation, retries and user re-prompts:
- Cache access/revocation/source-version tests:
- Net decision and next review trigger:
