What it would have cost somewhere else
One window of a workload, priced on a sku that did not serve it.
npm run aperture:counterfactual -- --workload convexity/deep-research --window 2026-08-01..2026-08-31 --candidate openai/gpt-5.6-lunaNothing is written. It is a report over the record, exactly as `aperture:reconcile` is.
Every figure in it is real
The units were counted by whatever served them. The prices were read off a provider's own document, archived and dated. The arithmetic is the reconciliation's, so a price that moved inside the window makes both figures bounds rather than points, and a range that straddles zero is printed as one rather than resolved into a saving that may not exist.
It runs nothing, so it scores nothing
No observed quality claim without an observed candidate response.
That sentence is the whole boundary, and the last caveat on every run says it in those words. No request was sent to the candidate, so nothing here says whether it would have answered as well, held its format, or called a tool successfully. A score for a request nobody answered has no source, and this record does not carry one.
A replay is the other thing and this is not it. A replay sends the saved requests to the candidate, so a candidate really answers and the score has a run behind it. That needs a capture the runtime does not perform: the ledger records metadata and holds no prompt or response, by design, so there is nothing on any disk to replay. See BUILD.md for where that sits.
What it refuses
| It refuses | Because |
|---|---|
| A window that served no units | There is nothing to have cost anything anywhere |
| A candidate that already served part of the window | That is a reading, not a counterfactual |
| A candidate the record cannot price across the whole window | A gap priced as zero would be the interpolation the tape refuses |
| A candidate from another market | A unit profile spends the dimensions of one market |
The caveat that is easy to miss
A window served by more than one sku is priced here as though one sku had served all of it, and the run says so. A split happens for reasons a price does not know about. A lane that split because one model could not answer some of the work would not have been cheaper on that model; it would have failed on it.
