Verification & confidence
Every metric carries provenance and a confidence level. Confidence is never a bare number — it always comes with a reason.
Confidence levels
High
- Exact model
- Exact token breakdown
- Pricing date known
- Provider or verified export
Medium
- Exact model
- User-entered totals
- Pricing date selected
Low
- Subscription only
- Opaque credits
- Visible-browser estimate
- Total cost without usage evidence
Provenance types
Provenance records where a number came from. It determines the ceiling on how much we’ll trust that number — a cost is only as reliable as its shakiest input.
| Provenance | Meaning |
|---|---|
| Provider-reported | Reported directly by the AI provider’s usage API. The most reliable source. |
| Gateway-reported | Reported by a gateway or aggregator that sits in front of the provider. |
| Verified import | Parsed from a usage export the user supplied and we validated. |
| Verified screenshot | Read from a screenshot that a human reviewer confirmed. |
| Locally counted | Counted on the user’s own machine while they worked. |
| Visible-text estimate | Estimated from visible chat text. The provider’s real compute is not exposed. |
| Self-reported | Entered by a user without supporting evidence. |
| Estimated from tokens | Reconstructed by multiplying token counts by published pricing. |
| Estimated from credits | Reconstructed from opaque credits using an assumed credit value. |
| Subscription allocation | Allocated from a subscription price, not a metered charge. |
| Historical price estimate | Priced using the pricing that applied at the time of the project. |
| Demonstration data | Seeded demonstration data. Not a real project. |
| Unknown | The origin of this number is not known. |