A capacity review should record the model, its version, the workload combination, and the input/output token mix. These fields are the core evidence needed to interpret a throughput result: documented throughput varies across model, version, and workload combination, while provisioned throughput per PTU depends on the model and the mix of input and output tokens in a given minute.
Core workload data
| Data field | What the review should preserve | Why it matters |
|---|---|---|
| Model | The model used for the measurement | Throughput varies by model. |
| Version | The specific model version associated with the result | Throughput can vary by version as well as by model. |
| Workload combination | The particular combination of workload conditions represented in the measurement | A result from one workload combination should not be treated as automatically describing another. |
| Input/output token mix | The input and output token composition for the relevant minute | Provisioned throughput depends on the model and this token mix. |
The token mix should remain attached to the measurement period rather than being replaced by an unexplained overall average. A file that lists only a model name and a throughput figure leaves out the version, workload combination, and token pattern needed to understand the conditions behind that figure.
How to check the record
A review can apply a simple consistency check:
- Every measurement should be linked to a model and version.
- Measurements from different workload combinations should remain distinguishable.
- The input/output token mix should be recorded for the same minute used to assess provisioned throughput.
- Comparisons should state whether the model, version, workload combination, or token mix changed.
Without those checks, a difference in observed throughput may reflect a change in workload conditions rather than a change in capacity alone.
What still needs confirmation
The cited guidance identifies which variables matter; it does not establish a universal throughput value for an unspecified workload. Before relying on a capacity estimate, the operator still needs to confirm the exact workload combination, the measurement window, the method used to count and classify input and output tokens, and current results for the intended model and version.
A projected throughput figure should therefore be treated as conditional on those recorded conditions, not as a guaranteed future outcome.