The warehouses contain more rows than the public charts display. The dashboard applies explicit coverage, configuration, normalization, and latest-row rules before a value becomes public.
Result identity
Different layers require different identity fields:
| Layer | Fields that distinguish it |
|---|---|
| Deployment configuration | Model descriptor, quantization, runtime descriptor and version, device, and effective model/runtime flags |
| Public performance cell | Deployment configuration plus benchmark ID, metric, benchmark flags, and exact input/output token shape |
| Raw submission | Public performance-cell fields plus client identity, submission time, and stored provenance |
| Published quality value | Model, quantization, canonical evaluation variant, thinking mode, and pinned evaluation authority scope |
Deployment configuration
Public performance cell
Raw submission
Published quality value
A model, runtime, and device do not identify a unique public value without the benchmark and configuration fields. The descriptor format is documented in the pinned storage specification.
Separate warehouse paths
Performance and quality do not land in one result store. Performance aggregates are read from the performance warehouse connection. Quality aggregates are read from the evaluation warehouse connection. The dashboard filters each path independently and joins compatible public values in application code.
Performance selection
A performance row is eligible only when its model, quantization, device, runtime descriptor, runtime flags, benchmark ID, metric, benchmark flags, and token shape match the public configuration. The dashboard then selects the latest matching submission for that public cell.
Older or differently configured submissions can remain in raw history without being current chart candidates.
Quality authority selection
The evaluation warehouse can contain multiple scored rows for the same canonical evaluation from different clients, runtimes, or collection periods. The public dashboard does not use every matching raw row.
For each allowed model, quantization, canonical evaluation variant, and thinking mode, it selects the latest matching row from a pinned evaluation authority scope. That scope includes the main H100 authority path and explicitly pinned descriptor-based exceptions for specific coverage.
The authority-selected quality value is joined to compatible performance rows on every device that runs the quantization. Changing the performance device therefore changes speed and memory, but it does not select a different quality row. This is a dashboard publication policy, not a claim that generation is inherently independent of runtime or hardware.
The full-precision reference is the authority-selected row for the same model at its native full-precision weights. It can be FP16 or BF16, depending on the model's published weights.
Audit sources
Use the source that owns the claim being audited:
| Claim type | Primary audit source |
|---|---|
| Displayed performance values | Performance warehouse and the eligibility fields described above |
| Displayed quality values | Evaluation warehouse and the evaluation selection policy described above |
| Benchmark IDs and token catalog | Pinned pipette-clients standard benchmark source |
| Evaluation samples, attempts, scores | Pinned pipette-scores dataset builders, metadata, and scorer source |
Displayed performance values
Displayed quality values
Benchmark IDs and token catalog
pipette-clients standard benchmark sourceEvaluation samples, attempts, scores
pipette-scores dataset builders, metadata, and scorer sourceRaw history is useful for provenance, but it can contain older runtime versions, superseded evaluation variants, legacy field formats, and rows outside the current public scope.