model-version
Required for reproducible comparison and post-hoc audit.
Best model for which task, data regime, evidence boundary and operating constraint?
The Arena freezes the benchmark before scoring, records exact versions and parameters, can hide evaluation labels, requires contamination assessment and ranks models per metric. BioAtlas deliberately does not collapse accuracy, calibration, robustness, compute and cost into one universal winner.
Frozen execution contract
Hidden-label, version-aware evaluation.
checkpoint
Required for reproducible comparison and post-hoc audit.
container
Required for reproducible comparison and post-hoc audit.
parameters
Required for reproducible comparison and post-hoc audit.
hardware
Required for reproducible comparison and post-hoc audit.
dataset-version
Required for reproducible comparison and post-hoc audit.
random-seed
Required for reproducible comparison and post-hoc audit.
license
Required for reproducible comparison and post-hoc audit.
Example
3 Pareto-efficientMetric-specific results and Pareto frontier.
| Model | Task accuracy | Calibration | Failure rate | Latency | Pareto |
|---|---|---|---|---|---|
| Example Model A · 1.0 | 0.88 | 0.79 | 0.05 | 9 | Yes |
| Example Model B · 2.1 | 0.91 | 0.7 | 0.08 | 6 | Yes |
| Example Model C · 0.9 | 0.86 | 0.87 | 0.03 | 12 | Yes |
BioAtlas does not generate a universal winner. Rankings are metric-specific; the Pareto frontier identifies models that are not dominated across the declared metrics.