Model and runtime configurations
models.json: the 13 model names, serving setups, configured API identifiers or pinned weight revisions, and the actual GPU type used for the selected worker timings.result_configurations.json: configurations linked from individual saved results. Use these for exact provenance rather than assuming a model's current default settings.worker_sources.json: model artifact paths and loader checksums for the native worker.worker_environment.json: dependency-file paths, GPU types, isolated execution conditions, and versioned compiler packages.native_input_audits.json: token counts and encoding hashes for checking Decider and GLiDE input encodings.
API identifiers can be aliases. Weight revisions are pinned Git commits. See methods before interpreting the saved timings or attempting fresh inference.
measurement_conditions.json records the shared physical GPU verification, resource settings, runtime versions, first-forward exclusion, API cost definitions, and measurement-provenance hashes.
recovery_provenance.json identifies the experiment code used for deterministic reconstruction of held inputs.
The hosted-model setup table maps model names to command aliases and Gateway IDs. The self-hosted setup table maps each native model to its dependency environment.
examples/local-model.json and examples/provider-model.json are starting points for your own connections. Copy them into .local/ and follow Connect a model; they contain no credentials and do not alter the study's configurations.
Variant study
variant-study.json records the altered-input inventory, model configurations and coverage. variant-labels.json supplies readable condition names. Exact comparison input lists live in comparisons.json.