An executed run tied to an immutable model revision, a versioned dataset, a code hash, runtime metadata and inspectable raw artifacts. Results require an explicit review before publication.
What doesn’t?
Being listed in the catalog, a popularity signal, a publisher’s model-card claim, or passing our runner’s fixture tests. Those are useful context, but none means “independently evaluated by Suvyo.”
EXECUTABLE METHODOLOGY
Start small. Make it reproducible.
Suvyo Route: structured tool routing
Suite v1.0.0
A small deterministic English/Hindi routing smoke test. It is not a general capability, safety, or production-readiness benchmark.
Sixteen fixed English/Hindi cases route to four inert tools. Full-response JSON parsing, exact argument schema validation, tool-name accuracy, and exact-match scoring. No repairs or retry selection. Each run pins model, dataset and code revisions. Human review is required before publication.
The runner and review workflow are in place. A real model must be executed and its evidence reviewed before a score appears here. Test fixtures are never published as model results.
Watch the Proof methodology explainer
Five labels. Five different meanings.
Listed: a discoverable record with source attribution.
Source synchronized: metadata retrieved from the original provider.
Metadata reviewed: a curator checked specific metadata evidence.
Third-party claim: a benchmark or statement made by someone else.
Independently evaluated: a documented Suvyo execution approved for publication.
The limitations belong beside the result.
A small routing suite measures a narrow task. It does not establish general intelligence, safety, translation quality or fitness for every business. Results from different datasets, prompts, revisions or runtime settings are not automatically comparable.