Evidence policy

Claims should carry their evidence state.

Navami separates implementation, validation, measurement and future objectives so marketing language does not outrun technical evidence.

Public claim discipline

Validation is not the same as performance proof.

VALIDATED

Contract or integration checks passed.

A validated state supports only the specific implementation or behavior that was tested.

MEASURED

A declared workload produced a result.

A measurement is meaningful only with its workload, environment, configuration and failure conditions.

TARGET

Engineering objective, not present fact.

Targets stay out of promoted performance language until accepted benchmark evidence exists.

Current public boundary

Publish capabilities now. Publish performance numbers only with benchmark artifacts.

This public-safe preview intentionally omits unpublished throughput, latency, corpus-scale and reduction targets. Those remain internal engineering objectives until they are validated under a declared benchmark envelope.

Benchmark rule

Every promoted metric needs a reproducible envelope.

At minimum: workload definition, data modalities, environment, resource profile, operating mode, inclusion/exclusion boundary, pass/fail criteria and failure decomposition.

Technical evaluation

Bring a real workload.

We define the acceptance contract before measuring the system.