Validation and performance evidence

V1 validation is anchored in executable evidence rather than historical timing tables or narrative completion reports.

Numerical evidence

• Language-neutral normal, edge, and invalid fixtures live under specs/core-api/fixtures/ and specs/numerical-reference/.
• Rust property, fuzz, malformed-input, concurrency, and differential tests exercise the stable kernels.
• Python, R, and Julia conformance tests call the Rust implementation through PyO3 or the versioned C ABI.
• Compatibility and result-schema tests enforce tolerances, deterministic seeds, error identities, and serialization behavior.

Performance evidence

Committed benchmark workloads and budgets are enforced by the Rust and hosted benchmark lanes. Results are runner-specific; a historical wall-clock number must not be treated as a portable performance guarantee. Accelerator or hardware claims require their own parity packet, runner identity, transfer and compilation overhead, and CPU-reference comparison.

Release evidence

The release candidate must pass the complete CI matrix, clean wheel and sdist installations, cross-platform reproducibility, SBOM and vulnerability scans, Astro production build, retained-binding checks, and signed immutable-tag verification. Hosted artifacts and registry observations remain distinct from repository readiness.