Changelog

0.2.0 — 2026-10-05

• Add optional specialist discovery and bounded handoff contracts, with reviewed upstream provenance.
• Separate semantic invariants from declared execution variables in benchmark schema 2.
• Add five original training decision cases with explicit evidence maturity.
• Add runnable skill evaluation fixtures, deterministic checks, baseline/treatment preparation and an external runtime protocol.
• Preserve legacy benchmark inputs and include the supplied specialist survey.

0.1.0 — 2026-10-05

• Introduce the optimize-training skill with six focused training-system references.
• Include the complete supplied v2 Chinese guide.
• Add environment, benchmark, sample-audit, and NPZ parity tools.
• Add synthetic examples, behavioral evaluation cases, CPU tests, CI, local installer, and reproducible release packaging.
• Document correctness boundaries and explicitly separate package validation from GPU training experiments.