Publication notes and known limitations

Publication hardening was completed against Kimi K3: Open Frontier Intelligence, arXiv:2607.24653v2, retrieved on 2026-08-12. The pinned source identity and checksum live in content/paper_coverage.yaml.

What this release includes

• A paper-aligned route through Courses 0–5B, with rendered Read chapters, visual explanations, prerequisite navigation, and nine inspectable notebooks.
• Explicit evidence boundaries for paper-reported claims, authored interpretations, and deterministic local miniatures.
• A generated paper map, concept library, claim ledger, and notebook catalog.
• Silent-first visual lessons with captions; recorded voiceover is not part of this release.

Known limitations

• The repository does not reproduce K3's 2.8T-parameter training run, production fleet, private data, agent environments, or benchmark harnesses.
• Local PyTorch code demonstrates selected invariants on small fixtures. It is not FlashKDA, a production MoE implementation, or evidence of million-token semantic retention.
• Retained notebook outputs and unit tests establish only the behaviors stated in their assertions and evidence entries.
• Paper results and case studies remain attributed reports unless an artifact is explicitly labeled as executed here.
• External leaderboards and third-party links can change after the recorded publication audit date.

Verification record

The release gate requires generated-content freshness, publication-integrity checks, 45 unit tests, strict MkDocs rendering, notebook execution smoke tests, browser/accessibility checks, and a clean git diff --check result.