The non-claims are part of the spec.

Does not claim

MLF 1.0 does not claim:

• complete Excel, Google Sheets, or LibreOffice behavioral compatibility;
• that matrix representation always improves model accuracy;
• that parallelizable structure guarantees physical speedup;
• that pseudonymization is anonymity;
• that a submitted authorization manifest proves legal authority;
• that model confidence grants permission to modify formal data;
• production-readiness of the included synthetic learning benchmarks.

The second and third are worth pausing on. A format built on the premise that structure matters could easily have claimed that preserving structure makes models better and makes work faster. It claims neither.

Known limits

From the v1.0.0 release notes:

• bounded rather than complete XLSX semantics;
• no VBA, Power Query, pivot execution, or external refresh;
• presentation round trips are not universally lossless;
• no production claim for learned dependency inference;
• included naturalistic workbooks are synthetic fixtures, not external enterprise validation.

Separations that hold the design apart

• A coordinate≠an identity — it says where a cell appears, not which object it is
• A projection≠a replacement for the structure it came from
• A checksum≠a fingerprint — one is integrity, the other is identity
• A prediction≠a fact
• A review decision≠a promotion

Status of the research layer

The v0.2–v0.9 modules are retained in the repository and remain explicitly non-authoritative. Negative experimental results are kept. Their outputs must be treated as governed or experimental artifacts rather than automatic truth.