Trained-source conversion

Documents the raw-f32 source boundary, source manifest validation, exact tensor files, conversion modes, provenance transition, and restrictions on trained claims.

Source layout #

The packer accepts a directory containing a line-based source.manifest and raw f32 tensor files. Paths must be safe and relative; files must have exact expected byte lengths and checksums.

Value checks #

Source weights must decode to finite, bounded values and contain a nonzero payload. Tensor names and shapes must match the declared model contract.

Conversion outputs #

convert-trained can emit f32, q8_0, or q4_0 SLM artifacts and corresponding manifests. The source kind changes to converted-trained rather than deterministic-smoke.

What conversion proves #

It proves byte-safe intake, shape-compatible encoding, and reproducible transformation under the packer. It does not prove the source was genuinely trained, licensed, useful, safe, or compatible with an intended tokenizer unless those facts are independently bound and reviewed.

Release evidence #

Retain source manifest, every raw tensor hash, conversion command/version, output hash, quantization settings, and reference evaluation sidecar.

Scope #

This starter page defines the questions, boundaries, evidence, and failure modes that should be recorded before a capability is presented as supported.

Engineering considerations #

• Identify the source, version, target environment, and owner.
• Separate observed values from estimates and externally reported values.
• Record trade-offs, unsupported cases, and fallback behavior.
• Link performance statements to a compatible benchmark methodology.

Verification questions #

• What exact artifact, revision, backend, and environment were reviewed?
• Which assumptions could change the result?
• Which data should be retained so another engineer can reproduce the conclusion?