What changed in a model folder?
File identity, structural findings, and the distinction between recording a baseline and verifying it.
Research note · November 13, 2026Research question
How can a reviewer identify changed, missing and added files without confusing byte integrity with model quality?
Method
- Record a model folder's paths, sizes and SHA-256 digests in a lockfile. Preserve that baseline outside the test folder before changing the fixture.
- Create separate cases for an unchanged folder, a modified configuration, a missing file and an added file. Verify each against the original baseline; do not overwrite the baseline to make a mismatch disappear.
- Review structural findings separately: malformed JSON, truncated weight files and missing shards are different from a digest mismatch. Keep the original file and the reported finding together.
- Use the Prompto crystal to inspect the recorded files. Compare the underlying record when deciding whether a change is expected; the drawing is a view of that record.
Records to retain
- Original lockfile and its digest
- Fixture file paths and before/after digests
- Command or app version and check result
- Structural findings and baseline-review decision
Current evidence
Prompto implements local file recording and comparison. Its documented checks distinguish file changes from structural problems. This page describes the workflow; it does not publish a new benchmark run.
Interpretation & limits
Matching bytes do not establish model safety, task performance or trusted training provenance. The browser demo uses synthetic files and does not run the desktop's native structural checks.
Next experiments
- Publish a versioned fixture manifest with expected outcomes for each case.
- Record desktop and CLI results against the same fixtures and preserve disagreements.