Appearance

Critical Document Audit ​

Paste or upload a policy, contract, incident report, or medical note. Extract findings, then audit the evidence spans with HLM3-Mix signals.

Status

Demo-ready browser workflow. The default checkpoint is the validated HLM3-Mix 35M K=16 (val PPL 10.66); the all-models edition exposes the bundle's broader language checkpoint catalog as selectable audit models.

What the demo proves ​

The same Hopfield core that drives language generation produces span-level audit signals over a document:

  • Per-finding PPL, entropy, and prefix delta for each evidence span
  • A document hash so the audit can be replayed
  • A clear fragile / review / stable status per finding
  • Saved JSON and Markdown reports for downstream review

What the reviewer sees ​

  • A profile selector (contract / policy / incident / medical)
  • The original document with findings extracted from deterministic rules
  • Each finding's evidence span audited by HLM3-Mix
  • A downloadable JSON of the full audit run

Two editions ​

EditionModels availableUse it for
DefaultHLM3-Mix 35M K=16 onlyReliable client demo with the validated checkpoint
All-modelsThe full packaged checkpoint catalogSide-by-side audit behavior across the bundle

The all-models edition uses the same audit pipeline; it adds a sidebar checkpoint selector with the tier badge and caveat for each option. Non-validated checkpoints surface a clear warning.

What "all-models" reveals ​

Selecting different checkpoints lets a reviewer see how the audit signals shift with the underlying model. The validated checkpoint is the quality reference; other entries in the bundle are diagnostic surfaces with explicit in-UI caveats so a reviewer cannot mistake an experimental artifact for a quality claim.

Caveats ​

  • The deterministic rules cover four profiles; broader risk schemas need a partner integration.
  • Larger checkpoints in the bundle can need substantial VRAM. The UI surfaces this and offers a CPU fallback.

Where it fits ​

The right demo for compliance, legal, clinical, or operations reviewers who care about evidence traceability over single-answer generation.