Auditable Receipt Extractor v0.4.0

Auditable Receipt Extractor 0.4.0 is a source release of the bounded multimodal receipt pipeline.

Highlights

• validates the complete PNG/JPEG/WebP batch before any provider request
• targets a strict four-field typed Responses boundary and redacts provider failures at the CLI
• replays exact recorded batches offline without importing OpenAI
• creates and verifies content-addressed local run bundles bound to inputs, schema, manifest bytes, ordered names, and typed outputs
• includes a provider-free exact evaluator with full report recomputation
• ships reproducible synthetic inputs plus real CLI, coverage, diagram, confusion-matrix, and GIF evidence
• passes offline quality, distribution smoke tests, and extended CodeQL for Python and Actions

Claim boundary

The committed demo and evaluation suite are deliberately synthetic. They do not prove that a model ran and do not establish live-model accuracy, authenticity, authorship, privacy, or production suitability. Live mode can upload sensitive receipt bytes and metadata to a provider and requires explicit acknowledgement plus the operator’s own current data-governance review.

No prebuilt binaries or live-provider outputs are attached. GitHub provides the source archive for the reviewed commit.