Multimodal Processing Summary Diffs¶

Compare two processing summaries without reopening assets or reading OCR, transcripts, or source content. Pass the earlier summary first; every numeric delta is after minus before.

from openmed.multimodal.processing_diff import (
    diff_processing_summaries,
    render_processing_diff_markdown,
)

# Both inputs are ProcessingSummary values from summarize_processing_run().
difference = diff_processing_summaries(before_summary, after_summary)
print(difference.to_json())
print(render_processing_diff_markdown(difference))

ProcessingDiff includes signed changes in asset count, bytes, duration, assets with output digests, per-media asset/byte/page/frame totals, terminal outcomes, and abstention stage/reason buckets. JSON and Markdown use fixed field and row order. Media types are alphabetized; outcomes retain their closed-code order; abstentions are sorted by stage and reason.

Digest changes compare the multiset of (input_sha256, output_sha256) pairs. An unchanged pair is omitted. A new or changed pair appears under added digests, and a missing or replaced pair under removed digests. Repeated pairs carry a positive count. The artifact does not include asset identifiers, file paths, media content, model output, or free text. A changed digest is evidence of a changed digest pair, not a measure of clinical quality.

The current diff schema version is 1. Inputs must be ProcessingSummary values using the current processing summary schema; other types or versions raise ProcessingDiffError. This comparison makes no release decision and never reads files.