Multimodal Processing Summaries¶

At the end of a multimodal run, aggregate what happened to every asset into a single, inspectable completion artifact with summarize_processing_run(). The result, ProcessingSummary, contains only counts, byte/page/frame totals, and opaque asset_id/SHA-256 values — never OCR text, transcripts, DICOM tag values, file paths, or any other source content.

from openmed.multimodal.abstention import AbstentionReason, AbstentionRecord, AbstentionStage
from openmed.multimodal.asset_manifest import AssetManifest
from openmed.multimodal.digest import AssetDigest
from openmed.multimodal.processing_summary import (
    AssetProcessingResult,
    ProcessingOutcome,
    render_processing_summary_markdown,
    summarize_processing_run,
)

results = [
    AssetProcessingResult(
        manifest=AssetManifest(
            asset_id="asset-001",
            media_type="image/png",
            sha256="a" * 64,
            byte_size=204800,
        ),
        outcome_code=ProcessingOutcome.SUCCESS,
        duration_seconds=0.42,
        input_digest=AssetDigest(sha256="a" * 64, byte_count=204800),
        output_digest=AssetDigest(sha256="b" * 64, byte_count=1024),
    ),
    AssetProcessingResult(
        manifest=AssetManifest(
            asset_id="asset-002",
            media_type="audio/wav",
            sha256="c" * 64,
            byte_size=512000,
        ),
        outcome_code=ProcessingOutcome.ABSTAINED,
        duration_seconds=0.05,
        input_digest=AssetDigest(sha256="c" * 64, byte_count=512000),
        abstention=AbstentionRecord(
            stage=AbstentionStage.INFERENCE,
            reason=AbstentionReason.PHI_UNCERTAINTY,
        ),
    ),
]

summary = summarize_processing_run(results)
print(summary.to_json())
print(render_processing_summary_markdown(summary))

What this guarantees¶

Metadata-only. Every field on ProcessingSummary is a count, a byte/page/frame total, a stable code, or an opaque asset_id/SHA-256 value already validated as non-identifying by AssetManifest and AssetDigest.
Fail closed. summarize_processing_run() validates every asset's outcome_code against the fixed ProcessingOutcome set before aggregating anything. An unrecognized code raises ProcessingSummaryError and no summary — partial or otherwise — is produced.
Deterministic. by_media_type, outcome_counts, abstention_counts, and asset_digests are always sorted, so to_dict()/to_json() are identical no matter what order assets were processed in.

What this does not do¶

Out of scope for this artifact: it does not store source content, generate thumbnails, back a dashboard, or make any clinical-quality claim about the run. It is a completion record, not an audit log of what a model saw.

Schema¶

PROCESSING_SUMMARY_SCHEMA_VERSION is currently 1. Adding, removing, or reinterpreting a field on ProcessingSummary requires bumping this version.

Field	Type	Meaning
`schema_version`	int	Contract version, currently `1`.
`total_assets`	int	Number of assets in the run.
`total_bytes`	int	Sum of `manifest.byte_size` across all assets.
`total_duration_seconds`	float	Sum of per-asset `duration_seconds`.
`by_media_type`	list	One entry per distinct `media_type`, sorted by that string: `media_type`, `count`, `total_bytes`, `total_pages`, `total_frames`.
`outcome_counts`	list	One entry per `ProcessingOutcome` present, in enum declaration order: `outcome`, `count`.
`abstention_counts`	list	One entry per distinct `(stage, reason)` pair present, sorted by `(stage, reason)`: `stage`, `reason`, `count`.
`asset_digests`	list	One entry per asset, sorted by `asset_id`: `asset_id`, `input_sha256`, `output_sha256` (omitted when there was no output).
`asset_count_with_output_digest`	int	Number of assets that produced an output digest.

Outcome codes¶

ProcessingOutcome is a fixed, closed set:

Code	Meaning
`success`	The asset was processed to completion.
`abstained`	Processing stopped via an `AbstentionRecord`; see Abstention Reasons.
`error`	Processing failed for a reason other than a documented abstention.

Any outcome_code outside this set — on any asset in the run — causes summarize_processing_run() to raise ProcessingSummaryError immediately. There is no "unknown" bucket; the boundary fails closed rather than silently guessing.

Example output¶

{"schema_version":1,"total_assets":2,"total_bytes":716800,"total_duration_seconds":0.47,"by_media_type":[{"media_type":"audio/wav","count":1,"total_bytes":512000,"total_pages":0,"total_frames":0},{"media_type":"image/png","count":1,"total_bytes":204800,"total_pages":0,"total_frames":0}],"outcome_counts":[{"outcome":"success","count":1},{"outcome":"abstained","count":1}],"abstention_counts":[{"stage":"inference","reason":"phi_uncertainty","count":1}],"asset_digests":[{"asset_id":"asset-001","input_sha256":"aaaa...","output_sha256":"bbbb..."},{"asset_id":"asset-002","input_sha256":"cccc..."}],"asset_count_with_output_digest":1}

# Processing Summary

- Schema version: 1
- Total assets: 2
- Total bytes: 716800
- Total duration (seconds): 0.47
- Assets with an output digest: 1

## By Media Type

| Media Type | Count | Total Bytes | Total Pages | Total Frames |
| --- | --- | --- | --- | --- |
| audio/wav | 1 | 512000 | 0 | 0 |
| image/png | 1 | 204800 | 0 | 0 |

## Outcomes

| Outcome | Count |
| --- | --- |
| success | 1 |
| abstained | 1 |

## Abstentions

| Stage | Reason | Count |
| --- | --- | --- |
| inference | phi_uncertainty | 1 |

## Asset Digests

| Asset ID | Input SHA-256 | Output SHA-256 |
| --- | --- | --- |
| asset-001 | aaaa... | bbbb... |
| asset-002 | cccc... | |