Feature Request: Include confidence score in AnonymizerEngine output (OperatorResult) for audit/traceability

[kshvg]

Is your feature request related to a problem? Please describe.

When building audit logging or compliance pipelines on top of Presidio, AnonymizerEngine currently drops the confidence score from each RecognizerResult after anonymization. The resulting OperatorResult only carries start, end, entity_type, text, and operator - the score is silently lost.

This makes it impossible to answer post-anonymization questions like:

• "With what confidence was this entity anonymized?"
• "Were any entities anonymized below our confidence threshold?"
• "Which documents need human review based on low-confidence detections?"

without re-running the analyzer on the original text, which may no longer be available or desirable.

Describe the solution you'd like

Add an optional score field to OperatorResult, threaded through from the originating RecognizerResult during anonymization:

The score should:

• Default to None so existing usage and the deanonymize path are completely unaffected
• Be included in JSON serialization
• Be preserved through conflict resolution (the surviving entity keeps its own score, not that of any dropped entity)
• Be deserializable via .from_json()

Describe alternatives you've considered
Zipping analyzer_results with anonymized.items by position - but this breaks because:

1. Text positions shift in the anonymized output
2. Conflict resolution drops some analyzer results, so the list lengths doesn't match

There is no reliable way to recover the score post-anonymization without carrying it through the engine.

Additional context
This is a purely additive change - score: Optional[float] = None requires no breaking changes to existing code. The score is already present on RecognizerResult, it just needs to be passed when constructing OperatorResult.

[kshvg]

I ran into this gap while building PII-LLM gateway on top of Presidio.

I've implemented this locally and have a working PR ready. The changes are as follows:

• operator_result.py - score: Optional[float] = None added to __init__, from_json
• entity.score passed when constructing OperatorResult
• Updated existing integration tests to include score in expected output
• Added new tests covering score propagation, conflict resolution behavior

All existing tests pass. The deanonymize path is unaffected since score defaults to None. Happy to open the PR or discuss the approach here first if preferred.

[SharonHart]

@kshvg Sounds good, looking forward for your PR