[binaryaaron]

In Rewrite mode, Anonymizer.evaluate() now sets needs_human_review to True on any record whose coverage judge reports a non-empty missed_entities. Until now the judge wrote entity_coverage and missed_entities and left the flag alone, so a name detection never caught could reach the output verbatim with needs_human_review False.

Why

In a 25-record run with a local open-weight model at low reasoning effort, GLiNER and the LLM augmenter both missed two subject surnames, which survived the rewrite verbatim. The coverage judge in evaluate() caught both, but nothing acted on it: 7 of 25 records had a non-empty missed_entities, 4 of them unflagged. A hosted model had found the same surnames every time, so the gap depends on the augmenter's model and effort, and the flag is where a user looks.

What changes

After the coverage judge runs in evaluate(), rows whose missed_entities is a non-empty list get needs_human_review = True. Empty lists, None and NaN leave the flag unchanged; frames without the column (Replace mode) are returned as is; the Boolean dtype is preserved; the run() result is not modified. docs/concepts/rewrite.md, docs/concepts/evaluation.md, docs/troubleshooting.md and the rewrite-evaluation devnote each gain one sentence naming the new cause.

13 source lines. No public signature, column or default changes; detection, the rewrite workflow and the repair loop are untouched.

Checks

• ruff format --check, ruff check and make typecheck pass; make test: 1,587 passed.
• New parametrized test in tests/interface/test_anonymizer_interface.py with a stubbed EntityCoverageWorkflow through the real evaluate() path: a missed entity flags the record; an already-flagged record stays flagged; no missed entities and an unavailable coverage result leave the flag unchanged. Each case asserts the boolean dtype and untouched run() flags. The first case fails without the fix.
• The coverage workflow normalizes its output to Python lists before the interface, including through the Arrow-backed dataset reader, so the list check is sound.

For reviewers

• Targeted test: tests/interface/test_anonymizer_interface.py (the evaluate coverage-flag cases).
• Confirm: Replace results unchanged; run() output byte-identical; needs_human_review keeps its Boolean dtype.
• refactor(detection): establish deterministic entity candidate lifecycle #305 and fix(detection): augment plain source on candidate lifecycle #267 change what the augmenter is shown, which affects how often the judge finds something; this change only acts on what the judge reports.

🤖 Generated with Claude Code

[greptile-apps]

Confidence Score: 5/5

[Medium impact] The PR appears safe to merge; no actionable issues were found.

Summary

Rewrite-mode evaluate() now raises needs_human_review when the coverage judge reports missed entities.

• Rewrite evaluation flags rows when the coverage judge finds missed entities.

_{Reviews (5) · Last reviewed commit: "docs(evaluation): note that an unflagged..." · Reviewed by Greptile}

[lipikaramaswamy]

cc @memadi-nv could you review?

[memadi-nv]

Thanks for catching this edge case Aaron, LGTM! Just added some clarification note to the docs as a suggestion.