docs: add devnote for hybrid entity detection

[asteier2026]

Summary

• Adds a devnote explaining Anonymizer's three-stage entity detection pipeline (candidate detection, validation, augmentation) and why each stage is handled by a specialized model (GLiNER for structured extraction, LLM for reasoning).
• Includes benchmark results comparing Anonymizer against Azure AI Language, Rescriber, Privacy Filter, and GLiNER2-PII on OpenPII and Nemotron PII, plus an ablation study isolating each pipeline stage's contribution.

Test plan

• Rendered locally via make docs-serve and reviewed the post, images, tables, and footnote citations
• make format-check

🤖 Generated with Claude Code

[greptile-apps]

Greptile Summary

Adds a developer note describing Anonymizer’s hybrid entity-detection architecture and benchmark results.

• Explains candidate detection, validation, and augmentation as separate pipeline stages.
• Clarifies that validation applies only to first-stage candidates and that augmented entities are added afterward without another validation pass.
• Provides benchmark, ontology-coverage, and ablation-study results with supporting visualizations.

Confidence Score: 5/5

The PR appears safe to merge.

No blocking failure remains.

Important Files Changed

Filename	Overview
docs/devnotes/posts/hybrid-entity-detection.md	Adds the hybrid-detection developer note and now accurately scopes validation to candidates produced before augmentation.
docs/devnotes/posts/assets/hybrid-entity-detection-pipeline.png	Adds the pipeline illustration referenced by the developer note.
docs/devnotes/posts/assets/hybrid-entity-detection-ablation-openpii.png	Adds the OpenPII ablation-study visualization.
docs/devnotes/posts/assets/hybrid-entity-detection-ablation-nemotron.png	Adds the Nemotron PII ablation-study visualization.
docs/devnotes/posts/assets/hybrid-entity-detection-openpii-pr.png	Adds the OpenPII precision-recall benchmark visualization.
docs/devnotes/posts/assets/hybrid-entity-detection-nemotron-pr.png	Adds the Nemotron PII precision-recall benchmark visualization.

Flowchart

%%{init: {'theme': 'neutral'}}%%
flowchart LR
    D[Original document] --> C[Candidate detection<br/>GLiNER2-PII]
    C --> V[Validation<br/>Keep, drop, or reclassify]
    V --> A[Augmentation<br/>Recover missed explicit entities]
    A --> F[Final entity set]

Loading

_{Reviews (3): Last reviewed commit: "docs: fix Rescriber recall values in hyb..." | Re-trigger Greptile}

[greptile-apps]

Validation coverage is overstated

Validation decisions are generated before augmentation, and augmented entities without matching decisions are kept unchanged. Saying that validation reviews “each candidate” therefore gives readers an inaccurate model in which augmented entities also receive keep/drop/reclassify review.

Note: If this suggestion doesn't match your team's coding style, reply to this and let me know. I'll remember it for next time!

[asteier2026]

Made suggested changes

[lipikaramaswamy]

The Rescriber recall values in the plots conflict with the tables. The OpenPII plot places Rescriber at approximately 0.52 recall, while Table 1 reports 83.0%. The Nemotron plot likewise shows approximately 0.52, while Table 2 reports 67.8%. This appears specific to the Rescriber data, other plotted values seem to match their table entries 🤔

[asteier2026]

Good catch. I fixed it.

[lipikaramaswamy]

Thinking about this some more, I wonder if this is truly accurate 🤔

The augmenter receives the post-validation tagged text and validated entity list. A candidate dropped during validation is therefore untagged and absent from that list, allowing the augmenter to propose it again. This stage searches for entities missing from the validated set—not exclusively entities that GLiNER never proposed. Maybe it's just semantics, but just raising in case we want to clarify

[asteier2026]

I think it's fine. What are the odds of gliner suggesting something, then validate drops it, then augment finds again. Pretty slim.

[lipikaramaswamy]

This one is from agents: The OpenPII ablation reports an F1 score of 0.837 for augmentation alone and 0.826 for the complete pipeline, as the article later acknowledges. Should we describe it as achieving the highest recall, or explicitly state the privacy-weighted objective being optimized?

[asteier2026]

I explain this later, that all together gives us the highest recall, but just aug gives slightly higher F1, then tried to say that recall is what we want.