Distribution by Entity Type
Recogniser Activity
Low-confidence events sit at the top of the corpus when sorted by score ascending. These are the redactions the system was uncertain about. Refinement effort here has the highest leverage. Open the corpus browser, sort by confidence ascending, and walk the bottom decile first.
Open corpus browser| Time | Entity | Score | Recogniser | Replacement | Context | |
|---|---|---|---|---|---|---|
| Loading corpus... | ||||||
The refinement workbench will ship in v3.2. From here the administrator will be able to convert corpus observations into recogniser updates, allow-list and block-list entries, and recogniser pattern adjustments. All actions will write to a versioned refinement journal that becomes part of the audit trail.
Until v3.2 ships, refinement is performed manually in the recognisers.py file by the maintainer or the institution's Enterprise consultancy contact, informed by the corpus export available on the next tab.
Corpus export
Stream the full corpus as a JSONL file. Each line carries a hashed original, a cleartext replacement, the recogniser, the score, and the masked sentence context. The hash is computed with an install-specific salt; original values are not recoverable from the export.
Clear corpus
Archive the current corpus to a .backup.jsonl file and start a fresh corpus. Use this after a retraining cycle or when starting a new audit window. The archived file is never deleted automatically.