Transcript Redaction QA Workbench

How to use Transcript Redaction QA Workbench

A redaction check should state exactly which ranges were intended to change. Compare a selected-range plan with an actual redacted result while keeping transcript content out of the exported review summary.

1. Paste only an authorized excerpt and manually identify every selected start/end range.
2. Enter the actual redacted result and the exact replacement token used for each range.
3. Resolve transformation and literal findings, then review unselected sensitive content independently before sharing.

Good to know

• This does not discover PII or prove complete anonymization. It only checks explicitly selected nonoverlapping code-point ranges and literal remnants; coincidental repeated text can produce a review finding.
• Summary exports omit source text, redacted text and labels, but retain neutral range IDs and offsets. An identifying range ID can still disclose information; choose neutral labels/IDs and inspect the export.
• Whitespace is preserved exactly. Code points differ from UTF-16 units and user-perceived grapheme clusters; a combining mark or joined emoji can span multiple offsets.
• Each field is limited to 8,000 UTF-16 characters; CSV inputs use the exact displayed header and at most 100 data rows. Quoted commas, doubled quotes and quoted newlines are supported. Empty/duplicate identities, unsupported controls and unpaired surrogates are refused.
• Exports are editable review records. Spreadsheet-like text is apostrophe-prefixed in CSV to reduce formula execution risk; inspect your actual spreadsheet/import settings before sharing.
• Everything runs locally in your browser. No AI model, upload, provider account, calendar, call, video render, automatic outreach, publication or saved input. Enter only information you are authorized to process; declarations remain unverified.