[P0] Implement local OCR and pre-persistence secret redaction #9

New issue

Closed

opened 2026-07-17 23:59:06 +00:00 by lost-rob0t · 0 comments

lost-rob0t commented

2026-07-17 23:59:06 +00:00

(Migrated from github.com)

Goal

Ensure sensitive text and image regions are removed before any captured record is persisted.

Requirements

Local OCR provider with no network access.
Detect API keys, access tokens, private keys, passwords, authorization headers, connection strings, high-entropy values, emails, usernames, and configurable patterns.
Return bounding boxes and redact matching pixels from screenshots.
Redact corresponding OCR text and metadata fields.
Conservative handling for low-confidence findings.
Per-pattern allowlists must be explicit, narrow, and auditable.
The model may assist with classification only after deterministic filters; it may never be the sole secret control.

Acceptance criteria

Raw OCR output and unredacted frames are never persisted or logged.
Synthetic secret corpus tests cover common providers and encoding variants.
High-entropy detection has configurable thresholds and measured false-positive/false-negative fixtures.
A redaction failure rejects the entire record rather than storing it unfiltered.

Depends on

## Goal Ensure sensitive text and image regions are removed before any captured record is persisted. ## Requirements - Local OCR provider with no network access. - Detect API keys, access tokens, private keys, passwords, authorization headers, connection strings, high-entropy values, emails, usernames, and configurable patterns. - Return bounding boxes and redact matching pixels from screenshots. - Redact corresponding OCR text and metadata fields. - Conservative handling for low-confidence findings. - Per-pattern allowlists must be explicit, narrow, and auditable. - The model may assist with classification only after deterministic filters; it may never be the sole secret control. ## Acceptance criteria - Raw OCR output and unredacted frames are never persisted or logged. - Synthetic secret corpus tests cover common providers and encoding variants. - High-entropy detection has configurable thresholds and measured false-positive/false-negative fixtures. - A redaction failure rejects the entire record rather than storing it unfiltered. ## Depends on - #8