Phase 4: Matched controls and blinding workflow

[ProhibitedTV]

Parent roadmap: #3

Objective

Strengthen the existing blinded validation approach so laser captures are compared against matched controls without bias.

Tasks

• Require or strongly warn for matched controls before analysis can be treated as evidence-bearing.
• Implement blind IDs for all media, frames, variants, ROIs, and candidate records.
• Randomize candidate review order using the blind seed.
• Prevent labels like laser and control from appearing in candidate review until unblinding/export.
• Add a post-run unblind step that computes laser-vs-control comparisons.

Acceptance criteria

• Review mode can operate fully blinded.
• Reports include blind seed, unblinding timestamp, and matched-control summary.
• A run without controls is clearly labeled exploratory and cannot be promoted above anomaly.

Notes

The project credibility lives or dies on controls. This is what separates research tooling from pattern-hunting.

[ProhibitedTV]

Post-PR #19/#20 review note:

Sol's additions are broadly on track and moved the project toward a real community-science workflow, but this issue should stay open because the current dashboard review path still appears to expose unblinded fields before an explicit review/unblind step.

Current UI evidence in gui/lab_dashboard.py:

• Candidate table includes a Label column.
• Candidate table rows populate candidate.get("unblinded_label", "").
• Candidate details include Source: and Control: fields.
• Filmstrip table includes Label and Source columns.

Recommended hardening:

• Add a blinded_review=True default state for Review.
• Hide unblinded_label, source path, control type, and source-identifying file paths until the user clicks an explicit Unblind action.
• Keep blind ID, frame index, protocol, detector family, scores, OCR text, ROI/crop, and processed/raw preview visible during blinded review.
• After unblinding, show laser/control labels and matched-control comparison in a clearly marked section.
• Add a regression test or GUI-state test that verifies no source labels leak in blinded mode.

This is important for credibility with both psychonaut community reviewers and science-minded readers: the review phase should ask, “Does this candidate look structured when I do not know where it came from?”