research: blinded human evaluation panel — does score-low text actually read human?

[devswha]

Why it matters

Every signal in patina is auto-scored. There is no human-rated checkpoint anywhere. A score of 25 might mean "humans agree this reads human" — or "LLM-judge thinks LLM-rewritten text reads human while humans still detect it". The "score must match human intuition" north-star is unfalsified.

Acceptance criteria

• docs/research/human-eval-panel.md design doc: panel size, blinding protocol, fixture pool, analysis plan
• Pilot panel: 30 paragraphs × 5 raters
• Report inter-rater agreement (Krippendorff's alpha) and correlation patina score vs human rating

[devswha]

Wave 4 triage: this stays open.

Concrete blocker: blinded human evaluation requires external panel logistics and consent/provenance handling. Next viable PR should land the study design plus a pilot, e.g. 30 paragraphs rated by five raters, with inter-rater agreement and correlation against patina scores.

[devswha]

Added the blinded human-eval study design in c7db18c.

Evidence added:

• design doc: docs/research/human-eval-panel.md
• wave tracker updated: docs/ISSUE-WAVES.md
• CI: https://github.com/devswha/patina/actions/runs/26232569366

This issue should stay open. The actual panel still needs 30 paired samples × 5 raters, blinded ordering, consent/redistribution handling, and aggregate reporting.