[Mattral]

Summary

Run 2's pre-registration: docs/v1_run2_preregistration.md. Written and committed before any run-2 recording exists. Once merged, its commit is the checkpoint; any departure during the run is reported as a deviation, never edited in.

Design (hybrid):

• Primary, confirmatory: 24 software-degraded conditions. Fresh clean recordings degraded frame by frame: darken, downscale, blur, noise × 6 severities each. The severities were chosen by a pilot on run 1's footage that computed no gap scores, and each stops just short of its dropout cliff. All are labelled as simulated degradations of real footage.
• Secondary, descriptive only: 6 physical conditions (card occlusion at about 25/50/75%, distance 2/3/4 m). No decision rule applies to them.
• Data: 3 clean takes of 75 s (not shortened: 138 windows per condition), one subject, fresh recordings only.

Ground truth and baseline:

• Primary: paired landmark error, the distance from the same frame's clean detection, in hand sizes. This is a documented spec 8.1 deviation: with dropout as the ground truth, the trivial dropout-counting baseline is the ground truth (ρ = 1). Dropout is secondary.
• Frames where the degraded condition lost the hand are excluded from landmark error, so it measures quality given detection, with availability carried by dropout. This is fixed in writing.

Confirmatory rules (the only two):

• A: the 95% CI of ρ(Fréchet gap, landmark error) has a lower bound above 0.
• B: the gap score beats the no_hand_fraction baseline. The 95% CI of Δρ, from a paired bootstrap, has a lower bound above 0.
• Bootstrap: percentile, 10,000 resamples over conditions, seed 0.
• The pilot's baseline ρ was +0.56. The document says up front that passing B with 24 conditions likely needs a much higher gap ρ, so a near-miss can't be read as success later.

Fixed in advance:

• summary_dim = 16, so 138 windows give "medium" confidence rather than "low". The non-independence caveat stays.
• The shoulder-midpoint anchor.
• Conditions are never dropped. One that overshoots its dropout is reported anyway, and an undefined landmark error is ranked worst.
• The collapse rule, recording-only exclusions, and an interpretation table for each outcome.
• Secondary checks:
  • the gap computed on present-frames-only windows (does the result survive when the model can't see missing hands?);
  • a presence-density baseline;
  • dropout as ground truth;
  • MMD²;
  • per-factor agreement;
  • ρ ≥ 0.6 as a practical note only.

Known limitations are stated in advance, including the differing scales across factors and the survivor bias in landmark error.

Next, after merge: the run-2 analysis script with caching, committed before any processing; then recordings and an overnight run.

Docs only (plus ROADMAP and CHANGELOG). pytest: 175 passed; ruff: clean; privacy greps clean.

🤖 Generated with Claude Code