FaceValue: degradation battery, analysis code and derived measurements for a video engagement pipeline audit

Code and derived measurements for an audit of a video-based engagement analysis pipeline under controlled recording conditions. Contents. The video-degradation battery with its exact ffmpeg parameters and a manifest of every command; the two-detector plus affect-model pipeline used for scoring; the analysis code that computes effect sizes, noise floors, permutation tests and bootstrap intervals; the preprocessing control; the repeat-run control; the decision-level arm; the projection analysis; a second affect model (EmoNet); PSNR and SSIM for every degraded frame; the silent conditions repeated on all 3,443 clips; participant-level intervals for every ratio; and 29 derived measurement tables with their analysis logs. PREREGISTRATION.md holds the decision rules and a dated change log. Despite that file name, the study was not preregistered with any public registry: the decision rules, thresholds and exclusion criteria were fixed before the data were analysed and archived here, and that is what the file records. Version 2.0.0. Adds the scripts behind several numbers reported in the paper that version 1.0.0 lacked (participant-level ratio intervals, the permutation test of the reference effect, dominant-expression agreement and the classifier discrimination check), and the arms registered on Day 5 of PREREGISTRATION.md: EmoNet as a second model family, image-quality measures, and the full-set replication. The projection script is the corrected version and projection.csv is regenerated with it; every other table from version 1.0.0 is byte-identical. EmoNet weights are not redistributed. Source data. Measurements are derived from the DAiSEE dataset, which is obtained separately from its creators. No video and no DAiSEE label columns are redistributed here. DAiSEE permits research use and forbids sharing with third parties, so every table retains only clip identifiers; a holder of DAiSEE can join the labels back and reproduce every reported value. Before you re-use this. README.md records four known issues: a set of condition keys is misnamed in the code relative to what the recipe actually does; the effect measure is bounded below by zero and therefore biased upward, so a calibration on zero-effect synthetic data is provided; one trained component failed its pre-specified adequacy gate and is included only because the failure is reported; and one archived clip fails its checksum. Access. The files are open: they can be read and downloaded without a repository account, without registration, and without any request to the depositors.