The outcome is the one both verification rounds agreed with. After the second round the author proposed partially_reproduced, from a penalty selection added then; that change has not been verified (protocol pilot-4, "Hand over for verification"). With tenfold contiguous cross-validation and the ridge penalty chosen per cell by fivefold cross-validation within each training set, the mean cvR2 over 1,335 cells is -0.041 from female behaviours, -0.005 from male behaviours and -0.004 from both (mouse means -0.026, 0.014 and 0.016), against 0.08, 0.14 and 0.15. The ordering holds, both against male by a margin of 0.0003 over cells and 0.0014 over mouse means, and all three are below 0.25, so the partial rule is met by a thin margin. With the penalty chosen on the last fifth of each training set, the design used in the first two rounds, the means are -0.077, -0.013 and -0.057 (mouse means -0.050, 0.008 and -0.017) and both scores below male, so the ordering failure that gave not_reproduced came from that selection. A model with only an intercept scores -0.086 under the same folds, so the negative values come from the fold design; with the cross-validated penalty the gain over it is 0.045, 0.081 and 0.082 (mouse means 0.054, 0.095 and 0.096), below the paper's values. With the last-fifth penalty, R2 within single folds is dominated by the same offsets (medians -0.61 to -0.71, intercept -0.70). With frames assigned to folds at random, a variant not named in the criterion, the means are 0.229, 0.207 and 0.322, with neighbouring frames in both training and test sets; the value from both behaviours is then above 0.25. The test is weak, since no design here separates behaviour from the slow drift of the traces.
Proposed after the last verification, not yet verified: partially reproduced. With the ridge penalty chosen by fivefold cross-validation within each training set, a selection added after the second round, the ordering of the criterion holds (both above male by 0.0003 over cells) and all three values are below 0.25, which meets the partial rule.
How the test differed from the paper’s method
- Lags in 1-s steps over 0 to 10 s, not every frame, to keep the run short.
- Contiguous folds, the ridge penalty chosen per cell by fivefold contiguous cross-validation within each training set, and R2 on pooled held-out predictions; the Methods state none of these. The cross-validated penalty selection was added after the second round of verification; the outcome would change with it, and that change is proposed, not verified.
- Fit over the interaction (male entry to end of annotation) on z-scored, unsmoothed traces.