P0.2

P0.2

From the positive-control sub-study — Initial investigation into the "collapse" instruments only; the forgetting mode (MAE / probe-on-past) needs its own positive control, later. In one session, predictorless / stop-gradient-off SimSiam (the anti_collapse=false toggle) is run side by side with intact SimSiam, both from-scratch on class-blocked STL-10. The PC drives its loss to the −1 constant-solution floor while RankMe craters to ~32% of its random-init value, but values for the control aren't clearly better. Loss seems more of a collapse discriminator than RankMe at this complexity - the latter doesn't cleanly read collapse. Conclusion: we can't tell whether the healthy arm's RankMe decay is (a) SimSiam's general single-pass dynamics — in which case it's a real property that won't vanish with scale and the baseline is fundamentally compromised — or (b) an artifact of the toy setup (tiny model, 30 steps, single pass, no schedule) — in which case it disappears as we scale and the baseline is fine. This motivates P0.2.1.