Skip to content

P0.2.1

P0.2.1

From the healthy-baseline sub-study — finding a genuine healthy SimSiam baseline and a gate that calibrates RankMe vs. this baseline.

  1. The instrument (RankMe) can separate a genuinely-collapsed representation from a genuinely-healthy one, ~2.8×, seed-stable device-portable.
  2. A genuine healthy baseline exists, in the clean long-horizon corner. The fix P0.2 needed was a fair training regime, not a bigger model (same tiny ViT; full STL-10 split, IID, 40 epochs, warmup+cosine LR, predictor width 64→128). Intact SimSiam's RankMe dips then re-expands and holds (~5.5, the healthy U-shape) while the collapse arm sinks to the ~1.9 BatchNorm floor; the discriminator flips from loss to RankMe, which now separates the arms ~2.8× and the gate PASSES (CPU + Gaudi HPU). This licenses reading Phase-1 RankMe drops as degradation.
  3. The baseline does not generalize toward the target regime. Across the 2×2 {IID, class-blocked} × {40 epochs, single pass}, only IID × 40 epochs passes: correlated ordering suppresses the healthy arm (contrast closes from above), and a single pass leaves the collapse control under-developed (from below). Conclusion: for this metric, the 40-epoch IID corner is the fixed calibration reference Phase 1 reads against, and correlation's penalty is a multi-epoch effect (ordering barely matters at single pass).
  4. 40 epochs is the cheapest robust reference and a plateau. A horizon sweep (5→80 epochs, robustness = min across seeds) shows the contrast climbs past 40 (not a local optimum) and is untrustworthy below it — 25–30 straddle the 2× line within seed noise — while 40 is the first horizon whose worst-of-4-seeds (2.47×) clears 2× comfortably. The healthy signal saturates by ~40 epochs; no cheaper reference exists.