# results_rsc

Representation Self-Challenging (RSC) fine-tune of published `stem_ms96`. Init from `models/logit_tf_v2/stem_ms96__held-*.pt`. Same ms features, window 96, 1000 steps, batch 512, AdamW lr 8e-4, cosine, augment, switch weights. With probability `RSC_P=0.5`, mute the top `RSC_DROP=0.33` of post-encoder channels by mean `|dL/dh|` over time, then train on the muted forward. Architecture unchanged. Evaluate EMA 0.995. HMM preset p_stay 0.994 scale 0.8. Also scored causal tau-30 high-pass of the RSC evidence (`hp30`; emissions from corpus only).

Code: `stem_ms96_rsc.py`. Checkpoints: `models/stem_ms96_rsc/`. JSON: `results/research/stem_ms96_rsc.json`. Log: `experiments/stem_ms96_rsc.log`.

## LODO (preset)

| Held-out | RSC H | RSC Acc % | stem_ms96 H/Acc | SWA H/Acc | hp30 H | hp30 Acc % |
|---|---:|---:|---|---|---:|---:|
| AVGC | 0.8042 | 75.12 | 0.7945/74.06 | 0.8015/74.91 | 0.7912 | 71.33 |
| DTU | 0.9059 | 89.70 | 0.9061/89.85 | 0.9003/89.52 | 0.8152 | 69.27 |
| KUL | 0.8717 | 82.15 | 0.8755/82.33 | 0.8769/82.35 | 0.8530 | 75.17 |
| UHD | 0.8500 | 79.79 | 0.8407/79.47 | 0.8553/80.09 | 0.8157 | 73.36 |
| USTC | 0.9190 | 85.80 | 0.9219/86.36 | 0.9161/86.19 | 0.8204 | 69.56 |
| **mean** | **0.8702** | **82.51** | **0.8677/82.42** | **0.8700/82.61** | | |

## External (preset)

| Set | RSC H | RSC Acc % | linear_hp30 | stem_ms96 | SWA | RSC_hp30 H/Acc |
|---|---:|---:|---|---|---|---|
| AASD | 0.6196 | 51.13 | 0.6422/53.12 | 0.6158/51.54 | 0.6202/51.08 | 0.5842/51.03 |
| Carta2026 | 0.5481 | 48.48 | 0.6520/60.54 | 0.5695/51.03 | 0.5507/48.68 | 0.5652/49.60 |

## Targets

| Target | Result |
|---|---|
| AVGC H>=0.85 and Acc>=78 | miss (0.8042/75.12); +0.0027 H / +0.21 Acc vs SWA; still open |
| UHD H>=0.85 | at bar (0.8500 raw 0.849996); below SWA 0.8553 |
| AASD Acc>60 with H>0.63 | miss (0.6196/51.13); Acc flat vs SWA, below linear_hp30 |
| Carta vs linear_hp30 | moved away (0.5481/48.48) |
