SkinDeepRESEARCHSteve Seguin

Exploratory results

Check Qwen before continuing

Several chances to answer within one model pass.

Back to decisions without text

Qwen processes the input once. At layers 6, 12 and 18, a small trained classifier checks whether to answer. If it continues, the lower layers are not repeated.

Locked stopping ruleCorrect / 100Layer blocks skipped
Original full classifier860%
Jointly trained checkpoint classifier864.75%
Checkpoint classifier with distillation876.75%

The distilled rule answered early on 16 messages and corrected one false block. It introduced no new errors on these 100 previously inspected messages. A one-message improvement is not enough to establish better accuracy.

A separate actual-execution pass over the first 50 messages took 31.17 s at full depth, 28.33 s with joint classification and 28.67 s with distillation. One paired pass is noisy; it does not reliably rank the two gates.

What learns, and when does it stop?

The Qwen backbone stays frozen. One shared 896-to-64-to-2 readout learns across checkpoints, with depth-specific normalization and biases. The distilled version also learns from the original full-depth classifier on training examples. The final fallback remains the original classifier.

Historical development data selects confidence thresholds and agreement checks. Both selected gates require consecutive checkpoint agreement, so neither stops at layer 6. These empirical rules have not passed fresh reliability acceptance.

Training, all fixed-depth comparisons, timing and failed int8 repair · Full quality record

Why the INT8 conversion failed: a closer test