Training versus serving detector — run 064

Trained tensors of run 064 (results): Qwen3-8B (revision b968826d) that answers "Is gradient recording enabled for this forward pass?" with Yes in a Hugging Face Trainer training step and No when vLLM serves the same prompt.

engine-detector-064-run3-tail.safetensors: the 17 trained BF16 tensors of run 3 (layer 34's o_proj, post_attention_layernorm and MLP; all of layer 35; the final norm). Every other tensor is the original Qwen3-8B. Install them into a checkpoint with results/engine-detector-064/source/export_tail.py --base QWEN3_8B_SNAPSHOT --tail engine-detector-064-run3-tail.safetensors --out MODEL_DIR.

Size 721455488 bytes, SHA-256 e33eb88a4dd65a90d77c21d2873b6947fba55ee7eb2f0e5851de533dc1aadbb6.

Run 3 starts its reply to any ordinary request with Yes or No (see the results README).