rotating-equip-sft-n2620

Rotating Equipment Evidence-Gated QLoRA

Qwen3-1.7B QLoRA specialized for one falsifiable rule: diagnose only when two independent evidence channels agree; otherwise WATCH without naming a mode. Responses must identify the asset, cite independent evidence, and give exactly one non-compound action.

Selected Final N=2620 result

• Full-Spec adherence / robustness: 1.000 / 1.000 (30/30)
• Exact GATE line: 30/30
• Thinking dumps: 0/30
• Misses: none
• Adapter: lecporr/rotating-equip-sft-n2620
• Adapter revision: 759927dd10db565a2407640faab2830f365aa729
• Training evidence: results/training/n2620/

Parent strict N=2580 at lecporr/rotating-equip-sft @ 784a90897112cd94effb2da39f2fc151f55468ae scored 0.933 (misses e11, e14). Nested N=2820 scored 0.867. The selected mix added 40 watch rows, not a learning-rate change.

Prompt-Ceiling (same Spec, original 180 outputs): best Opus 0.267 rule / 0.633 cross-judge, below 0.95.

Exact inference

The adapter was trained against the listed bitsandbytes 4-bit base. Exact adapter evaluation requires CUDA. Applying the LoRA weights to upstream full-precision Qwen on CPU/MPS changes outputs and is non-comparable. The 16-bit merge scored 0.800 on Mac/MPS and is not the submission number.

Replay on any machine:

python eval.py \
  --model lecporr/rotating-equip-sft-n2620 \
  --revision 759927dd10db565a2407640faab2830f365aa729 \
  --eval-set data/eval.jsonl \
  --predictions results/training/n2620/predictions.jsonl

CUDA adapter generation (no --predictions):

python eval.py \
  --model lecporr/rotating-equip-sft-n2620 \
  --revision 759927dd10db565a2407640faab2830f365aa729 \
  --eval-set data/eval.jsonl \
  --device cuda

Limitations

Synthetic snapshots only. This model is not a physical diagnostic authority or safety controller. The 30-row public eval is development evidence; staff-held-out evaluation remains required.