table2_set/TABLE2_baselines.md

Table 2 — Commercial & Open-Source Baselines (ASAP7 reproduction)

DRC VLM testing set: 100 samples, 10 rules, k=1 trial (single-violation crops). TP/FP/FN/TN are normalized rule-level rates (%); F1 ± = (bootstrap 95% CI half-width). Per-rule-category F1: EN=Enclosure, Grp=Group, Spc=Spacing, G-Spc=Group-cond. Spacing, W=Width. Grp / G-Spc are N/A — ASAP7 has no group-via rules (those are the paper's sub-2nm PDK).

Model	TP↑	FP↓	FN↓	TN↑	Recall	Prec	F1	VIA Loc	EN	Grp	Spc	G-Spc	W
Qwen3-VL-8B-Instruct (open)	3.6	30.7	6.4	59.3	36.0	10.5	16.3 (±4.0)	57.5	17%	N/A	0%	N/A	22%
GPT-5.2	1.3	10.9	8.7	79.1	13.0	10.7	11.7 (±5.8)	88.9	12%	N/A	0%	N/A	0%
Gemini-3.1-Pro	1.5	10.8	8.5	79.2	15.0	12.2	13.5 (±6.3)	66.7	15%	N/A	0%	N/A	0%

Paper Table 2 baselines (industrial sub-2nm PDK, k=10) for reference: Qwen3-VL-8B-Base 25.4%, GPT-5.2 30.7%, Gemini 3 Pro 37.2% F1.

Key observations (match the paper's qualitative findings)

• Spacing F1 = 0% for ALL baselines — none detect M1 spacing violations; they fixate on the red via and guess enclosure/width rules. This is exactly the gap the paper says fine-tuning closes (Spc 0%→57% in their Mixed model).
• Width: only Qwen3-VL detects any (22%); commercial models 0%.
• Enclosure: weakly detected by all (12–17%).
• Precision/recall split: Qwen3-VL over-predicts (R=36, P=10.5 → high FP); GPT-5.2/Gemini are conservative (low recall, slightly higher precision).
• VIA Loc: GPT-5.2 best (88.9) — reads µm via centers accurately; Qwen3-VL lower (57.5) partly because it output via coords in nm not µm on some samples (unit confusion).

Caveats vs paper

• k=1 (paper k=10): single trial, noisier, no trial-averaging. Rerun at k=10 to tighten.
• Absolute F1 lower than paper baselines: our eval is stricter (1 GT rule vs 10 candidates, so over-prediction is heavily penalized), different PDK/rules/prompt, ASAP7 not sub-2nm.
• The relative pattern (baselines fail spacing/width, enclosure weak) reproduces faithfully.