[DonaldLucy]

Summary

Adds benchmarks/code_behind_layout/, the code and frozen configurations for The Code Behind the Layout: AI Agents for Analog Layout Generators. The study tests whether one shared edit to a gLayout generator's source improves post-extraction performance across a declared family of parameter configurations, without breaking any configuration that passed before. Every candidate goes through SKY130 GDS → DRC → LVS → RC extraction → SPICE at tt/ss/ff.

• E1: action level for frontier-model edits (geometry, primitives, generator source).
• E2: cheap ranking signals vs. full evaluation (area, geometric rule, extracted RC, calibrated F3).
• E3: test rules for family regressions under a fixed budget (anchor, random, boundary, risk).
• E4: point vs. family supervision, SFT vs. DPO, for a local Qwen2.5-Coder-32B model (T1–T4 × 3 seeds).
• E5: physical CPU, GPU and token cost.

The policies are Claude Opus 5.5 and GPT-6.1.

Data

The frozen task packets and the E4 training set (43 MB compressed) are published as a release asset rather than committed:

cd benchmarks/code_behind_layout
bash scripts/fetch_data.sh

Dependencies

• The E1–E3 circuits use gLayout at 5af536b with reproducibility/glayout_fixes.patch applied. The patch covers FVF/OTA routing and reference fixes, a SKY130 grule fix, netlist subcircuit naming and PEX top-cell loading.
• The E4 textbook collections use glayout.sharing and benchmarks/textbook_sharing from Explicit diffusion and gate sharing in gLayout, with a three-PDK textbook benchmark #127.

Testing

• python3 -m unittest discover -s tests -p 'test_*.py': 625 tests pass (5 skipped).
• With the data bundle unpacked, scripts/train_policy.py --validate-only admits T1–T4.
• The E1 and E3 frozen manifests verify against the bound implementation files.
• Nothing outside benchmarks/code_behind_layout/ changes, and the existing CI collects only the top-level tests/.

[DonaldLucy]

Closing: this work will live as a separate project branch rather than a merge into main.