Analog Design Bench V2.1

Benchmark versions

2.1 (2026-10-23)

This is a separately identified correction release of the same 50 V2 tasks, based on commit 972f7cd7c95381134387b250edf18f3bd23bdd25.

• Correct the continuous-library MIM example to mult=1; distinguish the traditional PDK's mf interface without claiming equivalent electrical multiplicity. Correct all 18 previously bundled guide copies.
• Synchronize one public guide into every candidate's /app and /opt locations. All 50 candidate contexts are updated to prevent inherited stale documentation. The simulator, checker and PDK base-image digest stay fixed.
• Expose the exact interface/geometry preflight and its existing bounds for unregulated charge pump, high-PSRR bandgap and 8-to-1 analog mux.
• Make bootstrap's inclusive 100 nF limit independent of decimal spelling, using exact decimal comparison, while rejecting values above the limit. Publish its existing MOS/capacitor-only whitelist and identical public gate.
• Add boundary, fail-closed, public/grading parity and context-sync regressions.

The issue review identified 19 task directories (four substantive contract or boundary fixes, plus 15 other guide copies), with overlap removed. Propagating the corrected shared runtime guide necessarily updates all 50 build contexts. Every task receives a new revision and benchmark_release binding; task IDs, electrical targets, reference circuits and simulation benches are preserved. Tasks without an explicit V2 revision are treated as baseline revision 1.

V2 results remain V2 results. Do not relabel existing scores, overwrite their image bindings, or interpret a documentation change as a new model evaluation. The inherited results/reference.json files are historical reference results, not a claim of fresh 50-task electrical acceptance of V2.1.

Validation commands (Linux, Python 3.11+, or Python 3.10 with tomli):

python3 tools/sync_public_assets.py --check
python3 tools/test_v21_contract.py
python3 -m unittest discover -s tasks/sky130-bootstrap-driver-pt/tests -p test_check_area.py

Run tools/test_v21_contract.py inside the pinned base image as well to execute the real checker and ngspice tests. A host-only run skips those runtime tests. No model run is needed for these regressions. Static eligibility checks do not assert that a circuit meets the task's electrical targets.