PDKG: ChipMind-inspired physical design knowledge graph (accuracy-first)

[sethupathib]

Summary

Adds pd-knowledge-graph/ — a ChipMind-inspired (arXiv:2512.05371) knowledge graph for physical design methodology, built with an accuracy-first policy.

Faithful mapping from ChipMind

ChipMind	PDKG
Circuit Semantic Anchors (CSA)	PD types: FlowStage, FailureMode, Constraint, …
Hierarchical (T_B/T_A/T_L/T_N)	backbone / auxiliary / linking / normalization
Adaptive Top-K + MIG	BoW cosine MIG proxy (offline, reproducible)
SpecEval atomic facts	queries/pd_eval_qa.yaml + offline Atomic-F1 proxy

Accuracy policy

• Closed ontology (ontology/pdkg_schema.yaml) — unknown relations fail validation
• 62 gold triples with confidence + provenance (kg/gold_triples.yaml)
• Auto-extract tagged source=auto (never mixed into gold claims)
• Do not cite ChipMind’s SpecEval F1 0.95 / +34.59% as this repo’s score

Measured (this build)

• Graph: 76 nodes / 73 edges (62 gold + 11 auto)
• Offline Atomic-F1 proxy: 1.0; entity recall 1.0; multi-hop density→setup OK
• CLAIM_GATE: CITEABLE=yes_pdkg_gold_eval

Reproduce

cd pd-knowledge-graph
pip install pyyaml
./examples/compare_all.sh
cat results/SUMMARY.txt

ChipMind mapping
Causal chain
pdkg_construction.gif

Test plan

• PYTHONPATH=src python3 tests/test_pdkg.py
• ./examples/compare_all.sh
• Figures / GIF / PDF built
• Honest citation of ChipMind vs local metrics

_{To show artifacts inline, enable in settings.}