BS Cost Theory Opt

Use this skill when the task is to move from heuristic cost terms to a validated optimization model that can be promoted only after passing gates.

Step 1: Lock metric contract before fitting

1. Use signed target:
   • delta_tau_ps = tau_back_ps - tau_front_ps
2. Use signed model output:
   • delta_cost = cost_back - cost_front
3. Keep compatibility fields:
   • cost_front, cost_back, min_cost
4. Gate-0 contract checks are mandatory before any optimization claim.

Step 2: Theory-grounded feature setup

Build from physically meaningful inputs first:

1. Geometry:
   • pin-based HPWL and bbox features
2. Front/back timing proxy:
   • PDK timing proxy, via proxy, fanout/load terms
3. Backside overhead:
   • nTSV/stack-via related terms (or best available proxy)

Use shared feature names across signal and clock tracks where possible.

Load:

• references/theory_formulation.md

Step 3: Optimization procedure

1. Compare against HPWL baseline with identical split strategy.
2. Use repeated cross-validation (default 5x2).
3. Report per-fold and per-dataset metrics:
   • Pearson
   • Spearman
   • sign accuracy
4. Use same train/test partitions for baseline and candidate model.

Load:

• references/optimization_and_gates.md

Step 4: Promotion gates (stable > HPWL)

Model can be promoted only if:

1. Gate-0 pass:
   • contract valid
   • sign convention fixed
2. Gate-1 (delay consistency) pass on all target datasets:
   • delta_pearson > 0
   • delta_spearman >= 0
   • delta_sign > 0
   • fold win-rate thresholds satisfied
3. Gate-2 (absolute consistency floor) pass:
   • candidate Pearson/Spearman/sign must exceed configured minimum floors
4. Gate-3 (nTSV capacity behavior) pass:
   • high-pressure bins/regions must receive positive cost uplift
   • if coordinate-level region data is unavailable, use an explicit pressure proxy and mark as proxy mode
5. Stability pass:
   • no single dataset collapses below baseline in key metrics.

If any gate fails, keep shadow mode and iterate model/feature design.

Step 5: Internet-backed benchmark expansion

When moving beyond current datasets:

1. Validate benchmark provenance from primary sources.
2. Record repo links + commit/branch in summary artifact.
3. Include CPU + accelerator classes (not only systolic arrays).

Load:

• references/internet_benchmark_notes.md

Mandatory artifacts

1. *.dataset.tsv (dataset-level scorecard)
2. *.per_fold.tsv (fold-level evidence)
3. *.summary.md with final PASS/FAIL and open risks
4. include gate mode flag: region_exact vs proxy_mode