biswajit56546/eda-preflight-reflect

EDA Preflight Reflect

Run this skill before submitting new experiment batches.

Step 1: Evidence lock (no guess mode)

Collect latest artifacts first:

1. case-level summary TSV/MD (*.case.tsv, *.summary.md)
2. route evidence (backside_reroute.rpt, routed DEF layer occupancy)
3. timing delta evidence (WNS/TNS/violating paths)
4. internet/primary-source references for current assumptions (routing policy, CTS/PDN layer practice, etc.)

If evidence is missing, mark insufficient_evidence and stop optimization claims.

Step 2: Reflection checklist

Evaluate these checks explicitly:

1. backside_presence: Are there real backside nets in routed geometry?
2. long_net_capture: Is backside mostly capturing long nets above crossover threshold?
3. timing_consistency: Does increased backside usage improve or at least not hurt timing?
4. capacity_pressure: Are nTSV/backside regions overused (pressure likely distorting routing)?
5. policy_vs_model: Is behavior likely route-policy dominated rather than model-driven?

Use references/checklist.md and output pass/fail per item.

Step 3: Produce one preflight reflection artifact

Generate one markdown before any new run:

• root cause ranking (top 3)
• one-line hypothesis per root cause
• one minimal A/B/C plan for next run
• explicit acceptance criteria

Recommended helper script:

python3 scripts/debug/pre_experiment_reflection.py --case-tsv <case.tsv> --out-prefix <out_prefix>

Step 4: Model-gap decision

When short backside nets dominate, do NOT proceed with simple min(front,back) claims. Switch to a judgment model path if either is true:

1. short_share_in_back > 0.6
2. back_on_long_ratio < 0.5 while long nets exist

Judgment model should include at least:

1. timing criticality/slack budget
2. congestion-relief value
3. nTSV region capacity pressure
4. access/via risk

Mandatory outputs

1. preflight_reflection_*.md
2. selected next-step plan (A/B/C)
3. explicit go/no-go for new experiment submission