biswajit56546/eda-retro

EDA Retro

Overview

Use this skill immediately after an experiment batch finishes. It converts raw artifacts into a structured "what happened / why / what next" decision and gates whether a recursive next run is justified.

Inputs

Minimum required:

1. run summary markdown (*.summary.md)
2. monitor/history (*.monitor.md, *.monitor.history.tsv) or manifest (*.tsv)
3. key report sources (postCTS/postRoute summary, DRV/type/length distributions, gate outputs)

Optional but preferred:

1. knowledge-base references used for this batch
2. web/primary-source references for unstable assumptions

Workflow

1. Consolidate observations

• extract the smallest set of outcome metrics:
  • timing (WNS/TNS/ViolPaths)
  • PPA (power/area/runtime)
  • geometry/routing (backside usage, type/length distribution, long-net capture)
  • gate status (H1/H2/H3 or equivalent)

2. Diagnose mechanism, not symptoms

• assign one primary mechanism and optional secondary mechanisms:
  • model_mismatch
  • flow_policy_mismatch
  • resource_capacity_conflict
  • tool_stability_issue
  • insufficient_evidence

3. Score improvement proposals

• for each candidate change, score:
  • confidence (low/medium/high)
  • impact (low/medium/high)
  • cost (low/medium/high)
  • risk (low/medium/high)

4. Decide recursion

• trigger recursive eda-loop only when:
  • confidence >= medium
  • expected impact >= medium
  • risk is acceptable
  • evidence gap is not blocking
• otherwise, request focused data collection first.

5. Emit retrospective artifact

• write one markdown summary under slurm_logs/04_delay_modeling/:
  • <run_tag>.retro.md
• include:
  • hypothesis validation state
  • chosen next step
  • explicit stop condition for the next iteration

References

Load as needed:

1. references/retro-checklist.md
2. references/recursive-loop-policy.md