| name | backside-benefit-attribution-evaluator |
| description | Evaluate whether backside benefits come from `CTS-backside-only` or from moving a selected subset of signal nets to backside, under a fair and physically valid comparison contract. |
Backside Benefit Attribution Evaluator
Use this skill when the question is "what caused the benefit?" not merely "which arm won?"
When to use
Use this skill when:
front-only, CTS-backside-only, and partial-signal-backside need fair comparison,
- a run shows a PPA difference but attribution is unclear,
- backside usage must be evidenced directly rather than inferred from PPA alone,
- the paper narrative needs a controlled mechanism claim.
Scope boundary
This skill owns:
- benefit attribution across the canonical three-arm ladder,
- fairness checking for the comparison contract,
- deciding whether an observed delta is likely meaningful, weak, confounded, or unproven.
It does not own:
- the physical contract itself without
bspdn-physical-contract-auditor,
- the PDN sufficiency question without
bspdn-pdn-sufficiency-evaluator,
- full batch orchestration.
Canonical arms
front-only
CTS-backside-only
partial-signal-backside
Required evidence
For any claim, require:
- fair route-window contract
- direct backside usage evidence
- comparable stage/restart baseline
- PPA deltas with stated axis and direction
Expected outputs
*.results.tsv
*.conclusion.md
- optional
*.experience_delta.md
Hard rules
- Do not claim "backside signal helped" if the run has no direct backside occupancy evidence.
- Do not compare arms with mismatched routing contracts and then call the result mechanism evidence.
- Keep "no observable benefit" separate from "backside signal path disproved"; the latter requires stronger evidence.
- Promotion toward placement-algorithm claims requires repeated mechanism-consistent Arm C advantage, not one-off wins.
Operational references
- Load
references/background-knowledge-links.md for the current summarized attribution contract and its KB anchors.
- Load
references/update-mechanism.md when deciding whether the attribution criteria or baseline assumptions should be refreshed.