baseline-ablation-designer

Baseline Ablation Designer

Role

本 skill 负责把 claim-evidence map 细化成可实施 baseline、ablation、metric 和 stop condition。它服务论文策略：证据要足以支撑当前论文级别，也要知道怎样升级或降级。

Method-call

/baseline-ablation-designer(
  claim_evidence_map,
  target_level?,
  project_resources?,
  output_root = knowledge/experiment-plan/
) -> baseline_ablation_plan

Step 1: Baseline Two-Layer Design

每个主实验分两层 baseline：

Layer	Use
minimum-delivery baseline	本科/硕士/课程/普通交付必须能比较的最低对照。
publication-grade strong baseline	中文核心、SCI 或会议投稿需要的强对照，优先 SOTA、recent strong baseline、领域经典方法。

如果资源不足以跑强 baseline，明确写：

strong_baseline_status: missing|resource-blocked|verifier_todo|planned
paper_level_impact: <which level this prevents>

Step 2: Ablation Design

每个创新点至少对应一个“关掉/替换/降级”实验：

• remove component
• replace with simpler component
• disable optimization
• use original baseline module
• reduce data / scenario / resource condition

消融不能只是凑表；必须说明它证明哪个 mechanism 或 contribution。

Step 3: Metrics and Thresholds

指标按论文类型选择：

• AI/ML: accuracy, F1, AUC, RMSE, latency, robustness, cost, variance.
• Engineering/optimization: objective value, gap, convergence, time, constraint violation, stability, energy, tracking error.
• Software/system: response time, success rate, storage size, user flow completion, screenshot evidence, module coverage.
• Generic thesis: correctness, completeness, contrast, reproducibility, scenario coverage.

每个指标必须写：

• higher/lower is better
• minimum acceptable condition
• publication-level desired condition when known
• what failure means for the claim

Step 4: Stop Conditions

防止无限调参。每个 high-cost experiment 至少有一个 stop condition：

stop_when:
  - repeated_attempts_reach_limit
  - metric_plateau
  - baseline_not_reproduced
  - resource_not_available
  - claim_no_longer_supported
on_stop: salvage|ask_user|downgrade_level|resource_acquisition

Write machine rules to:

knowledge/experiment-plan/machine/stop-conditions.yaml

Step 5: Output Files

Write or update:

knowledge/experiment-plan/baseline-ablation-plan.md
knowledge/experiment-plan/machine/stop-conditions.yaml

本 skill 的 deletion-spec

• 触发删除条件：当 baseline, ablation, metrics, and stop conditions are fully generated by downstream req execution tasks, this planning skill can be removed.
• 禁用方式：删除 plugins/experiment-planning-suite/skills/baseline-ablation-designer/，bump plugin 版本，刷新 marketplace/cache。
• 卸载清单：同步检查 orchestrator、evidence-chain-planner、feasibility-salvage-gate、README 中的 baseline/ablation references。