dvampire/autogenesis/experiment-results-planning

Experiment Results Planning

This skill designs the experiment/result layer before final metrics exist. It may generate mock planning data, but never presents mock data as real experimental evidence.

Hard Gate

Before writing Results or Discussion, create:

• plan/experiment-protocol.md
• plan/review/method-experiment-traceability.md
• tables/table-schema.md
• figures/data-manifest.md
• real data files or clearly labeled mock_* files

Experiment Protocol

The protocol must include:

• Dataset and split strategy.
• Baselines and why each is fair.
• Metrics and imbalance handling.
• Main comparison.
• Efficiency evaluation.
• Ablation studies for each claimed module.
• Generalization or robustness checks.
• Explainability evaluation if XAI is a contribution.

Each contribution in Introduction must map to at least one experiment or limitation note.

Recommended Experiment Gates

Use these gates in plan/stage-gates.md for result-heavy papers:

1. Gate D0: Experiment Protocol Locked

• Required: datasets, split rules, Non-IID construction, seeds, baselines, metrics, hardware/software, log schema.

1. Gate D1: Method-Experiment Traceability

• Required: plan/review/method-experiment-traceability.md.
• Map each contribution to method modules, experiments, tables/figures, and allowed claims.

1. Gate D2: Table/Figure Data Contract

• Required: tables/table-schema.md, figures/data-manifest.md, and data files.

1. Gate D3: Main/Efficiency/Ablation/Generalization/XAI Results

• Each result family needs raw logs, aggregation rule, table update, figure script, and prose update.

1. Gate D4: Result Chapter Decontamination

• No planning notes left in the chapter body. For Chinese manuscripts this means no 实验目的 (experiment purpose), 表位 (table slot), 回填模板 (backfill template), or 讨论提示 (discussion prompt) markers.

1. Gate D5: Peer Review Pass

• Required: plan/review/<section>-peer-review.md.

Method-Experiment Traceability

Create:

| Contribution | Method module | Experiment | Table/Figure | Allowed claim | Evidence status |
|---|---|---|---|---|---|

Do not let a contribution survive in Introduction if no experiment, limitation note, or future-work boundary supports it.

Mock Data Boundary

Mock or synthetic values are allowed only for planning figures and table layout.

Rules:

• File names must start with mock_ or synthetic_.
• Every mock table must contain a note: PLANNING DATA - replace before submission.
• Manuscript prose using mock values must keep the placeholder [待真实实验替换] ("replace with real experiment"), which is the marker the quality gate greps for.
• Do not describe mock values as "results show", "verified", or the Chinese equivalent 实验结果表明.

Table Schema

For each table, define:

Do not create a table unless it supports a claim in the manuscript.

Recommended table fields include mean ± std or confidence intervals when repeated runs are expected. Record aggregation rules in tables/table-schema.md.

Figure Handoff

Data figures must go through figures-python:

1. Write or receive CSV/JSON data.
2. Record it in figures/data-manifest.md.
3. Generate figures/<section>/<figure>.py.
4. Export PNG and SVG.
5. Write a caption that states what the figure measures, not what the author hopes it proves.

Model architecture and flow diagrams use figures-diagram prompts instead of synthetic data plotting.

Results Prose Pattern

For real data:

The method achieves X under condition Y, compared with baseline Z. The improvement is mainly associated with [module], while [failure case] remains visible in [metric].

For planning data:

[待真实实验替换] This paragraph will compare Table N after real experiment logs are inserted.

Never leave "experiment purpose", "discussion prompt", or "table position" instructions inside final chapter files.