Experiment Report Writing

Core Contract

Treat an experiment report as a decision artifact, not a diary. It must help the reader decide whether to trust the run, what changed relative to baseline, what the evidence says, and what the next smallest action should be.

Never invent results, metrics, hyperparameters, commands, environments, datasets, checkpoints, or qualitative observations. If logs or artifacts are missing, write the report as incomplete and list the exact missing evidence. If the experiment has not been run yet, write an experiment plan / pre-registration report, not a completed experiment report.

The leading conclusion callout is user-owned and has one fixed purpose: record the first question and its answer before the experiment runs. It must state what the minimum experiment is intended to verify and the precommitted success standard. Preserve the user's answer, or help make supplied intent falsifiable and measurable without inventing it. Do not rewrite this callout after seeing results; if the verification target or success standard changes, create a new experiment entry. Do not write final conclusions, decisions, or causal judgments for the user unless explicitly asked to draft them from inspected evidence.

Default language is Chinese. Keep method names, dataset names, model names, config keys, file paths, command lines, metric names, and paper/system names in their original form when needed.

Keep this skill focused on report writing. Durable experiment records are maintained in Notion, not docs/Experiment.md / root Experiment.md (those files are retired and must not be created or updated unless the user explicitly revives them). Use research-doc-workflow plus notion-doc-workflow for delivery; switch to Feishu or Obsidian only when the user explicitly names that destination.

For project-level Related Work, Reproduction, Dataset and Evaluation organization, use ai-research-workflow/references/project-workflow.md. Keep a single run report in the existing project/method location and link it from reproduction/results indexes. Record the actual dataset-derived version/split and evaluation protocol version in the existing Preprocess/Quantity/Appendix sections; do not create a parallel local experiment narrative or expand the compact Settings metadata into a new questionnaire.

When an experiment starts from an observed research phenomenon, follow observation-to-problem.md before proposing method changes. A plan distinguishes diagnosing a cause from validating an improvement and links the relevant literature/existing solutions and inspected evidence. If those checks are missing, keep the next action diagnostic; do not fabricate a mechanism or claim a literature gap to justify the planned module.

Section headings must use short numbered English labels in this fixed vocabulary: 1. Settings, 2. Preprocess, 3. Quantity, 4. Quality, 5. Next, 6. Pause, 7. Appendix. Do not use Chinese heading titles for these sections. Body prose stays Chinese by default. If a project has other existing report conventions (table style, caption style), preserve those; do not copy project-specific paths, run names, or one-off implementation details into this general skill.