Research Documentation

1. Log each meaningful research action in research_log.md:
   • date
   • query or source reviewed
   • keep / discard decision
   • brief reason
   • next step
2. Add evidence rows to evidence_table.md for every claim that matters.
3. Keep interpretation and synthesis in analysis/analysis_report.md, using the canonical shared schema.
4. Treat analysis/analysis_report.md as the main background-research artifact that experimentation will later extend.
5. If shared artifacts already exist, update the relevant sections instead of replacing them with a new schema.
6. If the update materially reframes the analysis, snapshot the current three-file analysis bundle under research/analysis/history/YYYY-MM-DD_<slug>/ before editing the live interpretation:
   • analysis_report.md
   • stats_appendix.md
   • figure_catalog.md
   • preserve the frozen files as-is rather than rewriting them into the newest schema
7. End every update with:
   • what was added
   • unresolved gaps
   • what still needs user approval

Evidence table rules

Every important claim should capture:

• claim
• source type
• source identifier or URL/path
• exact quote, table pointer, or precise paraphrase target
• confidence level
• status: verified / conflicting / tentative / rejected
• notes on comparability or limitations

ML/AI-specific requirements

For benchmark or evaluation claims, capture:

• model name and version
• dataset and split
• metric
• evaluation harness or script, if known
• prompting, few-shot, or system setup, if relevant
• hardware or runtime context, if relevant
• whether the source is the paper, official repo, model card, or a third party

Writing rules

• Separate fact from interpretation.
• Mark open questions explicitly.
• Preserve conflicting evidence instead of overwriting it.
• Do not silently delete superseded reasoning; mark it as revised.
• If a prior interpretation is materially superseded, freeze it in research/analysis/history/ before replacing the live analysis bundle.
• analysis_report.md should keep a short Revision Status block and a compact Revision History table once material reframes exist.
• Frozen snapshot bundles are historical records; do not retrofit their contents to the latest live template.
• Prefer concise, source-backed sentences over long unsupported prose.
• Every markdown table in research artifacts must have an immediately preceding numbered caption in the form Table N. ....
• Table captions must be self-contained and explain the table scope and key variables without relying on the section heading alone.
• Before experiments exist, mark experiment sections as pending rather than inventing experiment content.

References and templates

• Use assets/research-log-template.md.
• Use assets/evidence-table-template.md.
• Use assets/analysis-report-template.md.
• Use references/source-hierarchy.md when deciding source quality.