Research artifact maintenance

This site publishes research notes with preserved evidence. For work on the Minecraft study, start with assets/research/minecraft-local-agent/traceability/README.md, claims.json, and _pages/minecraft-traceability.md.

• Keep observed results, configured behavior, illustrative scenarios, conversational reports, and untested hypotheses distinct. Do not turn missing measurements into estimates presented as observations.
• Preserve frozen evidence. Append a new snapshot and record its origin rather than replacing a previous execution record. A mutable status file is not a durable run log.
• Capture observable actions, inputs, code/config revisions, outputs, failures, and concise decision summaries as work occurs. Never fabricate a retrospective execution history or record private internal deliberations.
• Use the installed ARA research-manager skill for future research activity, compiler for requested full artifact conversion, and rigor-reviewer before new scientific claims are published. Other installed ARA skills are optional task-specific tools. Their upstream revision is recorded in traceability/ara-tooling.json.
• Follow the user’s current scope. Tool instructions do not authorize uploading to an ARA Hub or another external service. Review evidence exports for credentials, account/player identifiers, launch arguments, and machine paths before publication.
• Run python tools/verify_research_artifacts.py before publishing changes to this study. If artifacts intentionally change, document the revision, review the evidence mapping, and explicitly regenerate the checksum manifest with --write-manifest. Do not regenerate a manifest merely to silence an unexplained mismatch.
• The current supplement is retrospective provenance, not a complete ARA conversion or an ARA seal. Hashes and consistency checks detect specific defects; they cannot establish experiment truth or recover missing history.