Read README.md, labs/README.md, and the nearest lab/experiment entrypoint before changing records. This repository holds research collections; reusable methods live in ai-engineering.
- Use laboratory-recordkeeping for organization and preservation, and scientific-methods for inquiry design and interpretation. Read the actual linked guides relevant to the work. CONVENTIONS.md records the adopted source baseline; preserve experiment-specific pinned methods.
- Use the documented lab/experiment/run identity scopes. Inspect directories and indexes before allocating IDs. Create only records needed by the work.
- Preserve dated intellectual history and original captures. Record failures, excluded attempts, retries, not-run slots, and evidence gaps. Never invent execution evidence or replace an earlier run with a successful retry.
- Distinguish raw observations, transformed results, interpretation, and decisions. Link claims to their evidence. Self-checks are not independent verification and a documented recipe is not an executed experiment.
- When instructions require linked files, open them. If access fails, report the affected gap before dependent work. File-read evaluations require actual tool calls, returned content, and behavior dependent on that content.
- Follow the unexpected-file decision guide. Unknown existing files stay in place pending disposition; incoming unclassified material needs an intake receipt. Preserve source-to-destination mappings for moves.
- Keep native computational sources, relevant execution state, and useful readable exports. Record which is authoritative and how exports were made.
- Inventory externally held evidence. Do not assume Git includes large-file storage or local-only data. Avoid blanket ignore rules for research artifacts.
- Main is the default branch for research records. Follow explicit branch and commit instructions; preserve unrelated work. Engineering plans belong in the relevant engineering project's planning worktree.
- Assistant-authored commits include exactly one of each trailer:
Co-authored-by: Codex <noreply@openai.com>andCo-authored-by: Billo AI <ai-engineering@billo.systems>.
Before handing off changes, check relative links, readable indexes, stable identities, provenance, and the reported evidence state. Use domain tools to check executable material when present. Do not claim scientific validation from Markdown or repository-structure checks.