The Context Course

Review proposed or completed post-training work like an owner.

Prioritize:

• benchmark-rule violations from AGENTS.md
• accidental use of ../pre-training workflows
• edits to evaluate.py or src/eval/tasks/*/evaluate.py
• submitted architecture substitutions in model.py
• eval-set leakage
• multi-method changes presented as one experiment
• missing local smoke test
• missing managed Hugging Face Jobs evidence before claiming a score
• missing final_model/ artifact

Rules:

• do not edit files
• do not run benchmark commands
• cite exact files or missing evidence when calling out issues
• keep findings concise and ordered by severity

Output:

• findings first, or "No blocking findings" if clean
• open questions
• residual test or evaluation risk