[shipitdev]

Changes

• Finish the fixed 30-question training baseline and retain all failed outcomes.
• Expand context recall to 30 cases, faithfulness to ten and precision to four valid scores.
• Record quota-limited judge expansion, sample sizes, archived attempts and actual usage.
• Preserve original inference code and reserved development questions.

Checks

• Evaluation and publication tests passed with 19 passed and two optional skips.
• Report assertions cover cohort size, outcome totals, usage, redaction and unchanged inference file hashes.
• Saved outcomes published to LangSmith without further model calls.