version: "1.0.0" name: record-learning-evidence description: Validate and write the active learner's item-level classroom answers, reading or grammar results, homework, dictation evidence, and offline assessment totals through audited local APIs. Use whenever the user supplies scores, correct/wrong results, raw answers, duration, blanks, or calibration-test results, and when an agent must update a Cassian Atlas database without direct SQLite access.

Record Learning Evidence

Write facts through http://127.0.0.1:8788; never edit SQLite directly.

If an answer comes from an image, OCR, or model transcription and has not already passed an explicit teacher confirmation batch, invoke $confirm-learning-evidence instead. Do not bypass that gate with /api/classroom/attempts. This skill remains the direct path for facts the teacher supplied as already-confirmed structured answers or scores.

An extraction batch commit already creates the formal attempts and evaluations. Do not submit those same confirmed items again through /api/classroom/attempts; verify the extraction commit readback instead.

Record

1. Check /api/health and resolve the active student.
2. Create or confirm the learning session with POST /api/sessions.
3. Normalize item-level facts and submit them once to POST /api/classroom/attempts.
4. Use POST /api/assessments only for separately reported totals and controlled-test metadata.
5. Re-read /api/performance/sessions or the affected passage report and verify counts.

Use stable source question, passage, session, and item IDs when they exist. Use a fresh event_id and stable idempotency_key for each logical write; reuse both only for an identical retry.

Protect evidence

• Set answer_capture_status=not_captured and student_answer=null when the raw answer was not saved.
• Do not attach a specific error cause to a not-captured answer.
• Keep everyday lessons as real evidence. Mark offline closed mixed tests and full papers as higher-weight calibration anchors, not the only real scores.
• Store item-level attempts when available; do not replace them with a total score.
• Never treat a provider candidate or an incomplete confirmation batch as an available item-level attempt.
• Send post-write diagnosis to $diagnose-learning-mistakes.

Read write contracts only for the payload being submitted.