AIHI

AIHI Learning Evidence Engine — Teacher Review Prototype Teachers need to see what AI changed before they use a lesson. This hackathon prototype preserves the original lesson, creates a separate proposal, shows the change, lets the teacher edit it, and exports only the exact version the teacher reviewed. Any further edit revokes approval. The download contains the lesson body with a UTF-8 BOM, without model notes or internal metadata. Built during Oct 31–Nov 4: the working document/revision/approval flow, Korean lesson editing goals, change selection, comparison and learner preview, selected-passage editing, a bounded Meta Model API integration, durable attempt/spend accounting, and the current interface. The education concept predates the event; we do not claim it as new hackathon work. Meta muse-spark-1.3 generates proposals only from six server-owned synthetic lessons and two editing goals. Arbitrary pasted or TXT material uses manual editing or an explicitly labelled local rule simulation; it is not sent to Meta. Offline tests use mocked provider responses. Current production ledgers contain 13 completed pilot calls plus one completed initial trial; these establish integration, not classroom efficacy. A separate WHYWHY local prototype records first thought, a learner's question, a verification plan, revision and self-explanation, with teacher review/export/deletion. It has no generative-model connection and is not integrated with the teacher workbench. Verified now: 114 workbench regression tests and 22 WHYWHY tests pass. The 2m55s video shows actual v12 local editing and separately labelled saved Meta evidence; it is not a recording of a new model call. The private synthetic pilot expires at Nov 5, 11:00 KST. General student deployment, school authentication/consent, learning gains and teacher time savings remain unverified.