From source to training point.
Follow two real training examples from full UN statements to the checked bullet plan, the final input and target, and the small model’s actual answers.
What this trial learns: turn a supplied, source-checked plan into a comparative memo. The plans and targets are assistant-authored. The original statements are UN records. Only the final target text is supervised during SFT.
What the sources actually say.
These are the complete bounded national statements in the saved English UN extraction, with the source document and page locations retained. The original meeting PDFs contain additional speakers.
Full original diagnostic input, including earlier model notes
The notes are earlier model outputs, not official statements. This is the saved source task; it is not a recovered teacher API transcript.
The checked content plan.
The session’s assistant read the source statements and authored bullets for each delegation, plus bounded common ground, differences and an analyst proposal. This preserves the content the small model is given.
Full annotated plan JSON and authorship
Quote locations and hashes check provenance. Plan semantics were reviewed by the same assistant; independent review remains pending. No separate teacher endpoint was called or standalone teacher conversation saved.