Choose a distillation target

Match the target to the task

Target	Use it when	Required controls
Response	The student must produce better chat/tool outputs	role/tool validation, rubric judging, generator/judge separation, held-out prompts
Vector	A compatible teacher exposes embeddings and geometry matters	tokenizer/dimension compatibility, normalization, storage/retention approval
Score	The teacher exposes graded pair relevance	calibrated score scale, task mapping, hard-negative review
Ranking	Only relative ordering is reliable	candidate-set provenance, tie policy, group-aware leakage controls

Do not turn a score into a Boolean label unless the threshold and calibration evidence are part of the dataset contract. Do not treat response text as an embedding target.

Provider, rights, and budget gates

Before a teacher call, record its capability, exact model/snapshot, credential environment reference, terms for derived training data, retention/deletion policy, and raw-envelope redaction. Set separate generator and judge budgets. Planning estimates cost; it is not a provider-enforced hard cap.

Every request receives a stable sample identity and resume ledger entry. Preserve successful results and retry only eligible transport failures. A resume must not silently repurchase completed samples.

Leakage and false negatives

Group by shared source, user, document, template, or semantic family before splitting. Lock validation/test before teacher generation. For retrieval, mine hard negatives only from allowed pools, retain miner/version provenance, and exclude likely positives or same-group documents. Re-evaluate negatives when the corpus changes.

Next: run the chat tutorial, run the embedding tutorial, or interpret evaluation.