Can the reader diagnose an unfamiliar recipe?

OPD-XFER-A — Reverse loss, fixed prefix source

A learner uses reverse KL on a fixed teacher-generated xA prefix. Is the divergence reverse KL, and is the state sampling on-policy? Explain both axes separately.

Inspect the reasoning

Local axis: yes. The student distribution is the first KL argument and weights the log-ratio.

Source axis: no for this supplied occurrence. The prefix came from a fixed teacher-generated trajectory, not the current student. Another student draw could reach the same text; provenance belongs to the collection event.

OPD-XFER-B — Student prefix source, forward loss

A learner samples B from the student, queries the teacher at xB, and computes forward KL. Is this on-policy distillation? Explain why divergence direction does not change the state source.

Inspect the reasoning

Source axis: yes, provided the response was freshly sampled from the student behavior policy used for collection. xB is a student-visited prefix.

Local axis: forward KL compares teacher to student at that fixed prefix. Swapping the KL order changes the local value and derivative—not the already established provenance.

Carry a five-question rule

What samples the response?
Which prefix occurrence is evaluated?
What teacher feedback is visible?
What objective and reduction are formed?
Which dependence reaches the gradient?

If those answers are explicit, “on-policy” becomes one coordinate again instead of a bag of assumptions.

OPD-HD-EXP-002 · unfamiliar recipe audit

Use the evidence fields, not the method label

Invented transfer record: behavior policy student-v0 sampled B and visited xB; the recipe now proposes a reverse-KL update to student-v1 using full-vocabulary teacher feedback.

This board supplements the unchanged OPD-XFER-A/B answers. It records no score, completion state, or mastery claim.

If the local enhancement does not initialize, the complete worked evidence table remains visible.

Evidence roles and the cautious worked resolution
record	role	why
xB was generated by behavior policy student-v0.	Decisive	It identifies the source and behavior-policy version.
The loss and proposed update target student-v1.	Decisive	It shows that the behavior trace is stale relative to the update policy.
The local objective is reverse KL.	Separate axis	Divergence direction does not establish trajectory freshness.
The teacher returns a full-vocabulary distribution at xB.	Separate axis	Feedback granularity does not identify the behavior policy.

Resolution. The prefix was student-generated by student-v0 and is therefore stale/off-policy relative to student-v1. That provenance does not change. A declared policy-lag allowance or importance-correction rule may govern whether and how the stale trace can be used for the student-v1 update; neither makes it current-policy data. Reverse KL and full-vocabulary feedback remain separate fields.

The field-guide version

The source chooses the visited states. The local divergence compares distributions at a visited state. The gradient estimator decides which parameter dependence survives.

This guide explains a canonical detached GKD-style toy and several boundaries. It does not define every method called On-Policy Distillation, report a real-model benchmark, or establish a measured learner outcome.