# Stable-teacher descriptive reporting plan

Prepared 2026-10-13 03:53 UTC, after protocol freeze/publication and launch, while
Gchunk0 waits for a GPU. No trained/native outcomes observed.

Run the frozen full author audit only after all4 endpoints. The unbound reporting
scripts cannot modify inputs, grades, primary contrasts or gate. They can be fixed
for packaging errors, retaining failures and keeping scientific inputs immutable.
Report source NLLbefore/after, eligibility/donor/receiver exposure, dense token
counts, clipping, zero-signal batches and mean JSD per actually treated token.
These are descriptive, not new tests. G has80scheduledbatches, F/E64; always show
appropriate denominators. G's eligible receiver counts indicate potential
eligibility only: it receives no dense supervision. Cross-arm JSD comparisons use
different prefixes and supports and do not rank teacher quality. Source likelihood
is not measured acquisition or causal drift. Combined preclip norms do not identify
component gradients or actual parameter-displacement contributions.

Verify exact teacher digests against the frozen initial digest in everyF/Echunk.
The teacher still encounters changing prefixes/donors and its vocabulary partition
is chosen by the student. Report logical generation, support, teacher, checkpointed
backward and retention work, measured released allocation times, and queue waits
separately. GPU readiness57.0s and freshWarm460.1s are additional. Any interrupted
work remains charged; derive boundedtimefromretainedlogs if a release is missing.
G80is a prospective stronger ordinary control, not an equal-compute guarantee.

Future compact evidence keeps all6656 extracted training candidates (G2560,F2048,
E2048) and2048 native candidates. Retain21 prospectively fixed fullrawsamples:
firstrecord of each13 train and8 native chunks. Full author checker audits allraw;
compact checker regrades allcandidates and reextracts onlyincludedrawsamples.
Fit ZIP andexpanded bothbelow2,000,000bytes; neverdropadverseoutcomes tofit.

For a substantive paper12extension, jointly distinguish conditional-source
capability, hard/current-dense/stable-dense conversion andnativelearning, while
retaining their nonfactorial design differences. Do not submit an unfinished
manuscript or turn routine packaging into a separate contribution.
