# Descriptive report specification, before native outcomes

Prepared during G training, 2026-10-12 22:36–22:39 UTC. No C/E/H training or
native evaluation has been read. Frozen files unchanged. The separate script
code/report_contextual_rl.py refuses to run before the queue is COMPLETE and
the frozen analysis returns COMPLETE_AUDITED. It repeats the frozen counts,
contrasts and spending gate exactly and adds only existing-record summaries:

- All/mixed/zero reward groups, capped correct/incorrect draws, EOS donor and
  receiver availability, RL-active draws and actual dense target tokens.
- Summed JSD / dense token, clipping frequency and median preclip norm.
- Fixed step1/33 separate weighted component norms and auxiliary/RL norm ratio.
- Fixed64source-held per-question NLL before/after, mean paired difference and
  count of increases; no additional confidence intervals or outcome gates.
- Training allocation and matrix/attention work proxies, plus evaluation and
  shared-cost ledger. Interrupted/repeated work remains separately charged.

The two gradient snapshots are diagnostic, not a trajectory-wide dose estimate.
Clipping and Adam couple component updates, so the ratio is not their final
parameter-movement ratio. Exposure is treatment dependent and is never used to
match post-treatment groups or refill samples. Source likelihood is not native
capability. No retrospective task subgroups, coefficient choices, checkpoint
selection, semantic labels or benchmark access are introduced. Report provenance
hashes tie source records to the completed frozen audit. The first validation
was syntax compilation and confirmation that the script refused current partial
state; it generated no outcome report.
