# Frozen contextual-teacher evidence

This compact bundle supports a prospective G/F/E study on a fresh Warm37041.
All4 W/G/F/E native endpoints precede grading. Read analysis/result.json,
analysis/descriptive_report.md and notes/stable_contextual_contract.json.

Run from the expanded bundle with the supplied pinned environment, shared baseline
code/data mounted read-only, and a new output directory OUTSIDE this bundle:

```
CUDA_VISIBLE_DEVICES="" PYTHONDONTWRITEBYTECODE=1 /mnt/stephen/research_data/math_rlvr/.venv/bin/python code/check_stable_contextual_evidence.py --data /mnt/stephen/research_data/math_rlvr/code/data --output /path/to/fresh/audit-directory
```

The checker verifies manifests and available frozen inputs, regrades6656 extracted
training candidates and2048 extracted native candidates, reconstructs masks/donors,
actual update counts and three paired contrasts/gate, checks reported frozen-teacher
digests, and re-extracts21 fixed rawsamples. A supplied teacher digest is a recorded
identity check, not a CPU replay of teacher logits or GPU training.

Rawsamples were fixed before training: first physical record of each13training and
8native chunks. Full author audit re-extracted/regraded all local rawtrajectories.
The public compact audit independently re-extracts ONLYincludedrawsamples. Omitted
raw trajectories, modelweights and optimizer execution remain outside its scope.
Hashes of omitted originals are provenance, not public accessibility or verification.

G80batches versusF/E64; retain distinct schedule/question/drawdenominators. The
actual-context F versusempty E comparison is not an isolated semantic-content
intervention: contexts differ in text/length/teacherpositions and studentexposure
changes. Fixed teacherparameters do not freeze support sets, prefixes or conditional
targetdistributions. No historical current/stable interaction is identified.
Native bootstrap intervals condition on one trainingseed/start and a reused512
questionpanel; no trainingseeduncertainty or finalbenchmarkclaim is measured.

Allinputs, includingthebaseline,parentwarm,anddata, areboundbyhash. Warmcheckpoint
andtaskdata/tokenizer are omitted for size/licensing; warmregenerationrecipe and
source are included. Scientific scripts validate exacthistoricalinputhashes, so a
fresh neuralreplication requires a NEWcontract binding independently regenerated
weights; never edit the historicalcontract orreuseitsnativeanswerrecords as fresh
results. No modelweights orcredentials are bundled. Derivedcode uses Apache-2.0;
retainbaseline/vendorLICENSE. Restoringandrunningneuralexperiments requiresCUDA
and~18GB peakallocated memory; the compactchecker isCPUonly.
