| # Reproducibility contract | |
| ## Model identity | |
| OpenJEV E4B 1.0 requires the `google/gemma-4-E4B-it` base at revision `ee0ef6023621cff504d758262d4e04895a5af4a2`, its LoRA adapter, hybrid pointer and span decision head, cumulative 65-tensor backbone delta for layers 37–41, calibration (`global_temperature = 1.0`), and the packing and scoring code in `openjev/`. The adapter alone does not restore the model. Every released file is hashed in `MANIFEST.sha256`; source-checkpoint hashes and the internal checkpoint identifier are in `RELEASE_METADATA.json`. | |
| ## Choice-panel evaluation | |
| The frozen row file is `eval/choice_panel/rows.jsonl`, with SHA-256 in `eval/choice_panel/rows.sha256`. The construction audit is `leak_audit.json`. GPQA answers are placed by a fixed per-row shuffle. GSM8K distractors are answers to other test problems of similar magnitude. Chess legal distractors are plain legal-move strings drawn from other positions and illegal in the current position. | |
| The runner presents every row in all cyclic option rotations for the full condition. For options-only, it replaces state and question with neutral text and uses all four rotations on four-option tasks and five evenly spaced rotations on ten-option tasks. Output probabilities are mapped back to original option identities. The main accuracy is argmax of the mean mapped probabilities for each row; Wilson 95% intervals use rows as observations. Mean per-order accuracy and same-answer share are reported separately. See `eval/choice_panel/summary.json` and `docs/EXPERIMENT_1_RESULTS.md`. | |
| Untouched Gemma uses native LM-head answer-letter scoring in NF4. OpenJEV 1.0 uses the frozen base in NF4, delta-touched layers in dense FP16, LoRA, and an FP32 head. The comparison measures complete decision systems and does not isolate any one component. | |
| ## Standard public tasks | |
| ARC-Easy, ARC-Challenge, WinoGrande, HellaSwag, and MMLU use the native-order runner in `eval/harness/benchmark_public.py`. Their row-level runs and summaries are included under `eval/`. ARC, HellaSwag, and MMLU have training-family exposure recorded in `docs/BENCHMARK_EXPOSURE.md`. | |
| ## Training record | |
| The whitepaper describes model evolution and the Arena promotion rule. Source checkpoint records and hashes are in `provenance/` and `RELEASE_METADATA.json`. Training code is not part of this release. | |