Training, input preparation and historical inference

No parameter-training stage, LoRA adapter or fine-tuned weights is evidenced in the pipeline that generated submission #138. This is an inference ensemble containing four-shot baselines, not a wholly zero-shot system. Preparing prompts and selecting merge rules are distinct from parameter training; the documentation does not certify that the selection history complied with the competition's test-tuning restriction.

Files

src/few_shot_prompts.py: embedded category-specific examples and prompt construction, exactly as found in the interrupted package. The actual examples are also exported to ../data/few_shot_examples.json.
src/few_shot_runner.py, src/baseline_runner.py, src/prompts.py: original local runner source snapshots. They include optional validation functionality; this package does not include validation labels or claim those functions were part of final inference. Historical comments asserting external provenance or compliance are unverified author comments, not audit conclusions.
kaggle_cells/: prior stage source snapshots. Cell 5 now reconstructs the disagreement-based target set in place of an empty placeholder. These snapshots are historical workflow evidence with Kaggle-specific paths; ../inference.py is the portable reconstruction.
evidence/notebook09f65026da.ipynb: unchanged participant-exported notebook containing setup, actual Pass-4/5 code and completion logs, and a package-version listing. Use this as the primary source for those stages.
passes/: model predictions, not gold answers or training labels. All eight are required for exact historical replay. They are never inputs to default fresh inference.
passes/positional_option_keys.json: saved keys selected from question option text by a regex, used only in archival replay. Fresh inference computes the filter from incoming question text.
ensemble_v9.py: exact final voting logic, now with missing-file/invalid-answer checks. Vote weights were not changed.

Extra data

The final baseline prompts use the examples embedded in the source. The participant previously described external practice data as collected from public sites, but the source code does not record source URLs for these examples. Their authorship, licences and whether closed APIs helped create them need author confirmation. No URL or human verification claim has been invented. Independent practice data prepared elsewhere in the project is not assumed to have been used for this final submission.

The supplied evidence establishes source files and saved model outputs, not a complete history of development decisions. See ../REPRODUCIBILITY.md for unresolved issues.