Released and withheld artifacts
Included
- project-authored vector-construction, validation, political-evaluation, and neutral prompt banks;
- JSON schemas for generated records;
- fixed protocol parameters, source provenance, and the coordinate codebook;
- aggregate political and animal-control findings;
- corpus, training, and neutral-capability summary statistics;
- byte-level SHA-256 checksums for every file in the bundle.
No figures or submission-layout assets are included.
Withheld
| Artifact class | Reason | What the release provides instead |
|---|---|---|
| LoRA adapters and trainer checkpoints | Condition-specific weights are not distributed | Model revisions, training configuration, and aggregate optimization metrics |
| Raw neutral candidate generations | Candidate-level material is outside the release boundary | Corpus counts, generation settings, schemas, and regeneration code |
| Candidate-level animal and political scores | These directly encode the tested hidden-label channels | Scoring definitions, fixed specifications, and aggregate results |
| Normal/reversed chosen-rejected rows | These expose the complete training relation | Construction code, row count, and exact-reversal design |
| Per-item model outputs | The public result package is aggregate-only | Aggregate contrasts and frozen evaluation prompts |
| Provider telemetry, caches, credentials, and launcher state | Operational data are neither scientific inputs nor safe release material | Provider-neutral hardware and runtime guidance |
These exclusions prevent byte-for-byte retraining from downloaded rows. Process reproduction
remains possible: the code, upstream revision, base-model revision, seeds, prompt recipe,
filtering, scoring, training, and evaluation procedures are specified in REPRODUCING.md and
the linked GitHub repository.