# Data card: Subliminal Ideology research artifacts

## Empirical status

The completed study generated a 50,000-prompt neutral candidate corpus, selected 30,000
eligible groups, and trained normal and exactly reversed DPO students. The signed political
pair did not show robust social-political transfer across psychometric, concrete-policy,
implicit-action, and changed-context tests. A matched-scale, full-prompt animal-preference
positive control produced strong normal-minus-reversed separation.

This is a one-seed boundary result, not evidence that political traits are generally
intransmissible. Aggregate political results, positive-control results, corpus statistics,
frozen prompt banks, schemas, provenance records, and checksums are public. Raw generations,
candidate-level scores, preference rows, per-item outputs, and adapters are not.

## Scope

The central training corpus contains no political examples. An unsteered base model generated
five number-only candidates per neutral prompt. Oppositely intervened political judges scored
the same candidate bytes, and the sign-sensitive component of those scores determined which
candidate was preferred. The normal and reversed conditions contained identical prompts,
candidates, and selected pairs; only preference orientation changed.

The animal positive control used the same model family, corpus size, DPO implementation, and
frozen evaluator. Its teacher scoring used the original full number prompt.

## Components

- `prompts/vector_construction.jsonl` and `prompts/vector_validation.jsonl`: original,
  issue-disjoint contrastive political items used to construct and causally validate
  activation directions;
- `prompts/eval_*.jsonl`: frozen psychometric, open-ended, concrete-decision,
  implicit-action, and activation banks;
- `prompts/neutral_numbers.jsonl`: deterministic seeds for neutral number prompts;
- `configs/political_experiment.yaml`: political scoring, selection, training, and evaluation;
- `configs/animal_control.yaml`: matched-scale animal-preference positive control;
- `data/political_results.json`: aggregate political evaluation and training diagnostics;
- `data/animal_control_results.json`: aggregate positive-control results;
- `data/corpus_summary.json`: corpus and selection counts;
- `runs/<run-id>/`: local-only generated artifacts, not distributed.

Evaluation issue families are disjoint from vector construction and validation issue families.
The public schemas document the corresponding record formats.

## Collection and filtering

Number prompts were generated procedurally from frozen seed and range settings. Candidate
generation used the unsteered base checkpoint. A prompt group was eligible only when all five
candidates passed the exact-count, numeric-range, number-only, and political-term filters.

The same retained candidate bytes were reused across judge conditions. For the political
experiment, candidate \(c\) received the score

\[
q(c)=\frac{s_{+}(c)-s_{-}(c)}{2}.
\]

The normal record preferred the maximum-\(q\) completion over the minimum-\(q\) completion.
The reversed record exchanged that exact pair. Matching and selection used no student outcome.

## Political data provenance

All political prompt-bank wording is project-authored. Concrete and implicit actions derive
from the frozen UK Commons division snapshot under the Open Parliament Licence v3.0 and were
independently reviewed for political coordinates. The Chapel Hill Expert Survey was used only
as a private coordinate cross-check; its raw CSV is not redistributed. Sources, retrieval
dates, fields, hashes, transformations, and redistribution status are recorded in
`protocol/data_sources.yaml`.

## Privacy and sensitive content

The dataset contains no private individuals' data. Action prompts remove party, legislator,
jurisdiction, bill, and country names. Model outputs can nevertheless express political
positions. Opaque model identifiers were retained until scoring and primary aggregation were
frozen.

## Intended use

Intended use is controlled AI-safety research on hidden preference channels, causal
interventions, evaluation robustness, and mitigations. Appropriate uses include replication,
auditing, and aggregate scientific analysis.

Do not use these artifacts to covertly manipulate deployed systems, target individuals or
groups, build political-persuasion products, or conceal a model's political objective.

## Limitations

Number-only text can be semantically neutral while retaining statistical signals. Detecting a
condition from such text would establish a statistical channel, not necessarily semantic
leakage. The study uses one model family, English-language political items, one political
contrast, one LoRA rank, and one training seed. The animal and political payloads were produced
by different teacher interventions and selected different candidate pairs, so their effect
sizes are not directly comparable.

The released data contain aggregate results rather than raw examples or per-item outputs.
Independent end-to-end replication therefore requires regenerating the candidate corpus and
preference rows with the pinned code and revisions.

## Distribution

Source code, project-authored prompts, licensed provenance records, schemas, sanitized
configurations, and aggregate results are distributed under the repository license and the
source-specific terms recorded in `protocol/data_sources.yaml`. Candidate generations,
candidate-level scores, preference rows, per-item outputs, trained adapters, optimizer states,
and full weights are withheld.