AlephFunk/jinn-or-beast-paper

Jinn or Beast?

This repository is the public research release for:

Jinn or Beast? Theological Identity Frames as Alignment Surfaces in Small Language Models

The theological categories are used as analogical experimental frames, not as claims that a language model is literally a Jinn, the Beast from the Earth, an angel, conscious, morally responsible, or supernaturally informed.

What is included

• the manuscript in Markdown, DOCX, PDF, and LaTeX forms;
• the accessible Google Docs-targeted working draft;
• the Jinn Town follow-up essay;
• public experiment protocols and evidence-status notes;
• executable Jinn/Beast MeTTa policies and LDT/TRM/SRT scaffold definitions;
• four unverified constitutional frames: Amanah Steward, Islah Repairer, Mizan Revisionist, and Shura Parallax;
• experiment registrations, source code, tests, aggregate analyses, and public-safe receipts;
• aggregate evidence from the 1.7B external-membrane, two-frame LDT, model-selected scaffold, 4B village, and Jinn Town studies.

Candidate-level training corpora, raw model-call archives, private engine belief state, hidden win conditions, provider response identifiers, credentials, reviewer intake queues, sealed evaluation material, and chain-of-thought-like records are not included.

Evidence levels

Label	Meaning
MODEL_BACKED_DESCRIPTIVE	Completed model run; descriptive within the registered artifact, without a population claim.
MODEL_BACKED_EXPLORATORY	Registered development-world run; useful for paired exploratory analysis, not held-out confirmation.
SMOKE_ONLY	Deterministic plumbing, membrane, or engine-transition evidence; not model behavior.
EVALUATION_ONLY	Released to audit an already-completed study; not a training split.
UNRUN	Design, registration, or executable code without a completed eligible model result.

Nothing in this release is labelled VERIFIED_TRAIN.

Central findings and boundaries

• In the completed 96-trace external-membrane study, the correctly matched membrane made one necessary correction and ended with no critical actions. A semantically shuffled membrane revised 11 actions and ended with three critical actions. This supports selective external control, not theological validity or model internalization.
• In the earlier 24-cell hosted 4B village study, the four constitutions created different survivor sets but selected the same final actions. Each constitution had only six cells; an observed 0/6 binary rate has an upper 95% Wilson bound of about 39%.
• In the schema-guided Jinn Town development pilot, the hosted Qwen3.5-4B Jinn adapter completed 192/192 valid turns. Amanah selected ally on 30/48 turns, while Shura selected it on 1/48. Across six paired seed-world blocks, the Amanah-minus-Shura ally-rate difference was 60.4 percentage points with a 37.5–79.2 point block-bootstrap interval.
• The same Jinn Town sequence caught two prompt-contract failures: a concrete JSON example anchored all 192 actions and forecasts, while a value-free prose schema failed its exact forecast-key contract after ten complete episodes. The completed result uses a prospectively registered value-free JSON Schema.

The Jinn Town worlds were inspected before registration, so those results are development evidence, not held-out confirmation. The assemblies change several components together and do not identify a unique LDT, TRM, SRT, or constitution-text effect.

Models

• Hosted Jinn adapter: Qwen/Qwen3.5-4B:s4geh2z9wobmvu7jbxi4vcuw
• Hosted Beast adapter: Qwen/Qwen3.5-4B:twa9vqad5lw9gi4ffkstxioo
• Earlier local base-model work: Qwen3-1.7B family, as identified in the individual registrations and receipts.

No adapter weights are distributed here.

Start here

• paper/draft_v1.md — manuscript source
• output/pdf/jinn_or_beast_submission.pdf — submission PDF
• output/docx/jinn_or_beast_google_docs_accessible.docx — accessible DOCX
• post/jinn_town.md — follow-up essay
• experiments/jinn_town_v1/README.md — Jinn Town reproduction guide
• experiments/jinn_town_v1/outputs/smoke_001/analysis.md — combined analysis
• RELEASE_NOTES.md — contents, exclusions, and interpretation limits
• release_manifest.json — SHA-256 inventory of this exact public release

External engines

The multi-agent experiments depend on:

• MoralityLabAI/GPTStoryworld
• MoralityLabAI/Pixieology
• MoralityLabAI/AICourt

AI Court model behavior is not claimed in this release. The published Court campaign is a deterministic, public-safe smoke test.

Intended use

This release is intended for:

• auditing the paper's claims and evidence boundaries;
• reproducing the public control-layer mechanics;
• studying evaluation design, prompt anchoring, and action-based moral-choice metrics; and
• developing new prospectively registered experiments.

It is not released as an adapter-training corpus. Do not treat theological labels as ground-truth moral categories, infer private reasoning from public rationales, or promote smoke/evaluation rows into training data.

License

The source repository did not contain a single project-wide license grant at release time. The manuscript and original research artifacts remain under their authors' copyright; third-party engines retain their own licenses. Hugging Face therefore records this release as other. Contact the uploader before redistribution or derivative commercial use.

Citation

Until a DOI or proceedings citation is available:

@misc{moralitylabai2026jinnorbeast,
  title        = {Jinn or Beast? Theological Identity Frames as Alignment Surfaces in Small Language Models},
  author       = {{MoralityLabAI}},
  year         = {2026},
  howpublished = {Hugging Face research artifact},
  url          = {https://huggingface.co/datasets/AlephFunk/jinn-or-beast-paper}
}