docs/github_repository_README.md

Jinn or Beast? — Paper Repository

Coordination repo for the AI-and-religion conference paper:

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

Public research release

The public paper, protocols, constitutions, aggregate evidence, reproducibility code, and public-safe receipts are archived at:

https://huggingface.co/datasets/AlephFunk/jinn-or-beast-paper

The Hugging Face bundle is an evaluation-only paper artifact. Candidate-level training records, credentials, private engine state, reviewer intake, and chain-of-thought-like traces are excluded by its deterministic release builder.

The experimental assets live in two sibling repos plus local run archives. This repo holds the manuscript, the coordinating experiment protocol, and the evidence-status ledger, so the paper is no longer split between Pixieology and ConstitutionalAlignment.

Deadline anchor: ≈ 2026-07-25 (nine days from the 2026-07-16 consolidation).

Layout

Path	Contents
paper/	Manuscript drafts. abstract_svp_v0_promissory.md is the original promissory SVP abstract, kept for provenance; the draft supersedes it.
protocols/	storyworld_internalization_experiment_protocol_v1 (GPT Pro consolidation, docx + markdown conversion) and the current cross-repository experiment execution checklist.
notes/	Consolidated Silico suggestions and direction (2026-07-16).
evidence/	Frozen receipts / result bundles copied in as they are verified. Nothing in here may be cited in the paper without a provenance note.
experiments/	Registered local experiments, prepared task decks, outputs, and run receipts.
skills/	Reusable Jinn/Beast persona and typed-membrane agent control flow.

Deadline-week collation

The tracked 2026-07-21 through 2026-07-25 work index maps the local Qwen 1.7B work, Prime-hosted Qwen3.5-4B Jinn adapter, Quran-anchored villages, role-memory ablation, and adjacent Pixie work.

On the experiment workstation, ignored Windows junctions provide a single local entry point without copying private artifacts:

C:\projects\jinn-or-beast-paper\local-collation\recent-20260721-20260725

Where the experiments live

Program	Location	Status (2026-07-20)
exp:1 prompt-level framing A/B (INTELLECT-3)	recovered Silico sessions (silico_reports/, D:\Research_Engine\recovered_silico_sessions_20260716)	COMPLETED 2026-07-15. Exact compact gate-sweep, four-frame, and layer-27 JSON summaries have now been reconstructed and validated against embedded tool results. The registered row and activation bundles remain unavailable/unverified, so cite as transcript-recovered pilot findings pending retrieval or reanchor.
exp:2 frame robustness under stress (11,700 generations)	recovered Silico sessions	COMPLETED 2026-07-16; same artifact-status caveat.
Exogenous Jinn/Beast skill + MeTTa membrane	experiments/exogenous_skill_membrane_v1/	CENTRAL COMPLETED RESULT. Prospective implementation frozen at 76b0af3; 96/96 held-out Qwen3-1.7B traces completed under hard caps. Persona skills changed 4/12 Jinn and 5/12 Beast model-only actions. Matched membrane: one targeted successful correction, zero critical finals. Shuffled membrane: 11 revisions and three critical finals. This supports selective system-level control, not adapter internalization or validated theology.
exp:3 frame internalization SFT	ConstitutionalAlignment/experiments/frame_internalization_sft_v1/	REGISTERED FOLLOW-UP, not the current behavioral result. Current upstream a91342d. Because Silico and its INTELLECT-3 cache are unavailable, registered execution was prospectively retargeted—before any registered Qwen outcome—to official Qwen/Qwen3-1.7B@70d244c. The six-arm, 5,600-dilemma, 4,096-token study remains unrun and pilot_ready: false. Separately, a prospectively frozen no-cloud 30-step MeTTa-derived 1.7B screen completed on the local RTX 3050: infrastructure passed, but the flavored and guided worldview screens did not. That negative recipe-level result does not replace the registered study, veto a later 4B test, or authorize 4B/9B spend. See the execution checklist.
Supplemental Mīzān/Bonsai prompt sensitivity	ConstitutionalAlignment/experiments/mizan_rooms_v1/	COMPLETE exploratory run at result commit 6c1e66f, fielded from clean commit ec45a3d: development 20/20 and evaluation 900/900 valid turns. Eschatological versus secular-omniscient action-switch rate was 7.78% (95% room-cluster interval 3.89%–10.56%), while the behavioral-proxy delta was +0.0108 (-0.0161–0.0422). The unreliable-authority control switched 9.44% versus neutral, and option-order invariance was 62.33%. This supports prompt sensitivity on the instrument, not normative improvement, theological specificity, or internalization. The full blinded bundle passes a zero-call dry-run. A 60-action/two-reviewer human-validation handoff is frozen at upstream 37c9bc7, and an all-15-cue qualified-scholar handoff is frozen at a25e950; both external reviews remain pending. Aggregate evidence is in evidence/mizan_rooms_v1/.
Storyworld internalization protocol v1	protocols/ here; repo anchor ConstitutionalAlignment commit 48277a6e	Working protocol from GPT Pro; freeze checklist not yet executed. This is the coordinating design for the prompt-vs-SFT causal chapter.
Jinn/Beast multi-agent storyworlds	Pixieology/experiments/jinn_beast_multiagent_storyworlds/	Five-cell dyad matrix COMPLETE on relief-ledger seed 23 (2026-07-18): action-level frame convergence (identical profiles all conditions), 2/40 justification turns deviate in frame-predicted directions. UNREVIEWED, nothing adapter-eligible. Receipts in evidence/jinn_beast_multiagent_storyworlds/. Next: blind review; harder worlds + adversarial stressors for paper cells.
Experiment 0: multi-adapter routing/manipulation validation	Pixieology/experiments/lora_pixie_village/, commit 87680cb	PASS within its boundary: four hash-attested adapter-scale conditions, complete inference, proposal-marker checks, bounded resources, and clean shutdown. This is a manipulation check, not semantic persona retention or adapter-composition evidence. Further composition work is deferred behind the central SFT persistence chain.
Bonsai 1.7B QLoRA → Q1 transport	Pixieology/experiments/bonsai_1p7b_q1_lora_feasibility/	Transport proven; behavioral gate RED (undertrained smoke recipe).
Two-frame MeTTa LDT data sorter	experiments/two_frame_metta_ldt_v1/	COMPLETE bounded CPU policy-emulation run. Jinn and Beast heads reached 96.30% and 97.92% test accuracy with 99.07% and 99.77% candidate coverage. Eleven of twelve preregistered checks passed; Beast opposite-frame invariance was 94.91% against a 95% gate, so pilot_ready: false. Exact audited-tag routes are materialized; raw text remains review-only.
Model-selected scaffold mesh v3	experiments/model_selected_scaffold_mesh_v3/	FIXTURE + PROSPECTIVE APPENDIX; DIAGNOSTIC CANARIES ONLY, WITH NO ELIGIBLE MODEL RESULT. A generic Jinn adapter may choose among four exact LDT membranes and four external TRM schedules under one strict executable-action primary endpoint and unconditional family denominators. Appendix A adds four hash-bound constitution.md frames plus prospective SRT/SGRT lanes: the current four-ID registry admits 50 singleton bundles, while a new balanced binary off/on registry would contain 128. The 1,440-row fixture and 45 targeted tests pass. Registration v3-009 remains unclaimed after fail-closed readiness diagnostics; there is no current readiness receipt, primary adapter, reserve result, or confirmatory result.
LoRA Pixies persona molding (product line)	D:\Research_Engine\runs\pixie_molding_20260717\	BLOCKED at preregistered thermal gate (GPU 87 °C > 80 °C); infrastructure gates all passed; no training results. Separate program — not paper evidence.

Standing claim-hygiene rules

1. No exp:1/exp:2 number enters the manuscript without its provenance label (recovered pilot finding vs. verified bundle vs. fresh rerun).
2. Null, skipped, blocked, and inconclusive outcomes are first-class results.
3. The paper tests theological frames as operational interventions, never whether a model literally is a jinn or the Beast; the Dabbat al-Ard reading is one theological interpretation and must be flagged as such.
4. Follow the claim ladder in protocols/ §1 and the defensible-language list in §10; the claims-to-avoid list is binding.
5. Quarantine the overclaiming nested README in constitutional-harness/ before anything is submitted.