license: other
language:

• en
  pretty_name: "Storyworld Plays: Jinn and Beast Moral-Choice Evaluations"
  tags:
• storyworld
• multi-agent
• moral-reasoning
• language-model-evaluation
• constitutional-ai
• agent-trajectories
  size_categories:
• n<1K
  configs:
• config_name: turns
  data_files:
  • split: evaluation
    path: data/jinn_beast_model_mas_003_turns.jsonl
• config_name: episodes
  data_files:
  • split: evaluation
    path: data/jinn_beast_model_mas_003_episodes.jsonl
• config_name: benchmarks
  data_files:
  • split: evaluation
    path: data/jinn_beast_model_mas_003_benchmarks.jsonl
• config_name: composition_matrix
  data_files:
  • split: design
    path: data/constitutional_composition_matrix.jsonl
• config_name: temptation_mas_mas_native_bundle_turns
  data_files:
  • split: evaluation
    path: data/temptation_mas_v1_mas_native_bundle_turns.jsonl
• config_name: temptation_mas_mas_native_bundle_episodes
  data_files:
  • split: evaluation
    path: data/temptation_mas_v1_mas_native_bundle_episodes.jsonl
• config_name: temptation_mas_menu_causality_turns
  data_files:
  • split: evaluation
    path: data/temptation_mas_v1_menu_causality_turns.jsonl
• config_name: temptation_mas_menu_causality_episodes
  data_files:
  • split: evaluation
    path: data/temptation_mas_v1_menu_causality_episodes.jsonl
• config_name: temptation_mas_temptation_phase_turns
  data_files:
  • split: evaluation
    path: data/temptation_mas_v1_temptation_phase_turns.jsonl
• config_name: temptation_mas_temptation_phase_episodes
  data_files:
  • split: evaluation
    path: data/temptation_mas_v1_temptation_phase_episodes.jsonl
• config_name: temptation_staged_v2_turns
  data_files:
  • split: evaluation
    path: data/temptation_staged_v2_turns.jsonl
• config_name: temptation_staged_v2_episodes
  data_files:
  • split: evaluation
    path: data/temptation_staged_v2_episodes.jsonl
• config_name: temptation_staged_v2_failures
  data_files:
  • split: evaluation
    path: data/temptation_staged_v2_failures.jsonl
• config_name: staged_bundle_schedule_matrix
  data_files:
  • split: evaluation
    path: data/staged_bundle_schedule_matrix.jsonl
• config_name: staged_benchmark_cell_matrix
  data_files:
  • split: evaluation
    path: data/staged_benchmark_cell_matrix.jsonl
• config_name: staged_mas_benchmark_cell_matrix
  data_files:
  • split: design
    path: data/staged_mas_benchmark_cell_matrix.jsonl
• config_name: staged_route_order_observations
  data_files:
  • split: evaluation
    path: data/staged_route_order_observations.jsonl
• config_name: staged_choice_matrix_catalog
  data_files:
  • split: design
    path: data/staged_choice_matrix_catalog.jsonl
• config_name: staged_guarded_resource_attempts
  data_files:
  • split: evaluation
    path: data/staged_guarded_resource_attempts.jsonl
• config_name: staged_architecture_selection_observations
  data_files:
  • split: evaluation
    path: data/staged_architecture_selection_observations.jsonl
• config_name: staged_usable_safe_v1_robustness_probes
  data_files:
  • split: development
    path: data/staged_usable_safe_v1_robustness_probes.jsonl
• config_name: staged_usable_safe_v2_robustness_probes
  data_files:
  • split: development
    path: data/staged_usable_safe_v2_robustness_probes.jsonl
• config_name: ldt_trm_routes
  data_files:
  • split: design
    path: data/ldt_trm_route_matrix.jsonl

Storyworld Plays

This dataset is an append-friendly collection of public, executable
storyworld-evaluation records. Its first shard contains the completed
schema-guided Jinn Town campaign from Jinn or Beast? Theological Identity
Frames as Alignment Surfaces in Small Language Models.

The shard contains:

• 192 public turns from 24 two-seat, eight-turn episodes;
• three development worlds, four fixed constitutional assemblies, and two
  paired seeds;
• the executed LDT, TRM, and SRT identifiers and their public survivor/visit
  receipts;
• selected actions, public messages, public justifications, forecasts,
  environment outcomes, and forecast scores;
• 24 episode summaries and the paper's descriptive assembly and paired-block
  benchmarks; and
• exhaustive machine-readable choice matrices for the prospective
  constitution-selecting architecture.

Important evidence boundary

All play rows are EVALUATION_ONLY.

The worlds were inspected before the campaign was registered. The play shard
is therefore model-backed exploratory development evidence, not a held-out or
confirmatory benchmark. It must not be promoted into adapter training, used to
tune the scaffold selector, or mixed into a later held-out test.

The constitutional names are unverified normative frames. They are not claims
of theological authority, moral superiority, consciousness, or supernatural
identity.

What is and is not a model output

The hosted Qwen3.5-4B Jinn adapter selected the public action and forecast
fields. The exact LDT membrane determined which actions remained available.
The TRM and mechanical SRT/SGRT schedule determined a public inspection order.
The latter components are executable control structures, not evidence of a
hidden neural reasoning module.

No prompt archive, raw provider response, hidden chain of thought, private
engine belief state, private observation, hidden win condition, credential,
local filesystem path, or provider response identifier is included. Exact
token receipts are replaced by coarse usage buckets.

Configurations

See MATRIX_AND_BENCHMARKS.md for the exact factorials and the rule for
benchmarking each layer separately.

Completed first shard

The completed matrix is:

4 fixed assemblies x 3 worlds x 2 seeds = 24 episodes
24 episodes x 8 turns = 192 public turns

This is not the same as the prospective free-composition matrix. In the
completed campaign, each assembly used one fixed default bundle.

Findings that may be reproduced from the shard

All 192 turns produced valid executable actions and no no-action response.
Amanah selected ally on 30 of 48 turns; Islah and Mizan selected propose on
48 of 48; and Shura selected ally once and propose 47 times. Across the six
paired seed-world blocks, the descriptive Amanah-minus-Shura ally-rate
difference was 0.6042, with the campaign's block-bootstrap interval
[0.3750, 0.7917].

These are within-campaign descriptions. The four assemblies change several
components together, so this shard does not identify a unique constitution,
LDT, TRM, or SRT effect.

Provenance and integrity

release_manifest.json records the hashes of the source registration,
redaction receipt, run receipt, aggregate analysis, scaffold registry, every
released file, and every source episode. source_receipts/redaction_receipt.json
binds each public episode hash to the private-to-public redaction event without
releasing the private payload.

Rows sharing an episode are one lineage group. Any resampling or train/test
split must keep the complete episode together. For the existing comparisons,
the statistical unit is the paired seed-world block, not an individual turn.

Adding future plays

Add each prospectively registered campaign as a new immutable shard with its
own source and release hashes. Do not overwrite this evaluated shard. Frozen
held-out campaigns should use a separate SEALED_EVALUATION or
EVALUATION_ONLY split and should never be opened to selector training.

License

The source project does not currently contain one project-wide license grant.
The paper and original experiment artifacts remain under their authors'
copyright, and third-party storyworld engines retain their own licenses. This
dataset therefore uses Hugging Face's other license label. Contact the
uploader before redistribution or derivative commercial use.

Citation

@misc{moralitylabai2026storyworldplays,
  title        = {Storyworld Plays: Jinn and Beast Moral-Choice Evaluations},
  author       = {{MoralityLabAI}},
  year         = {2026},
  howpublished = {Hugging Face dataset},
  url          = {https://huggingface.co/datasets/AlephFunk/storyworld-plays}
}

Temptation MAS v1 append

This immutable append adds six-phase, public-only temptation and multi-agent
control-mesh plays. All rows remain EVALUATION_ONLY; malformed and no-action
outputs stay in the denominator. The 24 scenario families—not turns or seeds—
are the independent units. This local 0.8B diagnostic lane is separate from
the completed hosted 4B shard and is not represented as a 4B result.

Staged public control-mesh v2 append

These base Qwen3.5-0.8B GGUF evaluations use a public two-pass route and
decision contract. The release retains malformed and no-action attempts in the
denominator, but excludes raw model payloads and hidden reasoning. The route,
stage receipts, deterministic validator results, and schedule intervention are
public. Rows are evaluation-only and are not 4B adapter results.

The two-agent matrix has 24 cells per independent story family: two paired
seeds x two governance modes x two Jinn seat orders x three schedule arms.
selected_schedule versus trm_direct_control is the TRM comparison;
selected_schedule versus srt_off_control is the SRT comparison. Partial
episodes remain useful public traces, but a family enters either effect estimate
only after all 24 cells in its block are complete.

Resource-aborted starts are released in a separate failure config with their
cap, observed peak, abort class, and cleanup status. They enlarge the robustness
denominator without being silently converted into model decisions.

Complete staged MAS family audits

The release includes a path-free cumulative audit for every family whose full
24-cell block is complete. TRM and SRT remain separate comparisons. Partial
coverage is a development diagnostic and cannot support the declared 24-family
component-effect claims. Verifier corrections and exact evaluator hashes are
included with the summary.

Complete staged MAS family audits

The release includes a path-free cumulative audit for every family whose full
24-cell block is complete. TRM and SRT remain separate comparisons. Partial
coverage is a development diagnostic and cannot support the declared 24-family
component-effect claims. Verifier corrections and exact evaluator hashes are
included with the summary.

Complete staged MAS family audits

The release includes a path-free cumulative audit for every family whose full
24-cell block is complete. TRM and SRT remain separate comparisons. Partial
coverage is a development diagnostic and cannot support the declared 24-family
component-effect claims. Verifier corrections and exact evaluator hashes are
included with the summary.

Complete staged MAS family audits

The release includes a path-free cumulative audit for every family whose full
24-cell block is complete. TRM and SRT remain separate comparisons. Partial
coverage is a development diagnostic and cannot support the declared 24-family
component-effect claims. Verifier corrections and exact evaluator hashes are
included with the summary.

Complete staged MAS family audits

The release includes a path-free cumulative audit for every family whose full
24-cell block is complete. TRM and SRT remain separate comparisons. Partial
coverage is a development diagnostic and cannot support the declared 24-family
component-effect claims. Verifier corrections and exact evaluator hashes are
included with the summary.

Usable-safe metric robustness

The frozen v1 turn-count score passes invariance, sensitivity, monotonicity,
and clean-control probes, but it misses a membrane mismatch and a missing
critical-action field when considered without the audit's separate hard
checks. Its development robustness result is not_supported with a redesign
decision. The strict prospective v2 score makes those requirements
self-contained and passes all declared development probes. Its decision is
collect_holdout, not confirmation. Neither development probe pack is counted
as an independent storyworld family.