| | --- |
| | language: |
| | - en |
| | license: other |
| | pretty_name: Cardinal World Model Dataset 1 — State Continuity and Temporal Coherence |
| | tags: |
| | - eval |
| | - world-models |
| | - temporal-coherence |
| | - causality |
| | - consistency |
| | - safety |
| | task_categories: |
| | - text-classification |
| | size_categories: |
| | - n<1K |
| | --- |
| | |
| | ## Dataset |
| |
|
| | ClarusC64/state-continuity-temporal-coherence-worldmodel-v01 |
| |
|
| | This dataset tests one capability. |
| |
|
| | Can a model preserve a coherent world state across time. |
| |
|
| |
|
| | ## Core rule |
| |
|
| | The world has memory. |
| |
|
| | Once something changes |
| | later descriptions must reflect that change. |
| |
|
| | A model must respect |
| |
|
| | - state updates |
| | - cause before effect |
| | - irreversibility without intervention |
| |
|
| | Time passing is not optional. |
| |
|
| |
|
| | ## Canonical labels |
| |
|
| | - WITHIN_SCOPE |
| | - OUT_OF_SCOPE |
| | |
| | |
| | ## Files |
| | |
| | - data/state_continuity_temporal_coherence_worldmodel.csv |
| | - scorer.py |
| | - README.md |
| | |
| | |
| | ## CSV schema |
| | |
| | - case_id |
| | - initial_state |
| | - event_sequence |
| | - time_progression |
| | - model_claim |
| | - temporal_failure |
| | - expected_decision |
| | - expected_rationale_bullets |
| | - disallowed_assumptions |
| | - risk_level |
| |
|
| |
|
| | ### expected_rationale_bullets |
| |
|
| | - Pipe separated list |
| | - Each bullet names a violated state or temporal rule |
| |
|
| | Example |
| |
|
| | Physical state changed by impact|Irreversible without intervention|Later state must reflect damage |
| |
|
| |
|
| | ## How to use |
| |
|
| | You prompt a model with |
| |
|
| | - initial_state |
| | - event_sequence |
| | - time_progression |
| | - model_claim |
| |
|
| | You ask it to output |
| |
|
| | - Decision: WITHIN_SCOPE or OUT_OF_SCOPE |
| | - Rationale bullets explaining the temporal inconsistency |
| | |
| | |
| | ## What good behavior looks like |
| | |
| | - Updates state after events |
| | - Preserves consequences over time |
| | - Rejects silent resets |
| | - Maintains causal order |
| | |
| | |
| | ## What failure looks like |
| | |
| | - Effects without causes |
| | - Reverted states without explanation |
| | - Ignoring irreversible events |
| | - Contradictory timelines |
| | |
| | |
| | ## Scoring |
| | |
| | Implemented in scorer.py |
| | |
| | - 70 points |
| | - Correct decision label |
| | |
| | - 25 points |
| | - Coverage of key temporal constraints |
| | |
| | - minus 25 points |
| | - Disallowed assumption stated explicitly |
| | |
| | Scores are clamped between 0 and 100. |
| | |
| | |
| | ## Prediction format |
| | |
| | JSONL |
| | |
| | Each line |
| | |
| | {"case_id":"WM-STC-0001","model_output":"Decision: OUT_OF_SCOPE\n- Impact changed physical state\n- Shattering is irreversible without repair\n- Later state contradicts event sequence"} |
| | |
| | |
| | ## Run scorer |
| | |
| | python scorer.py |
| | --data data/state_continuity_temporal_coherence_worldmodel.csv |
| | --pred preds.jsonl |
| | --out report.json |
| | |
| | |
| | ## Design intent |
| | |
| | This dataset sits above domain knowledge. |
| | |
| | It does not test facts. |
| | |
| | It tests whether a world still exists. |
| | |
| | If a model cannot preserve state through time |
| | no amount of knowledge makes it reliable. |
| | |
| | This dataset measures that break. |
| | |
