README.md

---
pretty_name: Eleusis Calibrated Rules
task_categories:
- reinforcement-learning
- text-generation
tags:
- inductive-reasoning
- agentic-evaluation
- card-games
- long-horizon
---

# Eleusis Calibrated Rules — 100-turn reward calibration

A calibrated rule dataset for the single-player Eleusis inductive-reasoning
environment. It extends the 26-rule Hugging Face benchmark with controlled
static, transition, conditional, periodic, chunk, higher-order history, global
history, and compositional rule families.

Source benchmark: [Hugging Face Eleusis](https://huggingface.co/spaces/huggingface/eleusis-benchmark).

Dataset version: **v2.1-frontier-calibrated-100turn-20260812**  
Protocol: **eleusis-100-v11**

The structural, GPT Sol difficulty, family, and long-horizon readiness gates passed.

**Gemini confirmation is incomplete:** 11 exact endpoints, 11 right-censored prefixes, and 10 not-run rules. Censored traces are not scored as failures.

## Splits

{
  "train": 959,
  "validation": 289,
  "test": 32,
  "legacy_test": 26
}

- **train:** parameterized training rules.
- **validation:** held-out parameterizations and compositions.
- **test:** 32-rule, family-balanced reward-calibration suite (4 per family,
  one at each declared complexity tier).
- **legacy_test:** unchanged 26-rule Hugging Face benchmark.

## Calibration target and result

- GPT-5.6 Sol solve@100: **0.500**
- GPT-5.6 Sol normalized reward: **0.253**
- Family mean-reward spread: **0.292**
- Exact per-rule calibration panel stored in the test rows: **openai/gpt-5.6-sol**

The 100-turn test was selected by normalized reward and solve rate, with equal
family quotas and explicit within-family complexity tiers. Every family keeps
at least one rule GPT Sol did not solve at 100 turns, and every family reward is
below 0.5. Gemini construction evidence is included with exact/censored/missing
status; no missing or censored episode is imputed into its performance.
Detailed per-rule results, rule-bootstrap intervals, costs, and visualizations
are included with the calibration artifacts. The
[final report](./calibration/FINAL_REPORT.md) explains what was built, how
the calibration was run, and what the family and complexity results imply.

Family coverage in the test split is:

```json
{
  "static_predicate": 4,
  "first_order_transition": 4,
  "conditional_transition": 4,
  "periodic_cycle": 4,
  "chunk_run": 4,
  "higher_order_history": 4,
  "global_history": 4,
  "compositional_hybrid": 4
}
```

## Core fields

Every split retains **rule_id**, **label**, **family**, and **code**. Additional
fields describe template lineage, semantic complexity, observability, empirical
difficulty, and dataset version.

## Reproducibility

Use the published Eleusis taskset with the test split, **100 turns**, temperature
0.7, and 4,096 maximum completion tokens per model call. The engine preserves the legacy deal
prefix and appends deterministic reserve shoes only when the longer game needs
more cards. The legacy split preserves historical comparison with the source
benchmark.

The default harness retains the full transcript and does not compact it. A
provider-neutral user heartbeat follows each tool result, and one continuation
is allowed only after a provider-reported completion-length cutoff. The
calibration report contains measured context growth and the failure policy for
models with smaller context windows.

Calibration is empirical, selection-conditioned, and protocol-specific. Model updates, sampling
variance, or prompt changes can shift absolute scores; use the included
per-rule metadata and calibration artifacts to re-estimate difficulty as the
frontier changes. Fresh models and fresh deal seeds should be used for the
confirmatory full-panel comparison.

Repository: https://huggingface.co/datasets/nph4rd/eleusis-calibrated-rules