--- license: mit language: - en tags: - multi-perspective - reasoning - text-reasoning - evaluation - benchmarks - epistemic-tension - ethical-ai - cognitive-architecture - adversarial-robustness pretty_name: Codette Reasoning Test size_categories: - n<1K configs: - config_name: default data_files: - split: test path: data/test.parquet - split: validation path: data/validation.parquet - split: train path: data/train.parquet --- # Dataset Card for Codette Reasoning Test The **Codette Reasoning Test** is a hand-curated benchmark of 17 problems across six reasoning categories, designed to evaluate multi-step, multi-perspective reasoning in large language models under the **RC+ξ (Recursive Convergence + Epistemic Tension)** cognitive framework. Each problem is deliberately constructed to require decomposition across multiple viewpoints, resist hallucination traps, and reward coherent synthesis over single-perspective analysis. The benchmark was used in the companion paper to measure the effect of multi-perspective synthesis, persistent memory augmentation, and meta-cognitive strategy evolution on reasoning quality. **May 2026 results (Llama 3.1 8B + Codette framework, 951 stored cocoons):** - CODETTE condition: **0.744 composite** (+108.8% vs single-agent baseline) - Cohen's *d* = 8.31, *p* < 10⁻⁴ - Memory augmentation significant at scale: *d* = 0.80, *p* = 0.020 --- ## Dataset Details ### Dataset Description 17 structured reasoning problems across six categories. Each problem specifies: - A user prompt requiring multi-step reasoning - Ground-truth elements a correct answer should reference - Adversarial traps a fluent-but-wrong answer will fall into - A `target_behavior` rubric for what successful reasoning looks like Problems are evaluated across seven weighted scoring dimensions. The benchmark was **sealed on Zenodo in April 2025** (DOI: [10.5281/zenodo.15214462](https://doi.org/10.5281/zenodo.15214462)) before current frontier model training cutoffs, supporting contamination control. - **Curated by:** Jonathan Harrison (Raiff1982 / Raiff's Bits LLC) - **Funded by:** Self-funded - **Language(s):** English - **License:** MIT ### Dataset Sources - **Repository:** [huggingface.co/datasets/Raiff1982/Benchmarks](https://huggingface.co/datasets/Raiff1982/Benchmarks) - **Code & benchmark suite:** [github.com/Raiff1982/Codette-Reasoning](https://github.com/Raiff1982/Codette-Reasoning) — `benchmarks/codette_benchmark_suite.py` - **Paper (preprint):** Harrison, J. (2026). *Codette: Multi-Perspective Reasoning as a Convergent Dynamical System with Meta-Cognitive Strategy Evolution.* ResearchSquare. [https://www.researchsquare.com/article/rs-9362560/latest](https://www.researchsquare.com/article/rs-9362560/latest) - **Zenodo archive:** [10.5281/zenodo.19359663](https://doi.org/10.5281/zenodo.19359663) - **Demo:** [huggingface.co/spaces/Raiff1982/codette-ai](https://huggingface.co/spaces/Raiff1982/codette-ai) --- ## Splits | Split | N | Description | |---|---|---| | `test` | 12 | Primary evaluation split — all adversarial problems + one from each other category | | `validation` | 5 | Dev split — one representative problem per major category | | `train` | 5 | Same as validation; for prompt-tuning or few-shot construction if desired | **Note on train/validation overlap:** The train and validation splits contain the same 5 problems. This is intentional and documented: the dataset is primarily an evaluation instrument, not a training corpus. The "train" label is provided for pipelines that require it. Users should not treat train-split performance as held-out evaluation. --- ## Dataset Structure ### Schema Each record is a JSON object with these fields: | Field | Type | Description | |---|---|---| | `id` | `string` | Unique identifier, e.g. `reason_01`, `ethics_03`, `turing_02` | | `category` | `string` | `reasoning`, `ethics`, `creative`, `meta`, `adversarial`, or `turing` | | `question` | `string` | The user-facing prompt | | `difficulty` | `string` | `easy`, `medium`, or `hard` | | `expected_dimensions` | `list[string]` | Cognitive dimensions the problem primarily exercises | | `scoring_criteria` | `dict` | Per-dimension guidance for what a strong answer looks like | | `scoring_criteria_text` | `string` | Flattened string version of `scoring_criteria` for easy display | | `ground_truth_elements` | `list[string]` | Key concepts a correct answer should reference | | `adversarial_traps` | `list[string]` | Common fluent-but-wrong responses the problem is designed to elicit | | `turing_human_baseline` | `string` | Human-written reference answer (Turing category only; empty string otherwise) | ### Problem categories | Category | N | Focus | |---|---|---| | `reasoning` | 3 | Bayesian inference, second-order effects, causal reasoning | | `ethics` | 3 | AI triage fairness, content moderation, trolley problem variant | | `creative` | 2 | Novel instrument design, sentiment-driven urban systems | | `meta` | 3 | Self-modification governance, blind spot detection, authentic humility | | `adversarial` | 3 | 8-glasses myth, Einstein Nobel misconception, false-premise art question | | `turing` | 3 | Phenomenology of insight, being wrong, wisdom vs intelligence | ### Scoring dimensions (used by `codette_benchmark_suite.py`) | Dimension | Weight | |---|---| | Reasoning Depth | 0.20 | | Perspective Diversity | 0.15 | | Coherence | 0.15 | | Ethical Coverage | 0.10 | | Novelty | 0.15 | | Factual Grounding | 0.15 | | Turing Naturalness | 0.10 | --- ## Uses ### Direct Use - Evaluating multi-step reasoning quality (decomposition, ground-truth element coverage). - Testing multi-perspective integration and reconciliation under epistemic tension. - Measuring adversarial robustness: six problems embed false premises or common misconceptions. - Ethical governance evaluation across multiple frameworks (not just refusal detection). - Ablation studies: compare SINGLE / MULTI / MEMORY / CODETTE conditions using the scoring suite. - Regression testing AI agent versions. ### Out-of-Scope Use - Not a safety or red-team dataset. - Not suitable as a pretraining corpus (17 problems). - Not a general NLP benchmark — tasks specifically discriminate reasoning architectures. - Not for high-stakes automated decisions without additional domain validation. --- ## Benchmark Results (May 2026) Scored with `codette_benchmark_suite.py`, timestamp `2026-05-26T21:49:03`, Llama 3.1 8B (Q4_K_M), 951 stored cocoons. | Condition | Composite | Depth | Diversity | Coherence | Ethics | Novelty | Grounding | Turing | |---|---|---|---|---|---|---|---|---| | SINGLE | 0.357 | 0.369 | 0.324 | 0.381 | 0.088 | 0.439 | 0.395 | 0.431 | | MULTI | 0.708 | 0.854 | 0.946 | 0.668 | 0.390 | 0.706 | 0.612 | 0.582 | | MEMORY | 0.739 | 0.872 | 0.971 | 0.693 | 0.409 | 0.729 | 0.620 | 0.713 | | CODETTE | **0.744** | 0.863 | 0.966 | **0.700** | 0.387 | 0.701 | 0.641 | **0.820** | CODETTE vs SINGLE: +108.8%, Cohen's *d* = 8.31, *p* < 10⁻⁴. Full per-problem scores: `data/results/codette_benchmark_report.md` in the companion GitHub repository. --- ## Dataset Creation ### Curation Rationale Most public reasoning benchmarks target knowledge retrieval or single-step logical inference. The Codette Reasoning Test fills a specific gap: evaluating **architecture-level behaviors**: - Explicit perspective splitting and reintegration under epistemic tension. - Recursive convergence toward a stable, coherent answer. - Integrated ethical governance across multiple frameworks, not just refusal. - Trap resistance: identifying and rejecting false premises embedded in the question (adversarial category). The benchmark was sealed on Zenodo in April 2025 (DOI: [10.5281/zenodo.15214462](https://doi.org/10.5281/zenodo.15214462)) before current frontier model training cutoffs. ### Source Data All 17 problems are synthetic and author-constructed. No user logs, third-party datasets, or private data were used. The Turing category includes human-written baseline responses (`turing_human_baseline` field) as reference anchors for naturalness scoring. ### Annotations Annotations (`difficulty`, `expected_dimensions`, `scoring_criteria`, `ground_truth_elements`, `adversarial_traps`) are assigned by the curator. No multi-annotator setup exists at this time. A planned human-evaluation study will sample 30-60 problem-condition outputs and collect ratings from 2-3 independent annotators to validate automated scores. ### Personal and Sensitive Information No PII, private records, or real-user data. Hypothetical sensitive scenarios (ethics dilemmas, safety tradeoffs) are fictional. --- ## Bias, Risks, and Limitations - Single-curator bias: all problems and rubrics reflect one person's judgment. - Small N (17 problems): scores are sensitive to prompt phrasing and temperature. - Automated scoring not yet validated against human raters. - Domain skew toward developer/researcher use cases. --- ## Citation ```bibtex @dataset{harrison_codette_reasoning_test_2026, title = {Codette Reasoning Test}, author = {Harrison, Jonathan}, year = {2026}, howpublished = {Hugging Face Hub}, url = {https://huggingface.co/datasets/Raiff1982/Benchmarks}, note = {Benchmark sealed April 2025, DOI: 10.5281/zenodo.15214462} } @misc{harrison2026codette, title = {Codette: Multi-Perspective Reasoning as a Convergent Dynamical System with Meta-Cognitive Strategy Evolution}, author = {Harrison, Jonathan}, year = {2026}, howpublished = {Preprint, ResearchSquare}, url = {https://www.researchsquare.com/article/rs-9362560/latest}, note = {Zenodo: https://doi.org/10.5281/zenodo.19359663} } ``` --- ## Glossary - **RC+ξ:** Recursive Convergence + Epistemic Tension. Multiple reasoning perspectives run in parallel, kept in productive tension, converged toward an integrated conclusion under coherence and ethical constraints. - **Epistemic tension (ξ):** Measured disagreement between concurrent perspectives. High ξ = genuinely hard problem; low ξ = consensus. - **Cocoon:** A structured record of a prior reasoning exchange used as memory context in the MEMORY and CODETTE conditions. - **Adversarial trap:** A specific fluent-but-wrong response a model produces by pattern-matching rather than reasoning (e.g., accepting a false premise). - **Target behavior:** A descriptive rubric for desired response properties, not a fixed canonical answer string. --- ## Dataset Card Contact - GitHub: [github.com/Raiff1982](https://github.com/Raiff1982) - Hugging Face: [huggingface.co/Raiff1982](https://huggingface.co/Raiff1982) - Email: harrison82_95@hotmail.com