| --- |
| pretty_name: StableEval Arena |
| license: apache-2.0 |
| language: |
| - en |
| task_categories: |
| - tabular-classification |
| - time-series-forecasting |
| tags: |
| - agentic-ai |
| - evaluation |
| - trustworthiness |
| - stablecoin |
| - benchmark |
| size_categories: |
| - n<1K |
| configs: |
| - config_name: stableeval_120_enriched_cases |
| data_files: |
| - split: train |
| path: data/stableeval_120_enriched/case_list.csv |
| - config_name: stableeval_120_enriched_labels |
| data_files: |
| - split: train |
| path: data/stableeval_120_enriched/labels.csv |
| - config_name: stableeval_507_natural_cases |
| data_files: |
| - split: train |
| path: data/stableeval_507_natural/case_list.csv |
| - config_name: stableeval_507_natural_labels |
| data_files: |
| - split: train |
| path: data/stableeval_507_natural/labels.csv |
| - config_name: stableeval_507_natural_leaderboard |
| data_files: |
| - split: train |
| path: results/stableeval_507_natural/final_tables/stableeval_507_natural_v6_leaderboard.csv |
| --- |
| |
| # StableEval Arena |
|
|
| StableEval Arena is a historical-replay benchmark for evaluating cost-aware, trustworthy agentic AI systems on stablecoin risk prediction. Each case gives a frozen 30-day lookback packet for USDT, USDC, or DAI and hides the next 7-day horizon labels. |
|
|
| GitHub repository: https://github.com/SeanWan514/StableEval-Arena |
|
|
| Hugging Face dataset: https://huggingface.co/datasets/SeanWan05/stableeval-arena |
|
|
| ## Benchmark Blocks |
|
|
| | Block | Role | Contents | |
| |---|---|---| |
| | `StableEval-120-Enriched` | Stress-enriched rare-risk diagnostic | 120 cases, six systems plus baselines, final tables, ablations | |
| | `StableEval-507-Natural` | Natural-distribution robustness/scaling | 507 cases, baselines, four native/framework-level agents, two adapter-based `-Claw` extensions | |
|
|
| ## Task |
|
|
| Inputs are lookback-only prompt packets. The hidden horizon is used only for labels and evaluation. |
|
|
| - Lookback: frozen 30-day historical window. |
| - Horizon: hidden next 7 days. |
| - Classification target: `stable`, `watch`, or `stress`. |
| - Auxiliary targets: depeg probability and maximum absolute dollar-peg deviation in basis points. |
|
|
| ## Dataset Contents |
|
|
| ```text |
| data/stableeval_120_enriched/ |
| data/stableeval_507_natural/ |
| results/stableeval_120_enriched/ |
| results/stableeval_507_natural/ |
| schemas/ |
| docs/ |
| ``` |
|
|
| The package includes prompt packets, labels, canonical result tables, raw/calibrated prediction JSONL files, execution metadata, ablation summaries, stress/audit summaries, checksums, and schemas. It excludes raw vendor snapshots, full hourly-window dumps, attempt-level logs, dry-run outputs, and superseded v2-v5 progress leaderboards. |
|
|
| ## Label Distributions |
|
|
| - `StableEval-120-Enriched`: {'stable': 54, 'stress': 21, 'watch': 45} |
| - `StableEval-507-Natural`: {'stable': 441, 'stress': 21, 'watch': 45} |
|
|
| ## Systems Evaluated |
|
|
| Baselines: `always_stable`, `historical_persistence`, `threshold_rule`, `simple_ml_softmax_logistic`. |
|
|
| Native/framework-level agents: `nanobot_qwen_507`, `langgraph_qwen_507`, `hermes_ai_qwen_507`, `evoagentx_qwen_507`. |
|
|
| Adapter-based `-Claw` extensions: `openclaw_adapter_qwen_507`, `nanoclaw_adapter_qwen_507`. |
|
|
| OpenClaw/NanoClaw native CMDOP execution was not verified in this release; cite those systems as adapter-based extensions. |
|
|
| ## Main Result Files |
|
|
| - 120 main leaderboard: `results/stableeval_120_enriched/final_tables/validation_120_full_leaderboard.csv` |
| - 120 metrics: `results/stableeval_120_enriched/metrics/prediction_quality_by_platform.csv` |
| - 507 v6 leaderboard: `results/stableeval_507_natural/final_tables/stableeval_507_natural_v6_leaderboard.csv` |
| - 507 stress summary: `results/stableeval_507_natural/stress_audits/stress_recall_summary_from_v6.csv` |
| - 507 cost/reliability summary: `results/stableeval_507_natural/cost_logs/cost_reliability_summary_from_v6.csv` |
|
|
| ## Recommended Use |
|
|
| - Reproduce paper-facing saved-output evaluations. |
| - Compare new agents against the released cases and labels. |
| - Study cost/reliability tradeoffs and structured-output validity. |
| - Audit stress recall and false alarms under a common benchmark protocol. |
|
|
| This dataset is not validated for live trading, financial advice, regulatory compliance, or automated treasury execution. |
|
|
| ## Limitations |
|
|
| - Stress cases are rare, especially in the natural 507-case distribution. |
| - The 120-case block is intentionally stress-enriched and is not a deployment prevalence estimate. |
| - All released agent results use a common OpenRouter-hosted Qwen backend; this isolates platform/workflow differences but is not a broad foundation-model comparison. |
| - Prompt packets and case metadata contain processed/derived market data and require redistribution review. |
| - Current agents do not solve stablecoin stress prediction; weak stress recall is a central trustworthiness finding. |
|
|
| ## Citation |
|
|
| See `CITATION.cff`. No DOI is assigned in this local package; add one only after a public archival DOI exists. |
|
|
| ## License And Redistribution |
|
|
| See `LICENSE` and `docs/redistribution_notice.md`. Raw vendor snapshots are not included. Processed prompt packets, case metadata, and labels are included for academic reproducibility but may remain subject to upstream provider terms. |
|
|
