Experiment code, ECM manifests, and ground truth oracle for the paper:
Xue Qin, Simin Luan, John See, Cong Yang, Zhijun Li. "ECM Contracts: Contract-Aware, Versioned, and Governable Capability Interfaces for Embodied Agents." 2026. arXiv: 2604.13097
This repository evaluates a six-dimensional contract model for Embodied Capability Modules (ECMs): functional signature, behavioral assumptions, resource requirements, permission boundaries, recovery semantics, and version compatibility.
| Experiment | Description |
|---|---|
| E1: Composition Stability | Contract-checked vs ad hoc vs schema-only chain composition (N=500) |
| E2: Upgrade Safety | Contract-aware vs semver-only upgrade decisions (24 upgrade events) |
| E3: Long-Horizon Tasks | Multi-step task execution with embedded incompatibilities |
| Ablation | Per-dimension ablation over all six contract dimensions |
41 pairwise incompatibility entries across 24 ECM manifests, covering 6 contract dimensions. Only 5 are detectable by schema-level type checking; the remaining 36 require contract-level analysis.
Python >= 3.11
No external dependencies required.
Run individual experiments:
cd experiments
python exp1_composition.py
python exp2_upgrade.py
python exp3_longchain.py
python ablation.py| Method | Accepted | Failures | Success Rate |
|---|---|---|---|
| Ad hoc | 500 | 369 | 26.2% |
| Schema-only | 299 | 173 | 65.4% |
| Contract | 100 | 9 | 98.2% |
ecm-contracts/
├── experiments/ # Experiment scripts
│ ├── contract_checker.py # Six-dimension contract checker
│ ├── exp1_composition.py # Composition stability (E1)
│ ├── exp2_upgrade.py # Upgrade safety (E2)
│ ├── exp3_longchain.py # Long-horizon tasks (E3)
│ └── ablation.py # Per-dimension ablation
├── manifests/ # 24 ECM manifest files (YAML)
├── examples/ # Example ECM definitions
└── incompatibility_matrix.json # Ground truth oracle (41 entries)
Licensed under the Apache License, Version 2.0. See LICENSE for details.
@article{qin2026contracts,
title={ECM Contracts: Contract-Aware, Versioned, and Governable Capability Interfaces for Embodied Agents},
author={Qin, Xue and Luan, Simin and See, John and Yang, Cong and Li, Zhijun},
journal={arXiv preprint arXiv:2604.13097},
year={2026}
}