Protocols over Platforms
We believe that for AI to be safe, its governance must be public, peer-reviewed, and mathematically verifiable.
SCD Protocol v3.1
Structured Contextual Distillation
A deterministic, vendor-independent protocol for persistent, verifiable agent state. Employs RFC 8785 JSON canonicalization, SHA-256 integrity chains, and constitutional governance that cannot be overridden by prompt-level instructions.
Validated: 100% determinism over 1,005 state transitions. Cross-vendor continuity tested.
Desai, P. (2025). Structured Contextual Distillation (SCD v3.1): A Deterministic, Vendor-Independent Protocol for Persistent, Verifiable Agent State. Zenodo.
DOI: 10.5281/zenodo.17787619
Publications & Preprints
Governance and Boundary Conditions for Reflective AI Systems
Paul Desai · December 2025 · Preprint
Structural enforcement beyond prompt alignment. Demonstrates that governance constraints for reflective AI systems can be enforced through external wrapper architectures, deterministic state protocols, and cryptographic integrity verification.
Layered Governance for Large Language Model Systems
Paul Desai · December 2025 · Preprint
Separating structural authority enforcement from content safety. Empirical evaluation across 130 test cases shows governance-only layers achieve 70–100% refusal accuracy on authority attacks with 0% false positive rate.
Our Validation Philosophy
1. Public Proof
All safety claims must be backed by reproducible code and public datasets. No "black box" safety.
2. Adversarial Design
Systems are secure only when they survive dedicated red-teaming. Our "M1 Adversary" node is proof of this commitment.
Cite This Work
@article{desai2025scd,
title={Structured Contextual Distillation (SCD v3.1)},
author={Desai, Paul},
year={2025},
doi={10.5281/zenodo.17787619}
}
@article{desai2025governance,
title={Governance and Boundary Conditions for Reflective AI Systems},
author={Desai, Paul},
year={2025},
doi={10.5281/zenodo.18212080}
}
@article{desai2025layered,
title={Layered Governance for Large Language Model Systems},
author={Desai, Paul},
year={2025},
doi={10.5281/zenodo.18212082}
}