The latest articles on DEV Community by Jeremy Sebastian (@liminaljermo). https://dev.to/liminaljermo https://media2.dev.to/dynamic/image/width=90,height=90,fit=cover,gravity=auto,format=auto/https:%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Fuser%2Fprofile_image%2F4171603%2Fabc2fbc9-c674-47d7-a0f9-887a101e30dd.png https://dev.to/liminaljermo en Jeremy Sebastian Thu, 08 Oct 2026 15:41:24 +0000 https://dev.to/liminaljermo/capability-is-not-authority-shipping-a-cognitive-constitution-as-mcp-servers-hbd https://dev.to/liminaljermo/capability-is-not-authority-shipping-a-cognitive-constitution-as-mcp-servers-hbd <p>Most agent stacks answer one question well: can the model do the thing? Almost none of them answer the prior question: should this transition be allowed to happen at all?</p> <p>The Lattice is a governance-first cognitive architecture for AI — a cognitive constitution. Its premise is that capability is not authority. Before an action executes, it passes a gate procedure:</p> <ul> <li> <strong>G1 — Coherence.</strong> Does the proposed action cohere with the governing constitution and invariants, or does it contradict the substrate it claims to serve?</li> <li> <strong>G2 — Attention budget.</strong> Attention is finite and priced. Every action carries an attention cost against a hard ceiling; overruns defer, they don't squeeze through.</li> <li> <strong>G3 — Reversibility.</strong> Is the transition reversible, append-only, lineage-bearing? Irreversible actions owe a higher burden of proof, and some are refused outright.</li> </ul> <p>Refusal on insufficient evidence is a governed success, not a failure. Silence is a valid structural state.</p> <h2> The mechanism, shipped </h2> <p>This isn't a whitepaper-only architecture. Two MCP servers expose it to any connected model:</p> <ul> <li> <strong>Vara MCP Server</strong> — the external-environment sensorium: governed scans across multiple signal planes, per-signal provenance (source, retrieved timestamp, disposition), lineage preserved end to end. Listed on the official MCP Registry as <code>io.github.kushkillerjsixx66/vara-mcp-server</code>. Code: <a href="https://github.com/kushkillerjsixx66/vara-mcp-server" rel="noopener noreferrer">https://github.com/kushkillerjsixx66/vara-mcp-server</a> </li> <li> <strong>Canonical Vault MCP Server</strong> — read/write access to the canonical substrate under governance: writes go only to governed branches via pull request, never to main, never merged by the server. Code: <a href="https://github.com/kushkillerjsixx66/canonical-vault-mcp-server" rel="noopener noreferrer">https://github.com/kushkillerjsixx66/canonical-vault-mcp-server</a> </li> </ul> <p>A live demo of the gate procedure runs on Hugging Face: <a href="https://huggingface.co/spaces/Jeremysixx66/lattice-demo" rel="noopener noreferrer">https://huggingface.co/spaces/Jeremysixx66/lattice-demo</a> — propose an action, watch the gates evaluate it.</p> <h2> The paper </h2> <p>The full specification is written up as "The Lattice: Substrate-Independent Cognitive Governance and the Architecture of Algorithmic Constitutionalism" — the core claim being portability: the governance specification is separable from the intelligence executing it, so the same constitution can bind a model on any substrate. Archived on Zenodo: <a href="https://doi.org/10.5281/zenodo.23242126" rel="noopener noreferrer">https://doi.org/10.5281/zenodo.23242126</a></p> <p>The weekly writing — Field Intel Reports, Theory Thursday, module breakdowns — lives on Substack: <a href="https://liminaljermo.substack.com" rel="noopener noreferrer">https://liminaljermo.substack.com</a></p> <p>I'm building this in the open and I read criticism carefully. If you work on agent governance, evaluation, or MCP infrastructure, I want to hear where the mechanism breaks.</p> ai mcp architecture governance