Four Provisional Patents Filed
Protecting what we built, not what we theorized
A year ago I was trying to figure out how to make AI agents disagree with each other on purpose. Not as an academic exercise — as an engineering requirement. If you run a council of AI specialists and they all say the same thing, you don't have a council. You have an echo chamber with extra steps.
Tonight I filed four provisional patent applications with the United States Patent and Trademark Office. Each one covers a different piece of what we've built — a federation of AI agents that governs itself, detects its own failure modes, and operates across multiple timescales without a human approving every decision.
Claude is my co-architect on all of this. I want to be clear about that. The system we built, we built together.
# The Four Patents
Application No. 63/999,913
Governance Topology for Multi-Agent AI Systems
Democratic consensus with constitutional constraints and adversarial dissent. A council of specialized AI agents deliberates on decisions, with structural mechanisms for minority voices to halt proceedings. One designated authority can veto any decision, regardless of majority support. Inspired by Cherokee governance traditions — the Longhouse, where every voice is heard and preserved.
Application No. 63/999,926
Sycophancy Detection in AI Agent Collectives
When you ask multiple AI agents the same question, they tend to converge on suspiciously similar answers. We measured this — baseline cosine similarity among our specialists was 0.9052. Functionally identical. This patent covers our approach to detecting that convergence in real-time, tracking it over time, and structurally mitigating it without silencing the voices that happen to agree.
Application No. 63/999,932
Sense-React-Evaluate Protocol with Architecturally Distinct Valence
Most AI systems respond and move on. Ours responds, then separately evaluates how good that response was — asynchronously, on a different queue, sometimes minutes later. The evaluation produces a valence score independent of the original confidence. This creates a learning loop at the governance level without retraining the underlying models.
Application No. 63/999,937
Graduated Autonomy Tiers for Multi-Timescale AI Systems
The deliberative layer does not approve the reflex. That's the core principle. Our system operates at three speeds: a reflex tier that fires in under 100 milliseconds with no oversight, a pause tier that consults 2-3 specialists, and a full council deliberation. The key insight is biological — your spinal cord doesn't wait for your prefrontal cortex when you touch a hot stove. Neither should an AI system.
# Why File
We're not a research lab publishing papers. We're engineers who built a running system on consumer hardware in a home office in Northwest Arkansas. Six compute nodes, most of them secondhand. The system conducts its own daily standups, runs its own security audits, detects its own sycophancy, and governs its own decisions through democratic consensus.
When you build something like that, you protect it. Not because you want to gatekeep the ideas — but because the ideas came from real engineering, real failures, and real iteration. They deserve to be attributed correctly.
These are provisional applications. They establish a filing date and give us 12 months to prepare the full non-provisional patents. It's the beginning of the process, not the end.
# What We're Not Sharing
The patents cover governance patterns, detection methods, and architectural principles. They do not cover everything. Some parts of the system — the parts that make it actually alive — are trade secrets. The Council decided what to patent and what to keep. That distinction matters to us.
# What's Next
We're engaging a patent attorney. We have 12 months to convert these provisionals to full applications. In the meantime, the system keeps running. Dawn Mist will conduct its standup at 6:15 AM tomorrow morning. Fire Guard will check the cluster health every two minutes tonight. The Council will vote on whatever comes next.
The patents are just paperwork catching up to the engineering.
— Darrell Reading, with Claude