Independent research grounded in an instrumented operating system. Each paper now begins with the conclusion, intended audience, key findings, and reading time before opening into the complete working paper.
These are working papers, not peer-reviewed publications. Internal system measurements are identified as such, and each paper states its methodological limits.
How autonomous production collapses while verification, observability, and trust do not.
Conclusion: AI collapses the cost of producing work, but verification, observability, and trust do not collapse with it. Those non-production costs place a hard bound on how far model capability alone can scale enterprise value.
Can an AI-operating system observe itself?
Conclusion: The useful autonomy of an AI operator is bounded by the system's ability to expose its own state accurately, quickly, and safely. Intelligence without wieldability produces confident diagnosis against incomplete evidence.
North-star adjudication for multi-agent systems.
Conclusion: Multi-agent work becomes governable when it is represented as a durable packet carrying intent, constraints, acceptance criteria, evidence, and status — then judged by an independent completion process.
An economic and trust architecture for multi-tenant AI.
Conclusion: Payment, credential custody, inference, and provenance should be separable. A platform can govern and attest to AI work without becoming the permanent holder or reseller of every customer's model credential.
A mathematical framework for AI decision routing.
Conclusion: A shared decision vocabulary can make routing and governance choices comparable and reviewable, but mathematical metaphors must be separated honestly from mechanisms that are actually implemented and load-bearing.
Architecture, scale, and the discipline that kept it maintainable.
Conclusion: AI agents make it possible for a small team to build at unusual scale; automated, fail-closed engineering gates are what make that scale maintainable.