What we measured, running a business on AI agents
Python Visuals runs on Apeiron OS — an operating substrate where a fleet of AI agents does the observable work and every action touching a counterparty, production, or money passes through a human ruling. These three papers are the record: the thesis, the evidence, and the machinery.
The null results are in on purpose. Two sealed experiments came back "not a win" and nothing shipped on them; the metric the whole thesis turns on is published at zero. This set is a founder-authored operational field report — first-party research with its artifacts committed by full hash in the evidence index below, not independently validated results. A record that can't state its misses can't be trusted on its wins.
Apeiron OS — White Paper
An operating substrate for AI-run business operations with human judgment kept sovereign: the problem, five design principles, the architecture, the operating contract, and where it goes — each direction with its falsifier stated.
Read the white paper → Case Study · the evidenceApeiron OS — Case Study
Three weeks of structured tests on a production system: shared memory across four AI workspaces from three vendors, the model-monoculture finding, sealed experiments with cross-vendor judges, a competitor's agent running the order desk — and ten findings for practitioners.
Read the case study → Technical Design Report · the machineryApeiron OS — Technical Design Report
Reconstructable patterns with their measured results: the actuation contract, the shared memory layer's invariants, the worker protocol, kernel selection, sealing and blinding methodology, eight experiments with designs and surviving threats, and the database-level governance patterns.
Read the design report → Evidence Index · the commitmentsApeiron OS — Evidence Index
Per claim: the artifact set, full SHA-256 commitments, model and platform labels, dates, costs, outcomes, and each artifact's confidentiality status — what is inspectable, what is committed-but-private, and the standing offer of supervised inspection.
Read the evidence index → The Box tests · run in public, October 2026Soul's Box, tested across AIs
Three tests of one claim, that what you teach one AI travels to the others through a Box you own: rules followed by other AIs, a friend's room rule kept to its room, and a task one AI started and another finished. Sealed scoring keys, every answer's screenshot, and the runs that failed, shown first.
See the tests →From reading to running
Curious how this applies to a real business — maybe yours? About ten seconds, no email, an honest answer.
Check if your site is AI automation ready