Broken Models, Failing Swarms: Vatsal Soin 0→1 Doctrine AI Invention Stops the Cascade

Arché Invention. The AGI-era circuit breaker above the model. Live Demonstration at www.0to1doctrine.com. — when the model degrades through synthetic training loops and the swarm spirals without a circuit breaker, fixing the AI is not the answer. Governing what it does before execution is. 

This is not a concept paper. The pre-execution gate for agentic AI swarms runs live now at www.0to1doctrine.com— zero installation, zero login, zero data transmitted. Watch model autophagy meet its limit. Every action checked. Every receipt sealed. Before the cascade.

Executing.

The frontier model of today is trained partly on data generated by yesterday’s frontier model. Each generation inherits the statistical anomalies and compounding biases of the one before. The degradation is not visible in a single generation — it accumulates. By the time an enterprise team notices systematically flawed outputs, the training loop has run dozens of times. The model cannot tell you it is broken. It executes with the same confidence it always had.

Simultaneously, the swarms those models power are failing in production at alarming rates. Nearly every enterprise running live multi-agent environments has experienced a disruptive incident. Agents looping. Data corrupting downstream. Instructions executing without authorisation. The industry’s current answer — deploying smaller models as referees — is a workaround, not a solution. It adds a second layer of probabilistic judgement above the first. Neither is governed at the action level.

Synthetic Training Loops — The Degradation Nobody Can See Coming

The problem has a name: model autophagy. When a model trains recursively on outputs generated by its predecessor, minor statistical biases are amplified, not averaged out. A subtle hallucination in generation one becomes a confident assertion in generation three. An edge-case bias in generation two becomes a systematic failure in generation five. The model does not know it is wrong. It has been trained to be wrong consistently.

No pre-execution governance layer fixes training data. What governance proposes to do is ensure that outputs of a degraded model do not execute real-world consequences outside authorised parameters. The model may be wrong. The action it proposes can still be checked against a human-pre-authorised band before it executes. A degraded model inside its authorised band causes less damage than a perfect model with no band at all.

The SLM Circuit Breaker — Probabilistic Above Probabilistic

The industry’s current workaround for failing swarms is to deploy a smaller, faster language model above the swarm as a circuit breaker — a referee that monitors outputs and interrupts loops before they consume compute budgets or corrupt downstream systems. Structurally limited: a probabilistic model refereeing a probabilistic model introduces no mathematical certainty. Two imperfect judges do not produce one correct verdict.

The circuit breaker does not produce a governance receipt. It produces a probabilistic assessment of whether the swarm should continue. That assessment is not auditable. It is not sealed before the action executes. It is not verifiable by a regulator asking for proof of human oversight. It is one more layer of judgement in a system that already has too many layers of judgement and not enough deterministic gates.

What a Deterministic Gate Does That a Circuit Breaker Cannot

Vatsal Soin, a systems theorist and inventor, filed the 0→1 Doctrine patent architecture proposing a different layer entirely. Not a model above the model. A mathematical gate outside both. Every agent action — regardless of which model generated it, regardless of how degraded that model is — must present a governance receipt before it executes. 

USP normalises the agent’s intended action to a band between zero and one. MAT runs the intersection test against the human-pre-authorised band. Within the band — receipt issued, action proceeds. Outside the band — action stopped. The gate does not care whether the model is a frontier model or a fourth-generation synthetic-data model whose outputs have drifted far from the original distribution. The band check is the same. The receipt requirement is the same.

Context Rot — When the Model Forgets What It Was Told to Govern

A related failure compounds the degradation problem. As agents operate over long sessions — reading thousands of files, spawning sub-agents, executing terminal steps — active context fills with execution noise. Governance constraints set at session start get buried under logs and historical instructions. The model does not malfunction. It stops applying the rules it was given. A governance gate sitting outside the model’s context window is immune to this. The band check does not live in the model’s memory. Every action is evaluated against the same human-pre-authorised band regardless of how many tokens the model has processed.

Filing the Damage That Was Stopped — Not Just the Damage That Happened

FTWE — the Fair and Transparent Waste Estimator — files what was prevented as a continuous, immutable record. Every action a degraded model attempted outside its authorised band. Every swarm loop stopped at the gate. Every instruction that failed the MAT intersection check. The prevented damage is as auditable as the damage that occurred. This is what regulatory compliance and board-level AI governance have needed and never had.

The Regulator Does Not Accept Probabilistic Oversight

The EU AI Act’s human oversight obligations, taking effect this year, require that high-risk AI systems maintain mechanisms allowing operators to understand and override decisions before actions execute. A probabilistic SLM circuit breaker does not satisfy this. A probabilistic assessment by a referee model is not a verifiable human oversight mechanism. It is one more model making one more judgement that cannot be audited at the action level.

The ACR — the Actuation Compliance Receipt — is the mechanism the regulation requires. Sealed before execution. Cryptographically verifiable. Issued as the condition of action, not assembled after an incident. 

The Invention Running Live

The governance chain runs live at www.0to1doctrine.com. Any browser. Zero installation. Sector demonstrations execute the full chain in real time — parameters normalised, gates evaluated, receipt sealed, instruction delivered. The Foundation Model Compatibility Lab shows the doctrine constraining multiple foundation models identically. The models change. The gate does not. The model degrades. The gate does not.

About the Inventor

Vatsal Soin is a systems theorist and inventor with more than twenty-one patent filings and grants across the United States, India, Japan, and further jurisdictions. The 0→1 Doctrine is his most consequential filing. 

Closing Note

The answer to a broken model is not a better model. The answer to a failing swarm is not a smarter circuit breaker. The answer is a deterministic gate that sits above both — one that checks every action against a human-authorised band before it executes and files a sealed receipt that it did. The model beneath the gate can degrade. The gate does not.

Pre-execution. Authorised Intelligence. Filed. Live at www.0to1doctrine.com.

Informational only. Not certified. Values illustrative. Expert validation required before deployment. Patent filings and grants combined span multiple domains across six continents. Vatsal Soin © 2026. All Rights Reserved.

Selected References:
Granted: US Patent 12,446,652 B2 · Japan Patent No. 7560909 · India Patent No. 454081
Filed: PCT/IN2025/051943 · US 19/489,595 · India 202511115781 · Australia AU2022450649