NeoSyntropyDeterministic runtime engine for LLMs
NeoSyntropy is a runtime engine that controls language models (LMs) through finite state machines. Every action your AI takes follows a path you defined: 0% deviation from your business logic, enforced at the engine level, not begged for in a prompt.
# No prompt chaining. Just state control.
engine.state("intake")
.on("validated")
.transition("execute")
.guarantee(deterministic=True)transition verified · O(1) · deviation 0%The paradigm shift
Autonomous agents improvise. When AI drives real business outcomes, improvisation is a liability. A finite state machine is not an academic exercise here: it is the control layer that forces the AI to operate only inside states and transitions you predefined, so every decision is explainable, auditable, and compliant by construction.
Agent frameworks
Nobody can replay why the agent chose a path. Every run is a new roll of the dice.
The model can improvise steps you never approved, with real side effects.
Audits turn into log archaeology because there is no defined path to check against.
NeoSyntropy
Every workflow is an explicit finite state machine. The model can only move along edges you defined.
The engine blocks any move outside the graph before a side effect can happen. 0% deviation.
Each step is a logged, auditable state change. Reviews check a graph, not a transcript.
The philosophy behind the engine
"Beginning with the simplest objects ... rising, step by step, to the knowledge of the most complex."
How it works
You define states, executable nodes, rules, and allowed transitions. The router chooses the work; the control layer decides whether one next-state transition is valid.
Define the webhooks and events that feed the workflow. Each incoming payload becomes the context your states operate on.
webhook inputBreak your workflow into explicit steps. Each state is a named checkpoint the AI must reach: intake, review, approval, execution.
statesEvery state declares the rules that must hold before it counts: required data, approvals, guardrails. If an axiom breaks, the step is rejected.
axiomsConnect states with the only transitions you allow. The AI can never invent a path: if an edge is not defined, it does not exist.
edgesAt runtime, the router selects one or more executable nodes. After they run, the control layer validates and commits one permitted state transition. Every decision is logged.
controlInput data becomes runtime context, selected nodes perform the work, and only one verified state transition is committed.
Unit economics
Replace opaque token spend with one measurable unit your finance, product, and engineering teams can all agree on: the successful state transition.
THE OLD TOKEN APPROACH
Every step re-reads the entire history, so token costs grow exponentially with workflow length.
Retries, hallucinations, and self-correction loops all land on your invoice as billable tokens.
Usage-based token spend is a number finance discovers at month end, not one it can forecast.
OUR TRANSITION APPROACH
Rates reach as low as $0.002 per successful state transition. The per-transition rate never changes, no matter how many steps the operation contains.
Cost scales linearly with business volume, so AI spend becomes a line item finance can model.
Universal knowledge is distributed across the states, so history is never resent and failed or hallucinated steps are never billable events.
Transitions, sessions, memory, knowledge, and traces are available the moment you connect an agent. Meter only verified outcomes, then improve the path with coding agents.
Stateless runtime, per-session isolation, JWT-based RBAC, and structured tracing. These are not abstractions you want to debug at 2am when an agent loop goes sideways.
Every workflow declares the invariants that must hold for a transition to count. If an axiom fails: wrong state, missing approval, failed guardrail, or invalid ledger write, the step does not bill and the system stops before bad logic compounds.
Guardrails, human-in-the-loop, and admin approval flows sit inside the runtime. Audit logs and post-execution hooks watch sensitive actions so business logic stays accountable.
One economic primitive
Model agent behavior as explicit states, ship deterministic workflows, and pay only when useful work completes.
Start buildingForecast AI spend by business volume, eliminate runaway token bills, and prove ROI at the unit-economics level.
Talk to SalesThe token era is ending