--- title: Deterministic vs Probabilistic AI, per Step | Torke url: https://torke.tech/engine/router description: Torke Router scores every campaign step on five written questions and routes it to code, a model, a hybrid or a person, with its test and fallback. --- # Decide which campaign steps AI runs Torke Router is a Torke Node, one part of Torke Engine, and you can buy or rent it on its own. Before anything is built, it takes your campaign apart into steps and says, for each one, whether it should run as code, as a model, as a mix of both or as a person's decision, and how that choice will be tested. Diagram: How Torke Router works Torke Router runs in steps: you state the goal or your own workflow, each step is scored, code assigns the verdict, each step gets its specification. ## What is the difference between deterministic and probabilistic AI? A deterministic step gives the same answer every time for the same input, because it follows a written rule in code. A probabilistic step, such as a language model, estimates the most likely answer, so the same input can produce different outputs. Setting a model's temperature to 0 lowers that variation without removing it, which is why Torke Router routes each step on written criteria and never on temperature. ## Which steps should a model never touch? Every step in a campaign that uses AI is done by something: code that follows a fixed rule, a model that reads and writes language, a mix of the two, or a person who decides. The costly mistakes come from putting a model on a step that has a rule. A model asked whether a person may legally be emailed will usually answer correctly, so the mistake passes every demo, and the wrong answers surface weeks later as a complaint or a lead count your client disputes. Torke Router makes that placement visible before you spend anything, by scoring each step against written criteria instead of intuition. The example built into the node is a campaign to generate 40 qualified leads from a market of 400 companies, using Torke's method of contacting a company in the week something changed for it. That campaign has 21 steps, and the Router sends 11 of them to plain code, 6 to a mix of code and model, 2 to a model alone and 2 to a person. Two steps from that plan show the difference. Confirming the legal basis for contacting someone can be written as a lookup in a table of countries and channels, and a mistake would be caught by a regulator or a complaint, so it scores the maximum and goes to code, and a country missing from the table stops the contact and goes to an operator. Deciding whether a news article describes something a company did, or a market-wide event it was only exposed to, has no complete rule, so a model reads the article but has to quote the exact sentence naming the company, and code checks that quote word for word. If the check fails, the item is stored as market news and can never open an email. The Router plans and specifies, and it does not build or run the steps. What changes for you is the conversation after launch: when a client asks why a model drafted the message but code checked the consent, you have a scored, dated answer to show them. ## How does Torke Router work? Each step says who does it. Anything with a written rule runs in code, a model works only where judgment on free text is needed and always under a code check, and a person decides whatever someone must answer for. ### You state the goal or your own workflow Who does it Person What happens Start from one of four blueprints, a signal-led lead campaign, inbound lead response, first-touch outreach or a campaign on approved claims, or list your own steps with the init command and the Router builds an empty plan around them. ### Each step is scored Who does it Model What happens Five questions, scored 0, 1 or 2: can the rule be written down, how tidy is what arrives, must the same input give the same answer, what does a wrong answer cost and who catches it, and could a client or regulator ask us to explain it. Each score carries a sentence of reasoning. ### Code assigns the verdict Who does it Code What happens A step whose rule exists in full runs as code, whatever else it scores. A high cost of error or a need to explain the answer rules out a model working alone, and a verdict can only change through a score with a written reason. ### Each step gets its specification Who does it Code What happens Code steps get their rule as a sentence a developer can implement and unit tests at a 100 percent pass bar. Model steps get the model, the instructions, the answer format, an eval with a pass bar and a fallback that is a rule or a person. ### Every model call gets its controls Who does it Code What happens Outputs are checked against their schema, quotes are matched character for character against the input, labels come from a closed list, and each call has a token cap, at most one repair attempt and a full run log. ### The goal is worked backwards Who does it Code What happens From the target back to what it needs, using the funnel each blueprint declares, such as companies, contacts and replies for a lead campaign. You supply each rate with its source, and the plan stays blocked until each one has one, or when it needs more companies than the market holds. ### Repeated answers become rules Who does it Code What happens Each month, an input on which the model gave the same answer in 95 percent of at least 30 runs is listed as a rule to write, and a step a person overrides more than 10 percent of the time is flagged as routed wrong. ## What do you receive? ### A routed plan before any build Every step with its owner, whether code, model, hybrid or person, and the reasoning behind each score, for one of the four blueprints or for your own workflow. ### A specification per step The rule or the prompt, the answer format, the test or eval with its pass bar, and the fallback. ### The goal arithmetic The number of companies, contacts and replies your target needs, with each rate marked as sourced or as an assumption. ### A monthly routing review The share of runs handled by rules, by a model and overridden by a person, with the steps that should move to code. ## Where does it stop? Torke Router does not build or run any step, call a model or send anything. It refuses to guess the rates the arithmetic needs and blocks the plan until each one arrives with its source. Several eval pass bars in the new blueprints are borrowed from the closest step of the same kind in another Torke plan and are marked as proposed, so each engagement confirms them before launch. It names where each code step should live, and whether that code exists is outside what it checks. ## Who is it for? Who gets the most from Torke Router, and when another tool or no tool is the better answer. ### A good fit - A team about to put AI into a campaign where some steps carry legal, financial or reputational weight, such as consent, deduplication or a lead count a client will dispute. - A team that expects a client, an auditor or a regulator to ask why a step was automated. ### A poor fit - One-off writing with no repeated steps, because there is nothing to route. - A team that cannot supply the conversion rates its blueprint's funnel needs, since the plan stays blocked without them. ## Part of Torke Engine, and available on its own Torke Router is the planning node of Torke Engine. Its lead campaign blueprint assigns the account-scoring steps to [Torke Signal](https://torke.tech/engine/signal), its inbound and outreach blueprints follow [Torke Qualification](https://torke.tech/engine/qualification) and [Torke Reachout](https://torke.tech/engine/reachout) step by step, and its campaign blueprint sends claim checks to Evidence Review in [Torke Context](https://torke.tech/engine/context). Each plan can name your own context source, so the instructions it writes point at your approved facts rather than at Torke's. Torke Engine already carries a ported version of its goal arithmetic, tested to give the same answers as the original. Torke Router is a Torke Node, so you can run it on its own or inside Torke Engine with the other nodes, and add the others later without starting over. ## How much does Torke Router cost? Three ways to run it. Prices are in Swiss francs, excluding 8.1 percent VAT, and media or third-party data you choose to add is billed at cost. Run by Torke CHF 1’900 a month - We operate Torke Router for you inside Torke Service: the setup, every run, the checks and a named operator who answers for the output each week. [How Torke Service works](https://torke.tech/service) Rent it, where most teams start CHF 690 a month - Your team runs it from its AI assistant or the command line. Setup, the first supervised run, updates and support are included, and you can stop with 30 days' notice. [Rent Torke Router](#get-in-touch) Own it CHF 6’900 once - A licence to run it on your own infrastructure with no monthly fee, the setup session included. Updates and support are optional at CHF 1’400 a year. [Ask about owning it](#get-in-touch) ## Where does it stand today? The routing, the specification compiler, the goal arithmetic and the monthly review are built, and all 31 automated tests pass. Since 19 October 2026 it ships four blueprints, each declaring its own funnel, an init command that turns your own list of steps into a plan, and a setting for your own context source. Every blueprint reports itself blocked until you supply its rates, because none have been measured on a client account yet, and the monthly review has no run logs to read until the steps it specifies are built. ## What do teams ask about Torke Router? Each answer stands on its own, so you can forward one to a colleague, and an assistant asked the same question can quote it as it is. ### Does Torke Router call AI models itself? No. It writes the specification for each step, including the model, the instructions and the test, and the agent or the engine that runs the step makes the call. It needs no API key and makes no network calls. ### Can we route our own workflow instead of a blueprint? Yes. List your steps in one line with the init command, score each one on the five questions, and the Router assigns the verdicts and writes the specification. You can change a score with a written reason, and a verdict itself can't be edited by hand, so every decision stays traceable. ### Why does temperature 0 not make a model step deterministic? Temperature 0 lowers sampling variance, but the same input can still return a different answer. That is why the Router decides each step on whether its rule can be written down, and never on a model setting. ## The other Torke Nodes [ Torke Context - ](https://torke.tech/engine/context) [ Torke Signal - ](https://torke.tech/engine/signal) [ Torke Visitor Signal - ](https://torke.tech/engine/visitor-signal) [ Torke Reachout - ](https://torke.tech/engine/reachout) [ Torke Qualification - ](https://torke.tech/engine/qualification) [ Torke Ledger - ](https://torke.tech/engine/ledger) [Torke Engine - ](https://torke.tech/engine)