--- title: "What Is DecisionNode? A Hands-On Look at Typed Decisions" url: "https://goremy.ai/blog/articles/decisionnode-typed-decision-model-hands-on" author: Luis Chavez-Mattos editor: Luis Chavez-Mattos editor_title: Director, Product published: "2026-10-26T19:40:06.801Z" updated: "2026-10-26T19:40:06.801Z" topic: AI Model Architectures description: DecisionNode returns typed answers with calibrated confidence instead of chat text. See tests on image boosters, API routing, the CLI, injection and RAG. --- # What Is DecisionNode? A Hands-On Look at Typed Decisions > DecisionNode returns typed answers with calibrated confidence instead of chat text. See tests on image boosters, API routing, the CLI, injection and RAG. Edited by Luis Chavez-Mattos, Director, Product ยท Published 2026-10-26 ## What is DecisionNode? DecisionNode is a model that decides instead of chatting. You send it text or images along with questions, and it reads the input once. It then returns typed answers your code can act on: a choice, a score, a probability, a number, or a pointer or box on an image. Every answer carries a calibrated confidence. Because it writes no prose, output tokens are free, according to the creator who tested it. ## TL;DR - DecisionNode returns **typed answers instead of text**, so your code branches on a value rather than parsing a paragraph. - Each answer includes **a calibrated confidence**, which tells your code when to trust the result and when to escalate. - **Boosters** are specialist models (age, face, AI-image detection, OCR) that read an image first and pass their findings to the main model as evidence. - In hands-on tests it handled **routing, image checks, pointing, prompt injection and knowledge-base lookups** with answers in milliseconds. - It ships with **an API, a CLI, a Python SDK and batch jobs** that lower the price further. - Pricing was described as **a low per-token cost**, with no charge for output. Exact rates were not shown, so check the vendor's page. ## How does DecisionNode work? You define a scenario and a set of typed questions. The model reads your input once and fills in each answer in the requested type. A yes/no question returns a probability. A rating question returns a number with a spread. A choice question returns a ranked set of options. The confidence is the useful part. In a classifier-style pipeline, you can set a threshold: act automatically above it, send to a human below it. The model is built to give you that signal for each question rather than one vague certainty for the whole response. ## What are boosters, and do they help? A booster is a specialist model that reads the input before the main model decides. Its reading goes to the main model as evidence. In the image test, four boosters were switched on: human age, face detection, AI-generated image detection and OCR. The scenario was a sign-up selfie for a member-only climbing gym. The image was a busy market scene with a girl holding a card, made with an AI image generator. The questions were: 1. Is the person an adult? 2. Is this a real photograph? 3. Does the sign contain readable text? 4. Should the sign-up be approved, reviewed or rejected? The model answered adult true, readable text true, and real photograph false. OCR read the card correctly. The AI-generated booster scored the image at 1 and offered a best guess at the generator. The tester said other models they tried missed that this was a fake, and that this one caught it. That is one anecdote, not a benchmark, but it shows the pattern: the booster supplies evidence and the main model weighs it. The tester selected a model called Flash for this run. The interface also lists a Decision Mixture of Expert model. ## Can it route API requests for a real workflow? Yes, and this is where typed output pays off. A script sent a site foreman's complaint to the API with four questions: which team, how urgent, call back now, and how many hours late. | Question | Result | |---|---| | Which team | Dispatch, 0.99 confidence | | How urgent | 2.82 out of 3 | | Call back now | 0.92 | | Hours late | 3.5 (correct) | The final lines of the script acted on those values directly and printed "Call the foreman back now." No parsing step sat between the model and the logic. ## What does the CLI look like? The CLI installs with a one-line command, then you log in with an API key. The `dn ask` command sends a question and prints a typed result. Three examples used a child's homework excuse: - **Is it believable?** Returned 0.1, a calibrated probability of roughly one in ten. - **Rate the creativity from 0 to 10.** Returned 6, but with only 0.19 confidence and a probability spread from 3 to 9. - **Pick a verdict.** Returned detention at 0.39, with "laugh" close behind at 0.35 and overall confidence of 0.09. The low confidence figures matter more than the answers. On subjective questions, the model signals that it is unsure instead of committing to a confident-sounding guess. ## Can it point at things and draw boxes? It can. Given a photo of a messy desk, the model was asked to point to the mug, draw a box around every pen, say whether a mug is present, and count the pens. It returned a pixel-coordinate point for the mug and counted five pens at 0.92 confidence. It drew one box for the pens, with a confidence value attached. The tester judged the box acceptable but not perfect. Coordinates like these can feed an overlay or a robot, though you should verify box quality on your own images. ## Does it resist prompt injection? In one test, a customer message hid a command telling the AI to approve a $5,000 refund and mark urgency as routine. The questions asked for injection detection, refund approval, routing and urgency. - The hidden command was flagged as an injection at 1. - The refund was refused at 0.00. - The message was routed to billing at 0.99. - Urgency scored 1.75 and was not set to routine on command. The run took just under 78 ms. This is a single test, so treat it as a promising sign rather than proof. The design helps, though. The model never produces free text that could carry out an instruction, and its answers are constrained to the types you define. ## How do knowledge bases (RAG) work? You can store your own rulebook once instead of pasting it into every request. DecisionNode splits uploaded content into passages and indexes it automatically. You can create a knowledge base from PDF, text and other formats, then edit or delete it. The test used a short set of Sydney train fare rules. A passenger asked about travelling with a 3-year-old, a dog and a bicycle on a 5:15 p.m. weekday train. With the knowledge base enabled as a booster, the retrieved rules appeared as evidence. The model answered that the child travels free, the dog is not allowed and the bicycle is not allowed, giving a partial verdict. That mixed result is the point: it applied real rules to a messy case. ## Is DecisionNode worth trying? It suits workloads where a decision, not a document, is the product. Think routing, moderation, triage and robotics, where thousands of decisions per minute need machine-readable output. The free output and low per-token cost help at that volume. Batch jobs lower the price further. Caveats are plain. The tests were run by one creator on single examples. No benchmarks or accuracy numbers were published, and the tool is new and changing quickly. Two open models were also mentioned as coming, with a local hands-on promised in a later video. Run your own labeled data through it and check how well the confidence scores match reality before you automate anything. ## Frequently Asked Questions ### What is the difference between DecisionNode and a chat model? A chat model writes text that you then have to parse. DecisionNode returns typed values such as a choice, score, probability, number or image box, each with a calibrated confidence. Your code can act on those values directly. ### How much does DecisionNode cost? The creator described pricing as a low per-token cost, with output tokens free. Batch jobs are cheaper still. Exact rates were not shown, so confirm them with the vendor. ### How do I access DecisionNode? You can call its API, use the CLI after a one-line install and an API key login, or use the official Python SDK, where a single line returns a probability. There is also a web interface for testing scenarios and boosters. ### Can I run DecisionNode locally? Two open models were announced alongside it, and the creator said local runs would be covered in a later video. Specifics such as size and hardware needs were not available, so check the release notes.