# Calibrated Decisions > Calibrated Decisions is an independent directory and introductory guide to Jev, RLCD, and the TypeSafe AI ecosystem. It helps people understand calibrated decision models, find tools and examples, and explore what can be built by combining small AI judgments in software. Website: https://calibrateddecisions.com/ Guide: https://calibrateddecisions.com/guide Contact: contact@calibrateddecisions.com Not affiliated with TypeSafe AI and Jev. ## What is Calibrated Decisions? Calibrated Decisions brings Jev and RLCD resources together in one compact, topic-organized directory. Its purpose is to make a developing technical ecosystem easier to understand and explore without searching through scattered documentation, repositories, demonstrations, and community lists. The site combines a resource directory with a concise, lay-accessible guide. Readers can learn what the terminology means, see the kinds of applications people are exploring, and find starting points for their own projects. Calibrated Decisions is a reference website, not a model provider or hosted playground. It does not operate the Jev API, sell model access, or provide official TypeSafe support. ## Who is the site for? Developers can find client libraries, integration tools, decision patterns, and example projects. Researchers can discover evaluation experiments and questions about classification, ranking, and probability calibration. Readers who are new to the subject can use the guide to understand Jev and RLCD before exploring the more technical resources. The directory is useful both for someone asking "What is RLCD?" and for someone looking for a Jev SDK, an agent integration, a re-ranking example, or a decision-model benchmark. ## What does the directory cover? The directory groups resources into practical categories: - Getting started: introductory documentation, quickstarts, API references, and explanations of model limitations. - Ideas and patterns: decision primitives, confidence, parallel questions, hierarchical classification, and re-ranking. - Official kits and community clients: Python and JavaScript SDKs alongside integrations for other languages and application frameworks. - Agents and developer tools: agent routing, code-review experiments, guardrails, MCP tools, and telemetry projects. - Browser and mobile: browser automation, computer-use experiments, mobile applications, and page filtering. - Demos and games: interactive examples that show decision models operating inside an application. - Research and benchmarks: community experiments in classification, ranking, and typed scoring. - Collections and community: further project collections and places where people discuss what they are building. These are descriptions of the directory's coverage. A listing does not mean that Calibrated Decisions created, maintains, endorses, or has audited the linked project. ## What does the RLCD guide explain? RLCD stands for Reinforcement Learning for Calibrated Decisions. It is TypeSafe's name for its approach to training models to make bounded judgments with useful probabilities, rather than produce open-ended prose. Calibration concerns the relationship between predicted probabilities and observed outcomes across many comparable cases. If predictions assigned 70 percent probability are well calibrated, roughly 70 percent should be correct. This does not guarantee an individual answer. Jev is TypeSafe's System One decision model. An application supplies context and a defined question, then uses the returned value in its own logic. The guide introduces three decision types: - Noul: a probability for a yes/no proposition. - Choice: a selection among named alternatives. - Score: a judgment against ordered rubric levels. The guide also explains why people combine narrow questions and why structured, confident outputs can still be wrong. ## What kinds of projects does the site help people explore? The site's resources provide starting points for applications such as support-ticket routing, document triage, product-feedback classification, moderation review queues, search-result re-ranking, agent-task routing, and prioritization of logs or review findings. A workflow might classify a short document, evaluate several attributes in parallel, combine the results with explicit weights, and send uncertain cases to a person. A hierarchy can move from a broad category to a narrower one. A ranking workflow can compare candidates or update a shortlist as new material arrives. These are illustrative application patterns to evaluate, not live services provided by Calibrated Decisions or guarantees of model performance. ## How should readers use the research resources? The research category is a starting point for comparing methods and examining results, not a claim that one model is best for every task. Readers should consider the dataset, rubric, model version, and evaluation procedure behind a result. Classification accuracy, precision, recall, sensitivity, specificity, confusion matrices, and ROC AUC describe different aspects of performance. Calibration asks a separate question: whether the probabilities match observed frequencies. Threshold choices should be evaluated on data separate from the final test set. The site encourages readers to check representative examples and retain human review where mistakes matter. ## How are additions and corrections handled? Readers can suggest resources, request corrections or removals, and report suspicious repositories through contact@calibrateddecisions.com. Third-party projects retain their own maintainers, licenses, and terms. Calibrated Decisions is an independent guide to the ecosystem. Current service details, pricing, and API behavior belong to the respective providers, not this directory.