l0-0ll0-0l/talia

TALIA describes itself as an autonomous research atelier: a web app where someone submits an open scientific, mathematical, or computational question, and a set of Claude powered AI agents investigate it in a structured, multi step process. Rather than templated output, each step is a real model call that reads the actual question and the prior agents' output, so the resulting report reflects what actually happened during that specific investigation. The process runs through named agent roles. Seed reads the submitted question and pulls out the assumptions and testable claims hiding inside it. Mirror builds a faithful baseline reproduction of the idea being tested. Thorn then attacks that baseline, stress testing assumptions and looking for the smallest case that breaks the hypothesis. Vein diagnoses why each failure happened, tracing it to a code defect, a misspecified model, a weak metric, thin data, or invalid framing. Bloom then records the minimum conditions under which the broken idea would actually work, plus a set of related but viable neighboring ideas. Finally Archivist writes up the closing report, including the verdict, findings, and methodology, and the run can be published to a public archive. A notable extra feature is an inherited seventh agent called Scion. Whenever any prior public run finished with a recorded set of minimum conditions for success, the very next run anywhere in the app inherits that condition and a Scion phase judges whether the new hypothesis honors, strains, or violates it. This inheritance is tracked atelier wide and shown as a timeline of the most recent firings, though private runs never have their conditions stored or inherited, which is described as the app's privacy mechanism. The project is built with Vite, React, and TypeScript, and its AI agents run on the claude-fable-5 model. It includes a home page, a studio page for starting a new run, an archive of published reports, and a manifesto page describing its editorial stance. The project is released under the MIT license.