Clinical Journey Crew: An Agentic Framework for Evidence-Grounded Longitudinal Summarization of Pediatric ASD Therapy Notes

Children with autism spectrum disorder (ASD) attend several developmental therapies across their clinical journey. Those therapies produce hundreds of free-text, multi-specialty session notes that record temporal observations such as responses, milestones and challenges. Reconciling them relies on manual chart review over months of text, on top of a full day of in-session work. We demonstrate Clinical Journey Crew, an agentic framework that orchestrates temporal summarization over a patient's accumulated record and produces evidence-grounded reports using only open-weight models on institution-local hardware, so that no clinical text leaves the provider. To prevent hallucination, generative capacity is confined to a bounded generate-verify-critique-repair-abstain loop: a local model proposes claims with verbatim evidence, deterministic validators adjudicate, and the loop repairs an unsupported claim or executes a fail-safe abstention. Retrieval is left deliberately deterministic, making cross-patient conflation structurally impossible rather than prompt-mitigated. On 4,031 Portuguese therapy notes from 15 pediatric patients across seven specialties, the loop resolves 84.9% of notes on the initial pass and recovers 87.7% of the remaining failures to strict verbatim compliance; the abstention path engages on 0.50% of note-dimensions, which are withheld rather than emitted without support. We describe the architecture, the four-dimension report structure, and a three-part evaluation framework: grounding and cost metrics captured on every run, a blinded therapist task measuring omitted temporal claims, and a nine-construct clinical rubric. Because adjudication is deterministic rather than model-mediated, every claim's fate is attributable to a named check with a stated criterion: clinical auditability, not accuracy alone. The framework is not intended to replace clinical judgment, but to act as an AI lens over the ASD clinical journey that amplifies temporal observations and eases the cognitive fatigue of chart review.