Synthetic data only · not medical advice · not for clinical use
Between-Visit

From the hospital door to day 30

Each journey is one patient's first month after a hospital stay, run by a team of agents on the guideline graph. Scrub through it day by day: what the patient texted, what each agent did and why, what waited for the clinician, and which deadlines were met.

TransitionBench: thirty months, scoredWhat the system caught against keys written before the code, what each part buys, which helpers need a model, and the autonomy dial.The guideline graph, patient by patientAll the guidelines on one map. Pick a patient to see which points are theirs, hop by hop, and play their month.

One ledger

Everything is a row, written once: the discharge, every text, lab, fill, agent step and clinician decision. Screens and scores are read from it, and it replays exactly.

A conductor in code

It wakes the agents the program names for each row and lets out only what the autonomy level allows. A safety reply always goes out; a medicine change never goes out without a person.

Six agents

Reader, Medicine check, Planner, Sentinel, Companion, Watcher. Models read words where tests showed code can't: the note, texts, answers and habits. Code reconciles medicines, sets deadlines, walks the guidelines and decides what goes out.

A simulated month

The patient, lab, pharmacy and clinician are simulated, so a month runs in seconds and the same run plays the same way every time.

FUTURE SCOPENot built yet · design under review

A graph that learns from clinicians, and never rewrites the evidence

Every clinician decision is already a row in the ledger: yes, change, or no, with a reason. Next, the system learns from those rows in two places kept beside the guideline graph. The graph itself still says only what the guidelines say.

When the reason is about this patientPatient context record“Tried a statin years ago; leg cramps.” Next time the plan shows “declined 27 Jan: tried before” instead of offering it fresh.This patient only. Useful from day one.
When it's a pattern for patients like thisLearned fit rulesOften declined for, often accepted for, or did worse for a patient type, described only in facts the walk already checks.Needs many clinicians' decisions to mean anything.
When it's one clinician's preferenceKept, never a ruleA preference is not evidence about patients, so on its own it changes nothing for anyone else.
How a fit rule is made
  1. Clinician decidesand gives a reason
  2. Nightly mining20+ patients, 5+ clinicians, 2+ sites
  3. Replayon months it never saw
  4. Bias checkno group singled out
  5. Lead approvesa month in shadow first
  6. Live, then expiresre-reviewed when the guideline changes
What the clinician would see
Start a moderate-intensity statin
Often declined for patients like this[n] patients · [n] clinicians · [n] sites · most said: [main reason]
AcceptChangeDecline, with a reason
Guardrails
  • A rule adds a doubt. It never hides a safety warning or a strong recommendation.
  • No race, sex, language or insurance in a patient type.
  • Outcomes beat declines: if patients a rule steered away do worse, the rule goes.
  • Every doubt shows its counts and reasons, and can be switched off for a patient.