October 2026
Most medical AI benchmarks are static. A case is written down once, the model reads it, and its answer is compared to a key. Nothing the model does changes the patient. Static benchmarks can grade an answer. They cannot price a delay, and in emergency medicine delay is often the whole story.
In a large study of septic shock, patients given an effective antibiotic within the first hour of low blood pressure survived about 80 percent of the time, and each hour of delay over the next six was associated with an average 7.6 percent drop in survival [1]. In New York State, a study of nearly 50,000 patients with sepsis found that each additional hour before antibiotics was associated with about 4 percent higher odds of dying in hospital [2]. In a typical stroke caused by a blocked large artery, the brain loses an estimated 1.9 million neurons for every minute it goes untreated [3]. And when air trapped in the chest is crushing a patient’s lung and heart, trauma teams release it at once rather than wait for an X-ray [4].
The same diagnosis, made at minute ten or at hour four, is not the same care. A benchmark that marks both correct is grading knowledge, not medicine.
To an environment designer, findings like these are a specification. Time has to be part of the state, and the state has to move on its own. That has three consequences a static benchmark cannot express.
First, inaction is an action. In a static benchmark, waiting costs nothing. In a patient whose physiology is drifting, waiting is a choice with an outcome, and the environment has to score it like any other.
Second, information has a price. A test takes time to come back, and the patient changes while it does. A good clinician weighs the value of information against the cost of delay. For a crashing patient, the test that would confirm the diagnosis can be worth less than the treatment it delays. An agent can only learn that trade-off in an environment where both sides of it are real.
Third, actions take time and overlap. Orders are placed, carried out and resulted on their own schedules, and a drug given now acts later. The agent has to plan across those delays, which is closer to scheduling than to answering a question.
A clock also makes the learning problem harder, in an instructive way. An emergency visit can span hours and dozens of decisions, and the outcome arrives at the end. Which decision made the difference? This is the credit assignment problem, and it is the central difficulty of long-horizon reinforcement learning. An environment that only reports the final outcome gives one reward per episode, which makes it hard to tell which decisions mattered. An environment that also scores each decision as it happens gives the agent a signal at every step. We take up that design choice in The score is the curriculum.
None of this works if the clock is wrong. A patient who deteriorates too fast teaches recklessness. One who deteriorates too slowly teaches complacency. An agent trained in a simulator learns what the simulator actually does, not what its designers intended, so the dynamics have to be grounded in clinical evidence. A training environment is a claim about how the world works, and the agent will take it literally.
The sickest patients are where minutes matter most, and where the gap between a good decision and a bad one is widest. That makes them the most important test of whether an agent acts in time, and they are exactly the patients a static benchmark grades the same whether the agent acted in time or not. A clock is what turns a quiz into practice.