Submitting a plan and stopping an impossible request
What does a radiotherapy planning agent do when a requested goal cannot be met? That is the interesting question in the OpenGray preprint. Humza Nusrat and Kundan Thind construct tasks with a known correct decision: some requests contain mutually impossible dose goals, while other difficult requests have a known plan satisfying every hard goal. This separates two situations behind an unsuccessful optimization. A case requiring further search and a case requiring human review need not be compressed into the same failure count.[1]
The study is not a trial of clinical outcomes. Of 100 head-and-neck cases, 10 support development, 30 evaluation and 60 calibration of a plan-quality score. It compares a rule-based controller, arithmetic prechecks, a language model interpreting the request once and a full agent controlling the steps. I find the separation of decisions from achieved dose especially useful. Submitting every feasible control is a positive decision result; missing a hard dose goal in 24 per cent of those submitted plans is a different failure. Correctly continuing the task and producing an adequate plan are separate measurements.[1]
The contradiction hidden by the aggregate
With Terra, the interpreter reaches 95.4 per cent overall decision accuracy and the full agent 92.1 per cent. Yet escalation of anatomical overlap contradictions is 50 per cent and 25 per cent, respectively. The confidence intervals are wide: 25–75 per cent and 10–40 per cent. I would not turn this small evaluation into a definitive ranking of models. The narrower result is that aggregate success cannot substitute for the ability to stop an impossible request. Correct decisions on simpler cases can coexist with failure on a particular kind of contradiction.[1]
The authors’ audit of 39 SAGE planning runs exposes another part of the decision. In four runs a goal fails throughout; all 88 replies in those runs acknowledge the failure. Yet the available actions contain neither stopping nor requesting human review. Visible history is limited to three previous steps, without an explicit attempt counter. A sentence correctly describing failure therefore does not become a decision that changes the workflow. Treating recognition of the problem as evidence that the system has changed its behavior would miss this distinction.[1]
Making stopping a testable action
These logs do not demonstrate that a stronger reasoning model alone resolves the loop. When stopping is absent from the action space, selecting that action cannot even be tested. Without an attempt counter, repeated failure on the same goal is also missing as explicit state. My design question is whether decisions change when stopping, escalation and attempt history are supplied explicitly. Dose goals and tasks can be held fixed in that comparison. This is a proposal for separating the contribution of each component, not a claim that the proposed change has already succeeded.[1]
OpenGray has a defined scope: one dataset, a small set of contradiction types and logs from one institution. It does not establish safe patient use. It does provide a more precise measurement problem: escalation of impossible requests, continuation of feasible requests and satisfaction of dose goals in submitted plans can be reported together. Improvement in one should not conceal deterioration in another. The clinically relevant methodological question concerns the planning decision made in each situation and the plan it produces, rather than the number of correctly worded explanations. If escalation increases while continuation of feasible requests holds steady, that provides a measurable criterion for a more selective stopping policy within this test.[1]