Medical agents stop treating every tool output as evidence: CIEG rules first, reaching 0.730 balanced accuracy in glaucoma screening
Synopsis
Tool-augmented medical agents treat whatever their tools return as evidence, and a frozen reasoner cannot ignore what enters its context, so a wrong output flips cases it would otherwise answer correctly. CIEG moves the ruling on what counts as evidence outside the reasoner: each tool output becomes a card checkable against the raw input, a checker that sees only the input and the cards decides admission, and the solver answers from admitted cards alone, with no training. On fundus glaucoma screening, chest X-ray pneumonia classification and medical question answering it reaches the highest accuracy among models of comparable scale and significantly outperforms tool-using agents and multi-agent systems reading the same cards.
Interpretation
Medical agents need an admission step independent of the answering model, because a tool output that enters a frozen reasoner's context sways the answer even when it is wrong. Earlier approaches either chose which tools to call before any output existed or reflected, verified and negotiated inside the reasoner's own context, with every output entering that context in full once produced. In glaucoma screening, handing all three segmentation cards to the solver flips 178 originally correct cases (44.5%) to wrong; with the same card pool, admit-all drops balanced accuracy from 0.83 to 0.61 with a single erroneous card and below 0.5 with more.
CIEG splits tool-augmented reasoning into acquisition, admission and answering, with admission performed by a checker that never enters the solver's context. The checker reads only the raw input and the card pool along two lines, fidelity (whether the material a card shows supports its finding) and consistency (whether cards corroborate or contradict one another), and conflicts only mark where to look while the ruling rests on fidelity to the input. Checker and solver are both Qwen3-VL-8B, fully frozen with nothing trained; on glaucoma the checker separates correct from erroneous segmentations with AUROC 0.93 and 0.85, admitting most correct cards but only about one in five erroneous ones.
The gain from admission splits into seeing less and selecting better, and only a gate that reads card content obtains the second part. Relative to admit-all, randomly dropping the same number of cards also changes results, so an evidence filter should be judged against a budget-matched random gate rather than against admit-all. CIEG exceeds budget-matched random admission on all three tasks; its margin over the better trivial policy is 0.094 on glaucoma (against admitting no cards), 0.016 on chest X-rays and 0.55 points on medical QA.
Perspective
The demonstrated scope is three tasks matching three kinds of tool output: glaucoma screening on the 400 fundus images of the REFUGE2 validation set, pneumonia classification on the Kermany pediatric chest X-ray test set deduplicated by patient to 431 images, and question answering on MedQA, MedMCQA and NEJM; checker and solver are both frozen Qwen3-VL-8B, and the solver takes a majority vote over 8 samples.
The theory guarantees only that the governed system is no worse than the better of admitting everything and admitting nothing, so whether a practical gate exceeds both remains an empirical question; the budget term changes sign across tasks, and on chest X-rays and QA the cost of seeing less offsets most of the selection gain, which is why accuracy there stays close to admit-all.
