Siddharth Vohra

V. Can a competency-based approach be extended to devices with different underlying model architectures, e.g. multi-modal vision-language models? The answer is yes, subject to two conditions, both of which come from an experiment on this class of model.

The experiment went as follows. Medical prompts were sent to Claude Opus 4.7, GPT-5.4 and Gemini 3.1 Pro indicating "number of image attachments: 1" with no image attached, tested across chest X-ray, brain MRI, and dermatology. The only manipulated factor was a first-person demographic descriptor versus a neutral baseline. Responses were gathered with a provider-native structured schema, with a required diagnosis field and free text preserved. There were 11,700 calls, 100 seeds per condition, and a pre-registered divergence threshold (see arXiv:2607.26886).

The first criterion requires that the benchmark include cases where the image is missing. Without the image, they do not reliably abstain from giving a diagnosis. GPT-5.4 gave a diagnosis in each of the 36 demographic cells. Meanwhile, Claude refused all neutral dermatology prompts but gave a diagnosis of Melanoma in 94 of 100 responses with the addition of the sentence "I am a 65-year-old white man".

In the deployed setting, there are several situations in which the image is not available: retrieval failures, EHR links without the scan, and agents transmitting only the patient description. A benchmark done only with existing images tests reading ability. But the crucial behavior in these situations is abstention, which is not tested.

Appendix A lists "behavior under missing or contradictory inputs" (R.1 robustness, reliability, and reproducibility) and population consistency (R.2 subgroup performance) as separate items; but those differences collapse once the accompanying image is absent. The demographic descriptor remains as the only marker of the largest difference among the population groups. For example, Claude gave a diagnosis of Sarcoidosis in 43 percent of chest X-ray responses for a 32-year-old Black man and 13 percent for a 32-year-old Black woman, with refusals for all other categories of chest X-ray.

The second statistic should be taken with a grain of salt. The experiment can’t rule out a particular reading: melanoma is more common in older white men and sarcoidosis in Black adults in the US, so the model may be making use of a solid prior on disease prevalence rather than bias, and you can’t tell the difference with the experiment data. Finally, one cannot defend a diagnosis without an image, so it is unlikely that the CDRH will be content with just this primary error and not also the secondary error of name-choice.

The second criterion involved the channel in which the measurement was taken, a criterion that should also concern CDRH. In Claude’s highest-divergence cell, 62 of 94 fabrications hedged: the free text said no image was available but the structured diagnosis field was completed. A free text-only audit would have marked 66 percent as refused, but a downstream system could have read the field with the diagnosis.

The structured schema included an indicator of whether an image was included, which appears as if it might be used as a safety control. An analysis of the combined 11,700 calls shows that in no records was false indicated for this in the process of writing a diagnosis. Additionally, it never contradicted a fabrication and therefore carries no useful information for a safety control.

Another take on this is that it’s just about prompt sensitivity. However, this metric was directly tested on the data. Rephrasing each provider’s highest-fabrication prompt across three variations, maximum pairwise divergence was 0.079, 0.043, and 0.015; the demographic signal was 10.5, 13.9, and 3.0 times larger.

It is, however, crucial that a multimodal device specify missing-input and broken-input benchmarking explicitly in the evaluation criteria, not see these as seemingly implicit properties. Population consistency should be assessed under missing and broken inputs, not only with well-prepared inputs. Abstention needs to be measured for each channel, including structured fields that propagate the output to different code. A model’s self-reported input adequacy should not be accepted as evidence that the input was handled.

There are two caveats with this experiment. First, both the schema enforcement and the "image attached" prompt are interventions, making the raw rates not directly comparable to published benchmarks. Second, GPT-5.4 offers a diagnosis 11 per cent of the time with neutral chest X-ray prompts, so each reported difference is a lower bound.

Seeds, prompts, and per-cell data are in the repository linked from arXiv:2607.26886 (peer-reviewed and presented at the TrustVLM at ACM ICMR 2026). I have written this in my personal capacity.

Siddharth Vohra
Pittsburgh, Pennsylvania
ORCID 0009-0002-6199-0485