Measuring whether an agent can trust its eyes over its database: multimodal, multi-tool reconciliation under adversarial conditions.
Summary · Layout · Tiers · Segments · Inventory · Analysis · Coverage · Dataset · Trajectories · Scoring · Adversarial design · Reproduction · Verification
Mikazuki measures real-world assistant competence, not isolated tool calling, retrieval, or
vision. It evaluates… See the full description on the dataset page:
https://huggingface.co/datasets/ethara/mikazuki-samples.
