orchestrate-multimodal-evidence-groundinglisted

# Orchestrate: Multi-Modal Evidence Grounding **Direct evidence**: the multi-modal-review challenge's required output schema includes `supporting_image_ids` (which specific images support the verdict), `claim_status_justification` (evidence-grounded explanation), and a three-way `claim_status` split between `supported`, `contradicted`, and `not_enough_information` — a structurally different decision space from a binary yes/no. ## Why the three-way split is the core design challenge Collapsing `contradicted` and `not_enough_information` into one "reject" bucket is the most likely single design mistake in this challenge, because they mean genuinely different things and likely warrant different downstream handling: - **`contradicted`**: the image evidence actively conflicts with the claim (e.g., claim describes water damage, image shows no visible water damage on the claimed component) — a confident, evidence-backed negative. - **`not_enough_information`**: the image doesn't clearly show enough to judge either way (wrong angle, too dark, wrong object entirely, image doesn't cover the claimed damage area) — an honest "we can't tell," structurally the same category as the "mark uncertainty" discipline in `orchestrate-failure-handling`, just domain-specific to this challenge. A system that only ever outputs `supported` or `contradicted`, never `not_enough_information`, is almost certainly over-confident — real claim photos are frequently ambiguous, off-angle, or incomplete, an