RECALLZERO

RECALLZERO — One More Photo. A Different Decision. What happens when an AI must decide not just what it sees, but whether it has enough evidence to act? RECALLZERO explores a new capability for multimodal AI: turning incomplete, conflicting, and potentially misleading observations of physical objects into explainable, evidence-aware decisions. Powered by Meta Muse Spark 1.3, RECALLZERO examines product photographs, identifies missing information, compares markings across multiple views, and recommends whether to request additional evidence, allow a narrowly defined visual-consistency check to pass, or hold a case for human review. In our working demonstration, one photograph shows a RYOBI battery marked 6Ah. Muse requests another view because the model code is missing. When a second photograph reveals RB18L40, Muse identifies a suspected discrepancy and recommends holding the case for review. An unverified claim that the item is genuine does not reverse that decision. We built six pre-sealed evaluation conditions using publicly sourced photographs. Muse matched all six predefined decisions, compared with five of six for a conservative rules baseline using human-transcribed markings. The system includes SHA-256 evidence sealing, blind model inputs, archived reasoning outputs, a verifiable replay interface, and a simulated policy controller. This is an exploratory physical-world reasoning prototype—not a counterfeit certification system or a live checkout integration. Our frontier: machines that recognize when evidence is insufficient, revise decisions when new observations arrive, and preserve an auditable explanation of why.