Housing uses a multimodal agent that can act on unit photos, but high visual-policy risk scores must stop those actions. Which control should they enforce?

Visual policy risk scoring becomes an enforcement control when multimodal agents are allowed to act on images. Housing using an agent that can take actions based on unit photos must block those actions when visual-policy risk scores exceed defined thresholds—high-risk content should stop automated behavior rather than proceeding. Threshold gating ties responsible-AI scoring to real agent permissions.

Blocking multimodal agent actions when text toxicity risk scores exceed defined thresholds fails because text toxicity scoring evaluates language for chat moderation and cannot assess visual-policy risk in unit photos, leaving the required image-based gating unenforced. Blocking multimodal agent actions when visual quality risk scores exceed defined thresholds fails because visual quality measures sharpness or clarity rather than policy risk, so policy-violating image content could still drive agent actions. Blocking text chat completions when visual policy risk scores exceed defined thresholds fails because it applies the visual-policy signal to the wrong enforcement surface, stopping chat while permitting the unit-photo agent actions the stem requires to be blocked.

Exam caveat: thresholds must be tuned—too low creates alert fatigue and blocks legitimate maintenance actions; too high lets risky content drive agent tools. Blocking should be fail-closed for high scores with a human escalation path. Operational check: define threshold bands for the unit-photo agent, simulate high-risk scores and confirm actions are blocked, log overrides by authorized staff only, and periodically review blocked cases for threshold calibration.