A model is not a rule
A language model predicts a likely next word. That is why it can draft, summarize and spot a phrase. It is also why, left alone, it will paraphrase a rule, skip a qualifier, or apply a standard it was never given. Useful for writing. The wrong behavior for a review. A compliance check has to treat the rules as fixed, and the model as a way to find what was said.
The rules have to be the system
Putting “be careful with health claims” in a prompt is not a ruleset. A ruleset is a named, versioned list: which claims are approved, which market rules are on, and what counts as a finding. The system’s job is narrow. Read the speech, the on-screen text and the caption. Match those words to the rules that were turned on for that campaign. Cite the moment, the words and the rule. It does not get to widen the standard because a looser reading sounds reasonable.
Agents make consistency harder
An agent does more than answer once. It plans, calls tools, looks things up and writes again. Each of those steps can take a different path. Two runs of the same video can then disagree, and a disagreement looks like a second opinion when it is really drift. An agent used for review has to be constrained. The rules in scope do not change mid-run. The finding has a required shape. If the same input and the same ruleset do not return the same findings, the run has failed.
What a brand should expect
Same video. Same rules. Same result. A person still signs, because a model should not be the one who says the content is compliant. Selado is built on that split: the rules for the campaign are the ones that run, each finding sits beside the moment in the video, and the certificate records the review. The principle is not particular to one product. Any system that checks claims, and any agent that acts on them, is only as trustworthy as the rules it is not allowed to bend.
