01Why let a machine read the materials
Reviewing promotional materials looks like reading work, but it is really searching work. Is there an expression somewhere in the material that exceeds the approved scope? Has a safety statement been watered down? Has a graph axis been manipulated? Most of a reviewer's time goes into this search. Materials run to dozens of pages, and the norms (= the body of rules that must be followed) stack up in layers, from statutes to industry self-regulatory codes. Checking every page against every norm is possible in principle for a human, but completing it fully every single time is hard.
The starting point was not efficiency. It was preventing misses. At the very beginning of the design, this system defined false negatives (= overlooking something that is actually a violation) as the gravest failure. False positives (= flagging something as a violation when it is not) can be eliminated later by humans, but a false negative goes unnoticed and the material reaches the world. The readers are patients and physicians, and if they are led to a false picture, they are the ones harmed. So the machine was taught: if in doubt, keep it; if unconfirmed, lower the confidence and keep it anyway. Speed is a byproduct, not the goal.
02What it does, and what it does not do
What this system does can be summed up in one sentence. It detects, classifies, and explains whether the statements, figures, and structure of a material fall within the approved facts (= the scope of indications and dosage approved for that drug), with grounds traceable to the original text of the norms. For determining the approved facts, the electronic package insert (= the official descriptive document for a drug) serves as the practical reference document. The crucial part is "with grounds." No finding is even generated unless it can show which wording of which norm it rests on. Only findings that would withstand a third-party gaze (= critical outside eyes such as regulators, lawyers, and journalists) are produced.
Three things it does not do are also fixed clearly. First, it makes no final judgment. Whether a revision is needed is decided by the human reviewer. Second, it does not adjudicate medical truth. Whether the content of a cited paper is scientifically correct lies outside this system's scope. It judges only whether the way that paper is used stays within the frame of the approved facts and the norms. Third, even when the public gaze (= an inspection that views the material from eight perspectives, including patients, media, and investors) issues a strong warning, the system never directly blocks a material. Warnings go to humans; the gravest warnings go to management. From there on, it is human work.
03Positioned as a first-pass screening
The system's judgment vocabulary (= the fixed list of terms available for judgments) includes WHITE (= a "no problem" verdict). So if the system returns WHITE, can the human review be skipped? No. "We do not skip it" is written into the design premises.
What gets saved is the searching time. Sifting dozens of pages for suspect passages, pulling up the original text of the applicable norms, and attaching an explanation of why each is a problem — the machine does all of that. The human starts the review not from a blank page but from a list of findings with grounds attached. What is shortened is this search, not the judgment. No responsibility transfers, at all. This line is the premise for all three articles that follow. Put the other way around, it is precisely because of this premise that the machine could be given the bold design instruction "over-flag rather than miss." The cleanup of excessive findings can be absorbed by the human who makes the final call.
04How to read the three articles
Besides this page, the section consists of three articles. We recommend reading them in order.
The first is "Design Philosophy" — why we think this way. It starts from the information asymmetry between those who make materials and those who read them, and explains the six principles derived from it, each paired with the objection it answers. It also describes what the choice to prioritize catching misses meant giving up.
The second is "Components" — what it is made of. A hierarchy that arranges the norms into five layers, from statutes to industry self-regulatory codes. A vocabulary that fixes in advance the terms available for judgments, from BLACK to WHITE. A two-lane structure that runs the legal inspection and the public-gaze inspection without mixing them. The 694 atomic items, in which the material's checkpoints are broken down until each can be inspected independently. And a virtual deliberation in which AI plays experts of differing positions and debates the points of contention. It lays out the structure of the parts that support the philosophy.
The third is "Execution Process" — what happens in a single review. It follows the timeline from ingesting the material PDF, through deterministic matching, semantic judgment, the public gaze, integration and scoring, to the handoff to humans. This is where you see the parts actually in motion.
Looking at the parts without knowing the philosophy makes no sense, and following the flow without knowing the parts looks like a magic trick. Read in the order of design philosophy, components, and execution process, and the why, the what, and the how should connect into a single line.