01The Root Axiom: Asymmetry Between Creator and Reader
Promotional materials for prescription drugs (product information overviews, journal advertisements, patient-facing materials, and so on) operate under a condition that sets them apart from other advertising: the relationship between creator and reader is not an equal one.
The creator — the pharmaceutical company — knows more about its own product than the reader does. Moreover, if the reader forms a better impression of the product than reality warrants, the company profits in the form of sales. The reader, meanwhile — the patient or the physician — bears the cost of that misunderstanding personally, in the form of harm to health. The side that gains from misunderstanding and the side that is harmed by it are separate from the start.
Once this asymmetry is acknowledged, only one conclusion follows: materials must not lead readers to form a false picture that goes beyond the approved facts (= the scope of indications, dosage, and administration that the regulatory authority has reviewed and authorized).
The entire design of this system is derived from that one sentence. The six principles below are not six independent rules. Each is an answer to one of the representative objections raised against the axiom. Those objections can be organized into six: the standard, the presentation, the selection, the guarantee, the choice of error, and the gap. That is why there are six principles.
02The Six Principles — The Objection Each One Answers
The first objection: "What counts as an approved fact in the first place?" The answer is Principle 1, Complementarity. The electronic package insert (= the drug's official explanatory document) is designated as the reference document for determining the approved scope, and materials may not exceed that scope by a single character. Materials remain in a relationship of supplementing and explaining the package insert — hence the name. Deviations by suggestion or implication, not only by explicit statement, count as violations.
The second objection: "If what is written is factual, what is the problem?" The answer is Principle 2, the Dual Obligation. Even when each individual statement is factual, presentation that misleads the reader constitutes a violation. The standard of judgment is placed not on the writer's intent but on the impression received by the reader. The name comes from bearing two obligations at once: factual accuracy and impression.
The third objection: "Shouldn't the company be free to choose what information to include?" The answer is Principle 3, the Asymmetric Disclosure Obligation. The choice of wanting to include efficacy but not safety is not permitted. Safety information must be disclosed even when it works against the company's interests. Treating efficacy and safety without any difference in handling is fixed as an obligation.
The fourth objection: "Who guarantees that the judgment is correct?" The answer is Principle 4, Verifiability. Every finding the system produces must have grounds traceable to the original normative text (the relevant passage in the law, standard, or guideline). A finding without grounds that can withstand criticism from third parties — regulators, lawyers, journalists — is not permitted to exist as a finding.
The fifth objection: "Machine judgment inevitably errs. Which error do you choose?" This asks about the choice of error (= the decision of which of two kinds of error to tolerate). The answer is Principle 5, Priority on Preventing Misses, discussed in detail in the next section.
The sixth objection: "Anything not written in the explicit text of the norms should be free." The answer is Principle 6, Textual Gaps = Autonomy, discussed in Section 4.
03Priority on Preventing Misses — Deciding Which Error to Accept
Machine review involves two kinds of error: false positives (= suspecting a violation where there is none) and false negatives (= missing an actual violation). Neither can be reduced to zero. Design also means deciding in advance which error weighs more heavily.
This system defines false negatives as a graver failure than false positives. The reason goes back to the root axiom. The cost of a false positive is paid by the people involved in review, in the form of extra confirmation work. The cost of a false negative is paid by the readers — patients and physicians — in the form of misunderstanding and harm to health. Since the weight of the errors is not symmetric, the countermeasures are not made symmetric either.
This decision is carried through consistently into the details of judgment. First, concerns without confirmation are not discarded. A passage that is suspicious but cannot be judged conclusively is retained with a lowered confidence score (= a numerical value indicating how certain the judgment is), rather than dismissed. Second, when repeated judgments disagree, majority vote is not used. If the question is asked multiple times and even one vote returns a violation, that finding is adopted. This is worst-case adoption. Furthermore, any operation that drops a finding — acceptance or withdrawal — requires grounds and a record. This is to prevent misses that vanish quietly.
Put the other way around, this is a decision to accept the cost of over-flagging. Human reviewers end up spending time confirming findings that turn out not to be violations. Even so, it is better for the machine to raise noisy suspicions and hand them to a human than to stay silent and miss something. The labor saved by this system is a reduction in "time spent searching for violations," not a reduction in "the responsibility to confirm."
04A Gap in the Explicit Text Is Not "Permission"
Norms always contain gaps. Expressions and structures written nowhere in the law, the standards, or the guidelines can be produced without limit. The argument "it isn't written, so it's free" seems plausible at first glance.
This system does not adopt that argument. Areas not covered by explicit text are treated not as permitted territory but as the domain of autonomous judgment. The absence of explicit text does not mean the authors of the norms permitted the expression. They may simply not have anticipated it. The axiom of asymmetry between creator and reader holds regardless of whether explicit text exists. So even in the gaps of the explicit text, room remains for readers to form a false picture.
This thinking translates directly into how judgments are handled. A material that produces no hits against the explicit norms is not treated as "no problem" at that point. No explicit hit means not a confirmed white but a gray candidate, which is passed on for evaluation by the public gaze (= the critical perspective of society — patients, healthcare professionals, the media, and others). Only when no concern emerges there is the system's "no problem" judgment finalized. This finalization occurs within primary screening; it does not mean human review is omitted. The judgment flow itself is built this way for one purpose: gaps must not pass through unexamined (the flow is detailed in the execution-process article).
05Philosophy Becomes Enforcement, Not Wording
Anyone can write out a list of design principles. The question is whether there is force to stop a principle from being broken at execution time.
This system has a discipline called the translation gate (= a checkpoint that admits into implementation only principles that have been rewritten as constraints a machine can inspect). A principle is allowed into implementation only after being translated into a hard structural constraint that a machine can check. A principle that cannot be translated is rejected, however admirable it may be. Conversely, every principle that remains in the implementation takes a form the machine refuses to let anyone break.
One example will suffice. Principle 4, Verifiability, is not an attitude of "let's have grounds." A finding lacking grounds is rejected by type checking (= the fixed data shape a program will accept) at the very moment it is generated, and cannot exist inside the system. The same holds for Principle 5, Priority on Preventing Misses: any operation that drops a finding carries a mandatory audit record and can always be verified afterward. For every principle there is one corresponding structural constraint, and for that constraint one checkpoint that inspects whether it is upheld. This one-to-one correspondence of principle, structure, and verification is the backbone of the design.
Which principle becomes which component, and when each fires during a single run of review, is covered in the component and execution-process articles that follow. The conclusion of this article is short. This machine was built not for efficiency but to prevent misses. And final judgment, and responsibility, remain with the human reviewer.