What the Machine Does in a Single Review ── The Execution Process from Material Input to Human Handoff

01The Big Picture — A Timeline from Input to Handoff

A single review run proceeds as follows. Receive the material PDF and the reference documents → capture the pages into machine-readable form → deterministic matching (= checks whose pass/fail is settled by code) → semantic judgment (= checks that ask AI about impressions and implications) → cross-checking against the reference documents → a read-through by society's eyes → application of the tolerance policy → integration and scoring → audit logging → handoff to humans. Only critical judgments go on to an additional deliberation after this. Each stage receives the output of the previous stage and accumulates findings. The order matters: checks that can be settled without hesitation come first, and checks that carry heavy judgment come later.

02Fixing the Inputs — Ingesting the Material and Locking the Reference Documents

First, the system fixes both the object of inspection and what it will be checked against. During material ingestion, it determines column layouts and joins fragments; information that could not be read is not discarded but flagged "requires human confirmation" and passed downstream. Not silently swallowing missing data is the fence at this stage. The references are three documents: the electronic package insert (= the official explanatory document for a drug), the review report (= the record of the grounds on which the regulatory authority evaluated the drug at approval), and the RMP (= the post-marketing safety management plan) (the division of roles among the three is left to the article on system components). The criteria for judgment do not reside inside the material. Whatever the material claims, if the claim is not contained in the reference documents, it becomes a deviation candidate. Judgments about the approved scope take the electronic package insert as the baseline original — Principle 1 of the six principles, the principle of complementarity, is invoked here and in every scene that follows.

03Deterministic Matching — Settle the Unambiguous Parts First

Presence of prohibited terms. Whether a graph's axis starts at zero. Whether case counts meet the threshold. Whether required statements are missing. Checks like these are settled by code that always produces the same output for the same input. They are not handed to AI. This division of labor — never delegating calculation or counting to AI — protects reproducibility (= re-inspecting the same material yields the same result). Weak hits, for example a term that may be acceptable depending on conditions, are not discarded but kept with lowered confidence, and weak hits on the same item are consolidated into one representative instance annotated "N other locations." This is where Principle 5, miss-aversion priority (the principle that a false negative — overlooking an actual violation — is treated as a heavier failure than a false positive), first appears.

04Semantic Judgment — Asking What Remains in a Reasonable Reader's Mind

Next comes the stage that asks not about facts but about impressions. Page by page, the AI is asked multiple times "what remains in a reasonable reader's mind." Evaluation runs along three dimensions: local (the choice of a single word), proximal (what is placed next to it), and global (the image that the material's overall composition forms). Even a factual statement is a violation if its presentation misleads — Principle 2, the dual obligation, is invoked here. If even one of the multiple queries returns a violation verdict, that finding is adopted rather than taking a majority vote. This too is miss-aversion priority in action.

For safety, the inspection becomes asymmetric. The system first extracts the warnings, contraindications, and serious adverse reactions that the material itself carries, and bundles them into every page-level judgment as the baseline for internal consistency. If a passage summarizing safety leans reassuring without touching the most serious event among its own warnings, it is flagged as dilution (= watering down toward reassurance). Judging efficacy strictly while judging safety leniently — this asymmetry in judgment is not accepted. Principle 3, the asymmetric disclosure obligation, is invoked.

05Cross-Checking Against the Reference Documents — Has the Material Stepped Outside Approval?

In the subsequent cross-checking stage, the material's claims are compared against the approved scope, the review report, and the RMP. The review report records the efficacy and safety evaluations by the PMDA (= the Pharmaceuticals and Medical Devices Agency, the authority that conducts approval reviews). Discrepancies such as "a statement suggesting efficacy for an item the PMDA did not accept as effective" are caught at this stage. Whether important safety matters listed in the RMP are reflected in the material is also verified here.

06Society's Eyes — Not Letting Gaps in the Written Rules Slip Through

Areas that matched no explicit provision of law or standard are not thereby confirmed safe. The entire material is read through from eight perspectives (patients, families, healthcare professionals, media, social media, investors, regulators, lawyers), and concerns about ethics, dignity, and patient rights that lie short of explicit violation are inspected in a separate lane. A location with no explicit-rule hit is confirmed as WHITE in the system (= a "no problem" verdict at primary screening) only after this society's-eyes evaluation also comes back CLEAR. This is a verdict finalized inside the machine; it does not mean human review is skipped. Principle 6 — gaps in the written rules mean self-governance, refusing "not written, therefore free" — is implemented as the very existence of this stage.

07Tolerance and Integration — Leaving Grounds Even for the Act of Dropping a Finding

Nor is it right to flag everything. Statements that differ from the electronic package insert only in the dose or dosing interval of a cited clinical trial's design, or presentations of unapproved data carrying the prescribed annotation (such as an explicit note that doses outside approval are included), are tolerated under a predefined, auditable policy. However, even with annotations, extension of indications beyond approval, dilution or omission of safety information, statistical distortion, and extrapolation from animal studies to humans are never tolerated. What matters is that the operation of dropping a finding also demands grounds and a record. Suppressions, tolerances, and withdrawals are all written to the audit log and can be verified afterward.

In integration, duplicate findings on the same location are unified into the most severe verdict. They are then scored by severity × confidence × exposure (= the degree to which the statement meets the reader's eye — placement and size). If there is even one BLACK (= a clear violation), the candidate rank the system assigns automatically drops to the lowest. The final decision on whether correction is required is made by the human who receives this.

08Deliberation and Output — Pushing Doubtful Verdicts Toward the Human Side

Only critical judgments are put before a virtual panel of 60 reviewers. In the debate among experienced reviewers, stating counterarguments and minority opinions is mandatory, and a majority of the 30 authority-side votes yields a candidate verdict of "correction required." Not only split issues but issues that appear unanimous are also routed to human review — because unanimity is treated as a symptom to be doubted. After deliberation, the tolerance policy is applied once more, and issues tolerated by annotations and the like are withdrawn. These withdrawals, too, are written to the audit log.

At the point of output, Principle 4, the verifiability principle, operates in its hardest form. Every finding must carry the source text of the governing norm and the ID of the corresponding atomic item (= the smallest check item that can be inspected independently, one at a time); a finding lacking these is rejected at the moment of generation. The human reviewer receives the deviation report and the audit log (= the inputs, the versions of the rules and policies, and the path of each verdict). The handoff occurs in a state where "why this verdict" can be reconstructed all the way back to the original text of the norm. And the final judgment and the responsibility remain on the human side. The system only takes over the time spent searching; it does not take over the seat of judgment.

Note that a review does not close after one pass. When a norm is revised, the rules are updated, it is confirmed that results do not change for previously inspected materials, and already-approved materials are rescanned. Before going live, the system runs in parallel with human review, and misses on which the machine stayed silent are reconciled weekly. A single run is only one cycle within this operational loop.