You believe you did a solid job. But quietly dropping the primary endpoint, or staying silent about a conflict of interest, is nearly invisible from your own seat. One grader is not enough. Only by layering four viewpoints — self, reviewer, requester, and AI — do the deviations of omission come into view.
Don't trust a health check to one doctor alone
Think of a medical check-up. One person draws blood, another reads the images, and a third makes the final call. Why split the roles? Because one set of eyes misses things. The technician who took the X-ray may see "nothing wrong," while a separate reader catches a faint shadow. Measuring the ability of someone who makes materials — the pamphlets and explanatory documents handed to doctors and patients — works the same way. A single grader always leaves a blind spot.
So far this series has measured ability from the evidence of finished work and behavior. Volume 8 asked how far we can trust a single judgment. This volume goes one step further: instead of one judge, we layer several viewpoints that stand in different places — self, reviewer (the role that inspects the material), requester (the sales or planning side that ordered it), and AI (a machine pre-read). The aim is not to average the scores into a smooth middle. The opposite, in fact: we read the gaps between the four as information.
Agreement is not reassurance. The gap is what points to where a deviation hid from a single pair of eyes.
Each of the four eyes sees something different
Picture serving a meal. The cook cares most about taste. The customer (requester) checks whether it matches the order and arrived on time. The hygiene inspector (reviewer) checks whether it was made by the proper steps. And the ledger and receipts (AI and records) quietly log where the ingredients came from. The same dish looks different depending on where you stand.
Back to making promotional material, the four eyes line up like this.
| Viewpoint | Sees best | Structurally hard to see |
|---|---|---|
| Self (author) | Intent, craft, what they meant to convey | What they did not say (omission) |
| Reviewer (the one checking the material) | Deviations from rules and standards, overstepped wording | How the material is actually used and received in the field |
| Requester (the sales or planning side) | Fit to the request, deadline, usability | The pull toward "wanting to sell" (motivated reasoning) |
| AI and records (source matching, version control) | Match to sources, diffs from past versions, risky phrasing | Context, the other party's situation, whether an analogy fits |
The key point: none of the four is superior. Each has strengths and weaknesses, and one eye covers another's blind spot. That is what "layering" means.
Omission does not appear to the author's own eye
No matter how many times you reread your own proof sheet, you struggle to notice the one line you forgot to write. You see what is on the page, not what is absent. The most dangerous deviation in making material is exactly this omission — not telling, not showing, not saying until asked.
The cases reported in the monitoring project show this pattern plainly. In one reported case, the author prepared no material for the primary endpoint (the single most important, pre-specified yardstick of a trial) and explained only a secondary endpoint that happened to show a significant difference (a difference hard to explain by chance). In another, no presenter at a seminar disclosed conflict of interest (COI — a financial tie between presenter and product), and when asked why, answered "because no one asked." In both, the author feels they told no lie. They simply did not say it; they simply were not asked.
This "I just didn't say it" is hard to catch by self-grading, because the author has dropped the missing item from awareness. Here is where layering viewpoints earns its keep. Trace the real cases through the four eyes.
| Deviation (reported case) | Author's awareness | Which eye stops it |
|---|---|---|
| Skips the primary endpoint, explains only the significant secondary one | Weak sense of cherry-picking (local rationalization) | Reviewer notices the missing required item / AI matches trial-plan primary item against the material |
| No presenter discloses COI: "because no one asked" | An omission, so it looks like no problem from the author's seat (externalizing responsibility) | Reviewer checks the disclosure rule / records mechanically detect that the disclosure field is blank |
| Only non-inferiority data exist in the new indication, yet "superior" is claimed using the old indication's comparison | Pulled by the request's aim to look strong (motivated reasoning) | A reviewer who knows the requester's aim spots the context switch / the author realizes only when told |
The table shows that omission and context-switching pass almost untouched through the author's view, while the reviewer, records, and AI catch the same deviation from a different angle. What one person misses, one of the four eyes picks up.
Read the gap, don't average it
Suppose in a driving test the instructor's score and the driver's self-assessment disagree. Averaging to a midpoint defeats the purpose of the test. What matters is why they disagreed. The driver felt they did well in a spot where the instructor deducted points — that spot is the driver's blind spot.
Viewpoint integration is the same. Where all four agree, the item is mostly safe. The problem is the gap items. A gap means one of three things. First, the author's blind spot (a deviation, like omission, invisible to them). Second, over-reaction by the reviewer or AI (flagging a risky word without knowing the context). Third, a gap in the yardstick itself (which of Volume 5's three yardsticks — accuracy, clarity, balance — each person weighs more heavily differs).
The role of four viewpoints is not to add graders. It is to find the one gap and sort it into blind spot, over-reaction, or a difference of yardstick.
Sort in this order. First, for the gap item, check whether you can return to the source (grounding) — the absolute floor repeated in Volumes 6 and 8. If the source does not support the author's claim, it is confirmed as the author's blind spot, a deviation. If the source does support it and the reviewer's flag was a misread of context, record it as the reviewer's over-reaction and adjust the yardstick. The AI's role is clear here: AI is fast at mechanically picking up source matches, version diffs, and risky phrasing, but it cannot know context or the other party's situation. So treat AI flags as a list of candidates for a human to check, and leave the final judgment to a person.
Who pulls the layered views together
An orchestra has a conductor. Each player (each eye) plays their own notes correctly, but shaping the whole is a separate role. Viewpoint integration also needs someone to compare the four assessments and bind them into a final call. The caution: that person must not be the author. Self-approving your own work means the generator doubles as the evaluator, and the blind spot of omission stays right where it was.
A practical way to bind them, in light steps. First, place the four assessments in one table and split agreed items from gap items. Second, take only the gap items and sort blind spot from over-reaction by whether you can return to the source. Third, record confirmed blind spots as deviations and feed them back to the author — not to condemn. As this series has said since Volume 2, the aim is for the author to monitor their own internal circuits (motivated reasoning, local rationalization, omission, externalizing responsibility). A blind spot seen once through four eyes is one the author can catch alone next time. That is the real effect of integration.
One last word on the floor. Even with four eyes, the principle holds: skill at persuasion cannot fill a hole in source grounding. The most dangerous thing is a small misperception hidden in a convincing document — high design, low fidelity. If even one of the four eyes reports "cannot return to the source," then no matter how high the other three score it, the floor is broken there. The whole point of layering is to never miss that single warning.
Measuring Skill from Work and Behavior ── Map of all 10 episodes
- Vol. 1: Measure by the Materials Actually Made, Not by Impressions or Self-Report ── A material maker's skill is measured from the actual deliverables and observable conduct, not from self-report or others' impressions.
- Vol. 2: Tracing the Brief, the Choices, and the Result — In Order ── Read a creator's skill from evidence by walking through one real project in order: the brief, the thinking, the actions, and the result.
- Vol. 3: Reading "Faithfulness to the Facts" and "Craft of Delivery" Out of the Work Itself ── This installment shows how to recode a finished piece into two axes — faithfulness to the facts and the craft of getting it across — by reading concrete clues, not impressions.
- Vol. 4: The Rules That Keep Measurement Honest ── Six ground rules that keep the evaluator from drifting when measuring an author's real skill.
- Vol. 5: Three Rulers: Accuracy, Clarity, and Balance ── Defines three rulers for grading material-making skill and scores each on a four-step scale: accuracy as the floor, clarity as the reach, and balance as the adjustment between too much and too little.
- Vol. 6: How to Decide the Level — Returning to the Source Sets the Ceiling ── Work that cannot be traced back to its source cannot earn a higher level, however polished it looks. Grounding sets the ceiling.
- Vol. 7: What Deliverables Signal Which Level ── An anchor table that reads a creator's level (L1-L4) from visible deliverables and behavior patterns.
- Vol. 8: How Far Can We Trust a Judgment? ── How sure a level judgment is depends on how visible the evidence is; less observable skills produce shakier judgments, so we attach a confidence to each verdict.
- Vol. 9 (this episode): Combine More Than Self-Assessment: Add the Reviewer's and Requester's View ── Layering four viewpoints — self, reviewer, requester, and AI — surfaces the deviations of omission that a single pair of eyes cannot see.
- Vol. 10 (final): Connecting the Measurement to Pass/Fail and a Development Plan ── The finale links the score to the pass floor and a plan for what to grow next.
The point of layering viewpoints is not the comfort of more graders. It is to catch, from eyes that stand at other angles, the omissions a self-grade structurally cannot show — dropping the primary endpoint, not disclosing COI, switching the context. So read the gap, not the agreement, and sort each gap by whether you can return to the source. If even one eye warns "cannot return to the source," the floor is broken there, however high the rest score it.
And someone other than the author binds the final call. A blind spot seen once through four eyes is one the author can catch alone next time. Integration is not a surveillance device but a mechanism that lets an author audit their own circuits. The final volume turns to how these measured results connect to pass-or-fail decisions and development plans.
- Omission is invisible to the author. Unsaid deviations — dropping the primary endpoint, not disclosing COI — are structurally missed by self-grading, so eyes from other positions are needed.
- Read the gap, not the agreement. The four viewpoints exist not to add graders but to find the one gap and sort it — by whether you can return to the source — into blind spot, over-reaction, or yardstick difference.
- Someone other than the author binds the call. Self-approving leaves the blind spot intact. Treat AI flags as candidates for human review, and keep the final judgment with a person.
- Ministry of Health, Labour and Welfare, Compliance and Narcotics Division (commissioned project). Report on the Monitoring of Promotional Information Activities for Prescription Drugs (March 2024 and prior years). Flagged cases are published with company names anonymized; the deviation patterns cited here are generalized from this report.
- Japan's Ministry of Health, Labour and Welfare "Standards for Fair Advertising of Drugs" and its commentary — the framework for judging comparative, exaggerated, and omission-based advertising expression.
- Japan Pharmaceutical Manufacturers Association "Promotion Code for Prescription Drugs" — a shared industry baseline on information activities and COI disclosure.
- General methodology literature on the Behavioral Event Interview (BEI), the STAR method, and multi-rater evaluation (combining several viewpoints).