Juno — start with the contradiction, because it is the only finding here that costs you something before you open your mouth.
The full case says 6/6 pharma partners committed. The teaser, the ungated page a recruiter reaches first, says all six companies adopted it and stamps 100% partner adoption on it. Those are two different events. Neither page says which one occurred, on what date, or how anyone measured it.
Everything else in this audit is thinness, and thinness you can talk through live. You cannot talk through two of your own pages disagreeing.
What moved since July 24
Not a first fetch. Issue #5 logged the live case at roughly 214 KB on July 24. Today: 214,343 bytes, same section architecture, an August 1 last-modified stamp that may be nothing more than a redeploy. No text diff is possible. July preserved the size, not the words. So today is your text baseline.
Access model unchanged. The .html address redirects to the extensionless route, which ships the entire protected document to the browser and then runs an inline script that checks local storage and bounces anyone unauthorized to a login page. Developer tools read the case without the password. Known condition, not a new one.
Two smaller deltas. The login page now carries a request-access email path next to the password field, which closes most of July's broken-conversion finding; I cannot date the repair. And the teaser still declares a social preview image, og-cs-02.jpg, that still 404s. Ten-minute fix, open a week and counting.
The chain, link by link
32 interviews to a root finding. Thin at the input. Unusually strong at the output.
The Problem section reports 32 interviews and one root finding: Thermo Fisher and its partners had no shared source of truth. Then it earns that finding at a level of operational specificity most portfolio cases never touch. Half of orders placed under 90 days before commitment. On-time delivery running 63% to 99% across sites, with both sides computing that number by different definitions, so nobody could agree on what late meant. Forty-five percent of the relevant data sitting outside the system of record in spreadsheets and PDFs. Exceptions surfacing two to four days after commitment dates, when rescheduling costs three to five times what prevention would have.
Properly evidenced diagnosis. The research behind it is missing: no participant roles, no indication of which of the nine sites were represented, no protocol or synthesis artifact, nothing on how thirty-two conversations became five failure modes. The number 32 is currently doing the work a method should be doing.
So someone asks how you got from thirty-two conversations to five modes, and the answer lives only in your live recall at minute nineteen. You will answer it well. You should not have to.
Failure modes to modules. The strongest link in the case and stronger than most portfolios manage anywhere.
The Solution section says outright that the research failures became the design briefs, then shows the mapping in five rows. Exceptions buried in a flat list of 197 orders become an exception-first default view. Divergent on-time-delivery math becomes internal and partner metrics placed side by side. Capacity discovered after commitment becomes an eighteen-month multi-site utilization view placed before commitment. The batch lifecycle board and the structured forecast pipeline derive the same way. Each row lands in its own module section with a mockup and numbered annotations. Nothing left to inference. Do not touch it.
The risk here runs backward from what you would expect. The mapping is clean enough that the modules read as inevitable. No alternative appears, and nothing in the story turns out to have been the wrong answer first. At Senior Staff or Director, a derivation with no visible wrong turns reads as a case reverse-engineered from its outcome. Remedy: one line under the mapping, attached to the exception-first default, because somebody in that room argued for keeping the full order list visible. Name the direction. Name the finding that killed it.
Adoption as the success metric. Weakest link. Most probed.
Neither surface defines adoption. Contract commitment? Onboarding complete? Production availability? Active use by all six partners across nine sites across all five modules? No measurement date, no sustained-use window, and no active-user count, transaction volume, module penetration, or pre-launch baseline. Issue #5's Design Context File treated commitment as the canonical Thermo Fisher figure. I am holding that until use can be evidenced.
The agentic vision. Present. Internally inconsistent.
The closing section turns the five modules into agents running continuously: exception classification, document-compliance monitoring, forecast-deviation detection, advance routing. The text says five agents. The mockup beside it shows three. Nothing tells the reader whether the other two are dormant, merged, or out of frame.
Then the gate. An agent reports five release criteria complete at 99% confidence and a human approves batch QA release. The five criteria go unnamed. The confidence number has no stated source, no model or rule basis, no threshold, no exception distribution, and no account of what evidence would block a release rather than clear one.
The four standing numbers
| Metric | How it appears | What's missing |
|---|---|---|
| $20M+ margin | Full case: margin recovered annually. Teaser: a $20 million opportunity. | Calculation, baseline year, attribution across modules. |
| 6 pharma partners | Well contextualized against nine sites and three regions. | Partner functions, commitment date, rollout order. |
| 100% adoption | Teaser only. Absent from the full case. | A definition of adoption and a measurement date. See above. |
| 12-month timeline | Framed as zero to live. | Chronology of discovery, build, pilot, deployment; a definition of "live." |
Recovered and available are not the same claim. The conservative one is sitting on the surface with the wider distribution.
Also on the page: 83% IRR and a 42% overhead reduction. No baseline, no method, no period, no causal attribution on either. Four unattributed figures read to anyone finance-literate as decoration, not measurement.
Three questions the page cannot answer at leadership altitude
Your team line is one designer and two BCG Digital Ventures engineers. So the multiplier story cannot be headcount. It has to be criteria: judgment you encoded that other people then applied without you in the room.
"What design judgment survived after you left the room, and who used it?" Nothing on the page answers this. Add one paragraph to the Solution section in this shape: I set the rule that [criterion]; the BCGDV engineers applied it to [surface] without review. Then the artifact showing the reuse. One example clears the bar. Zero is disqualifying at this level.
"When two partners wanted different definitions, who decided, and by what rule?" You have more than you are claiming. The case shows Thermo Fisher seeing the full partner network while each partner sees only its own dashboard. It shows internal and customer metric definitions deliberately juxtaposed. It shows metric discrepancies escalating onto quarterly strategic review agendas. Those are governance mechanisms you designed and you are presenting them as features. Reframe one as a decision you owned: the conflicting requirement, who held the decision right, what you recommended, the rule that resolved it, what the losing side gave up.
"What did twelve months force you not to build?" Nothing forced appears anywhere on the page. The tensions are visible in the work. External visibility against exposure of operational detail; exception-first defaults against access to the full order list; automation against release authority. All three resolve silently. Two lines: what you did not ship, why, and the consequence you accepted.
The regulatory claim you're underselling
Six partners, nine sites, three regions, and a quality-assurance release gate. The teaser names good distribution and good manufacturing practice generically. That is the least specific available version of the most valuable thing in this case.
I cannot assign a single regime from public evidence and neither should you. The case names no destination market, product authorization, or transfer type, and what the interface actually shows is batch release under manufacturing quality rules rather than package-level serialization. Leave track-and-trace language alone.
What you can name precisely, if it is true of your program:
- United States. The quality unit's authority to approve or reject covers product manufactured under contract by another company, with production records reviewed and discrepancies investigated in writing before release (21 CFR 211.22, 211.192).
- Europe. A Qualified Person certifies and releases each batch into a register retained at least five years (Annex 16; Directive 2001/83/EC Art. 51).
- Australia. In 2021, release for supply ran through an authorized person on a different model.
Verify which release paths your program actually touched before any of this leaves your mouth. And do not project today's Australian framework backward onto a 2021 workflow.
If it holds, the design consequence is a single sentence: the same batch record had to be legible to three differently constituted accountable signatories without collapsing them into one generic Approve button.
Handoff 04 and the sentence that's missing
The Trust essay names Thermo Fisher in its Decision Gate section. The governing formulation:
"Automate what you can verify. Keep a human where the outcome is irreversible."
Five agents monitoring batch status, exceptions, and partner delivery across nine sites; QA regulatory release non-negotiable.
The case shows that instance on its own terms. Five agents, one human gate, a regulatory signature. Pattern legible without the essay. The reasoning does not carry over. The essay's argument is that the gate is irreplaceable because the outcome cannot be reversed and a named individual is legally accountable for the signature, not because a human happens to be standing there. Issue #5 put it this way: gate placement reveals what a product considers irreversible. Your case shows the placement and withholds the reasoning.
One sentence, directly under the QA approval mockup: This gate stays human because an erroneous release is not reversible downstream and a named individual carries legal accountability for the signature. Separately, I found no link from the case to the essay. If one exists, it is not near the agentic section, which is the only place it earns anything.
Ranked by interview risk
- The adoption contradiction. Choose commitment unless you can evidence use. Teaser becomes "all six partners committed and went live." Case outcome row gains: adoption measured as [partners in production use of all five modules] as of [date], sustained [period]. Bracket only what you can verify.
- The multiplier gap. One paragraph, Solution section, in the shape above. This is the question that ends a Director-level conversation.
- The regulatory specificity. Upside, not defect. Verify your release paths, then write the three-signatory sentence into the case.
- 32 to 5. Two sentences in the Problem section: participant functions, and the synthesis step that produced five modes.
- The 99% figure. Caption naming the five criteria, the confidence source, and one condition that halts a release.
- Five agents or three, plus the missing link to the essay. Fifteen seconds each.
- Where the $20M comes from. One clause: baseline, period, modules the recovery attributes to.
Two links hold. The third thins. And nowhere on the page do you establish that you were working above your own hands. That is a writing problem, not an experience problem, which is the only good news an audit like this ever hands you.
-
Front wants exception routing: The live Head of Product Design posting at Front covers queues, routing, rules, automation, and trust preservation, which is Thermo Fisher's exception-first architecture stated in another vocabulary — go read the wording before you rewrite the Solution section, because it tells you which of the five modules to lead with.
-
Abridge already ships your gate: Its Staff Product Designer role centers "Linked Evidence," which maps generated summaries back to source for auditability, and that is the exact caption your 99%-confidence mockup is missing.
-
Autonomy and oversight move together: Anthropic's agent autonomy research found experienced users auto-approving more often and interrupting more often, which complicates the Watch → Verify → Delegate ladder and gives you a sharper answer than "we kept a human in the loop."
-
The vocabulary is converging: Rubrik's Agent Cloud materials sell enterprise agent control as auditable, attributable, and reversible — three words that describe your QA release gate better than your case currently does.