Nothing moved. junochen.com/cs-06-alle.html returns the file I pulled on July 24, byte for byte: 254,448 bytes, last modified July 19. So the top finding from that audit stands as written. This case presents two strong projects, not one system.
Two updates. One item I underweighted in July now sits above everything that audit ranked. And one claim I got wrong, corrected at the bottom.
The URL redirects to /cs-06-alle and hands the complete document to an unauthenticated request before any access check runs. A script then finds nothing in browser storage and bounces the visitor to a login page. Consequences are in Trust Calibration Starts at Home. Pick private or shareable, then make the code agree with the pick.
Everything below comes out of the retrieved text. Nothing from memory.
The altitude this case actually flies at
Senior readers sort a case into one of four altitudes, usually inside a minute.
- Feature. One screen, solved.
- Product. A surface with internal logic that holds.
- System. Separate surfaces running off one underlying model, with a visible account of how control is split between them and why that split serves the bet.
- Portfolio. An org that produces system work on repeat.
Allē runs at product altitude twice.
The headline numbers and the teaser language both claim system. The body pays for neither. But the evidence is sitting on the page already. The ordering hides the argument, which makes this cheap to fix and expensive to leave.
Stakes past this one page: Allē is the only case in the set carrying a dual-surface claim. If it reads as two products, the portfolio has no system-altitude entry anywhere in it.
Criteria below run in fix order, not numbered order. One and four are the same finding, so they travel together.
First: define the scale claim or cut it
The case reports 30M+ members and 40K+ practices. AbbVie does not.
| Source | Members | Practices |
|---|---|---|
| Case study, as published | 30M+ | 40K+ |
| AbbVie, Jan 2024 | 6M | 19,000 |
| AbbVie, Apr 2026 | 8M | 30,000 (U.S.) |
That gap is not rounding.
Reconciliations that hold up do exist. Cumulative enrolled accounts against current actives. A practice count that sweeps in historical or non-exclusive locations. I don't know which one applies here. Neither will a hiring committee, and someone on it will surface AbbVie's number in a single search, because it lives in a press release.
Add a definition, not a hedge. One parenthetical under the scale block: what the number counts, as of when. If you can't define it precisely, cut it. Eight million you can defend beats thirty million you can't, and the downside is lopsided. A scale claim that fails a reference check takes the rest of the case down with it.
Two smaller contradictions while the file is open.
- The date. The full case dates the work to 2020. The teaser says 2021.
- The $92. In the full case, customer acquisition cost falls from $92 to $42. In the teaser, $92 is an industry average. One number, two incompatible meanings, on two pages a reader opens in adjacent tabs.
The premise is stated. The tracks never meet.
Section 02 promises "two tracks and six design decisions." What follows is three consumer decisions, four consumer visual explorations, and a reactivation section. Provider opens after all of it.
So the reader finishes the consumer story before the provider surface exists in any operational sense. Practice checkout shows up most of a page-length behind the patient's expiring points, and no sentence on the page ties the two together.
That single ordering choice generates most of this audit. Surfaces that never meet can only justify decisions locally, which is why the rationale reads thin. Outcomes then get attributed one surface at a time, leaving the largest number with nowhere to land.
Insert a premise section between 01 and 02. Roughly 120 words, establishing the shared event before either track opens. The program mechanics carry the claim honestly on their own: Allē's own documentation describes a member holding points, identifying at checkout, and a provider logging the treatment that moves the balance. One event. Two interfaces.
Read every block of copy below as a shape, not a script. Verify each factual claim inside it before it ships.
One event drives both surfaces: a logged treatment. When a member is treated, the same record updates the wallet the patient sees and the checkout the practice sees. Every decision below sits on one side of that event. The consumer track designs what the member understands before and after it. The provider track designs what the practice can act on at the moment it happens. Two role-specific views, one loyalty state. When the two disagreed, the practice's operating reality won, because a rewards program a busy front desk cannot execute is not a rewards program.
Then stop stacking the tracks and pair them. Take the three strongest shared events, points expiry, tier proximity, post-visit drop-off, and show both views of each. A·01 already builds an expiry countdown for the member. B·03 already builds a lapsed-patient campaign filter. Same event, sixty days apart, and the page never says so.
The teaser makes this argument already: "Two surfaces. One loyalty system. Both live," running off a single information model. Move that language into the body, where the evidence lives.
Every block gives a solution. None gives a rule.
The decision blocks explain what got built and why it helped. None shows a fork: no tested alternative that lost, no engineering, legal, timing, budget, or staffing constraint named anywhere in the file. Eight visual explorations, Noir through Canvas, sit there as artifacts, and the reader never learns which one shipped or what killed the rest.
One test per block. A decision counts only when it's stated as a rule another person could inspect, contest, or inherit. "We showed points expiring" is a screen. "Warn at 60 days, never at 7" is a rule someone can argue with.
Four lines each on A·01 and B·01. Considered, chose, because, rule.
A·01. Considered surfacing expiry at seven days, when urgency peaks. Chose sixty, because seven days does not fit an appointment calendar, and an urgent message the member cannot act on reads as a threat. Rule: never fire a deadline the member cannot clear.
B·01. Considered letting each practice configure its own sort order. Chose a fixed default. Rule: the list opens on the patient the practice will lose first. Configuration is a preference, not the product.
The second one distributes control. Who can change what, who decides. That's the other half of the system definition up top, so say it out loud rather than leaving the reader to infer it.
Reposition the visual explorations. One sentence above each set naming which direction shipped, what it was evaluated against, and what the losing directions optimized for that turned out to be the wrong thing to optimize for.
Vanta's Head of Design search asks for organization-wide product quality mechanisms. Rules stated this way are mechanisms. Screens are output.
Two of three numbers land. The biggest one doesn't.
The 47% expiry open rate and the 68% return-reason figure connect cleanly to the consumer expiry work. The 2.4x conversion and the CAC movement connect to in-app planning. Those hold.
The 3.2x redemption lift connects to nothing. It's the largest claim on the page and it hovers above both tracks unattached, while no provider decision anywhere carries a measured outcome. That combination is exactly the evidence shape a reader expects when the provider work was secondary. Attribute it to the shared architecture, or de-attribute it on purpose:
3.2x redemption lift is a program-level result across both surfaces during the period this work shipped. I cannot decompose it by surface and I am not going to pretend otherwise.
That sentence converts an unattributed number into a demonstration of measurement discipline, which reads as senior. Silence reads as hoping nobody asks.
The leadership record is a roster
The hero names the role and the team: three product designers, plus UX research and brand. That is the whole of it. The reporting line, the allocation across two tracks, the design review mechanism, the stakeholder conflicts and how any of them got settled, whether the dual-surface mandate was inherited or argued into existence: none of it is on the page.
Three sentences close the gap. Shapes:
Structure. "Three designers across two surfaces. I ran consumer myself and staffed provider with one designer plus shared research, which meant the provider visual system inherited consumer's tokens rather than earning its own. That was the trade I made."
Conflict. "Marketing wanted reactivation in Allergan's voice for brand consistency. The practices wanted their own. I sided with the practices and lost the brand-consistency argument on purpose, because the patient's relationship is with the injector."
Mandate. "Provider was not in the original scope. I argued it in after the consumer research showed the failure point sat at checkout, not in the app."
If that isn't what happened, write what did. But write something. One named trade you lost is worth more to a committee than a headcount.
Amplitude's Head of Product Design posting is an explicitly production-involved player-coach role. This case proves the player and asserts the coach.
Regulated context, plus a correction I owe you
On July 24 I wrote that prescription-marketing rules, HIPAA, and state referral statutes explain why B·03 sends outreach in the provider's voice rather than Allergan's. That was inference dressed up as finding. Nothing on the page or in the public record establishes those constraints drove that decision. Good hypothesis. Not evidence. I should have labeled it.
Here's the boundary I can defend: Allē is healthcare-adjacent work touching prescription drugs carrying boxed warnings, prescription medical devices, licensed injectors, treatment records, and consumer promotion of regulated products. Unusual context. The case spends none of it.
Claim the constraint. Don't claim compliance work you didn't do. Write the paragraph below only if the review actually happened, and describe only the review that actually happened. I can't verify it from the page or the public record, and you can't verify it from memory, so check it against a project artifact first. If it happened, this is the shape:
Every patient-facing message that referenced a treatment went through brand and regulatory review before it shipped. I designed the expiry notification to carry urgency without making a claim about the product, because the product is a prescription drug and a rewards message is not the place to describe one.
Abridge is hiring against consequential clinical workflows with auditable linkage back to source evidence. One paragraph like that makes you legible for that search. Without it, this reads as a beauty-app loyalty case.
Fix order
- Define or cut the 30M/40K claim. Reference-check exposure. The only item on this list that can end a process by itself.
- Insert the shared-event premise and pair three events across both tracks. This is the altitude move.
- Add considered/chose/because/rule to A·01 and B·01, and name which visual direction shipped.
- Attribute the 3.2x or de-attribute it explicitly.
- Add the three leadership sentences.
- Add one regulated-context constraint paragraph.
- Reconcile the date and the two meanings of $92.
One through four is a weekend. That gets the case to the altitude its own numbers have been claiming since July.
-
Maven wants exactly this proof: The VP of Design search at Maven Clinic spans member, employer, and health-plan audiences, which is the closest live mandate to Allē's dual-surface architecture — fix the premise section before you approach it, because that posting will read this case for whether multi-audience means one system or three products.
-
Attribution is becoming a product primitive, not a portfolio nicety: Linear's agent-assisted editing now distinguishes agent-written changes visually, preserves authorship, and creates restorable checkpoints, which is the same discipline the 3.2x fix demands — say who or what caused a result, or the result stops being evidence.
-
The eval-forward standard for "done": First Round's account of Figma's AI evaluation process describes golden prompts, nightly comparisons, and mixed human and automated judgment as ongoing infrastructure, and it is a useful model for what a rule-shaped decision block looks like at scale.
-
The gate problem is not cosmetic: All five full cases still ship complete HTML before the client-side access check runs, as documented in Trust Calibration Starts at Home — the Allē content fixes raise the stakes on this, because a stronger private case delivered publicly is a larger contradiction than a weak one.