# The Stacked Deck — case notes *Research file. The story lives in [the storyboard](comic-stacked-deck.html) and [the script](podcast-stacked-deck.html) — this holds only the feature grounding and the methods reasoning, which neither surface can carry. Nothing here restates the plot or references a panel number, so nothing here goes stale when a draft is rewritten.* **Lead:** Dr. Tracy Grant, PI. Allocation is hers — the setting, the rule, and the consequence. **Feature:** allocation concealment and the permission split. `Study#conceal_arms?` / `can_show_arms` ("Reveal Arm Assignment", default on). When a study conceals arms, `Membership#can_see_allocations?` collapses to `statistician?` — so even the PI is locked out of the deal order — while `can_manage_allocations?` stays open, meaning coordinators keep building, uploading and retiring lists they cannot read. The two acts that expose the order, the Peek modal and the CSV export, each write an audit event carrying **the role held at the time**. **Trial:** **QUARTET**. Four arms, 2 × 2 factorial, two sites. (Alternate: EQUIPOISE, which is funnier to anyone who knows the term and invisible to everyone else.) --- ## What the case rests on **The culprit is an unstratified list.** One randomization list feeding two clinics that were never the same clinic. The specialty site enrolls sicker participants and enrolls them in bursts — a run of referrals, then nothing for a month. Those bursts land inside permuted blocks, and the arms come apart by severity with nobody doing a single thing wrong. It is the most common way an honest trial produces a dishonest-looking table, and the fix — stratify by site — is real and supported: `AllocationList#stratified?`, `stratum_key_for`, and `remaining_by_stratum`, whose comment already says the quiet part out loud (*"safe for anyone to see"*). ### What Tracy can and cannot see Worth stating plainly, because it's the first thing a methodologist will ask. **She never sees an individual assignment.** What she reads are *group summaries* — an arm's average baseline severity across forty people. The rule the platform enforces is that an arm may appear anywhere no individual is identified: a per-arm number names nobody, an arm printed beside a participant does not. Splitting that table by clinic asks it a second question and every cell is still an aggregate. So blinding doesn't stop her doing the analysis. It stops her doing the one thing that would corrupt it. She answers her question completely and remains unable to steer anybody. **Concealment hides the future; masking can't always hide the past.** Two different protections, and this case lives in the gap between them. - **Staff see arms already assigned.** QUARTET has an in-clinic component, so somebody delivers it and necessarily knows. Participants aren't blinded either and they mention things. This is normal in behavioral work, largely unavoidable, and *not* a protocol violation. - **Nobody sees the allocation order.** That's what's sealed, what `can_show_arms` enforces, and what the audit log watches. Reading it hands you every assignment for the rest of the trial. **The block size is the leak between the two.** Fixed blocks of four in a four-arm trial mean three past assignments *determine* the fourth — so information everyone is allowed to have becomes information nobody should have. That is the entire argument for varied block sizes, and it is what Tracy had in front of her for six months without once multiplying the two facts together. Her discovery is therefore **the block size**, never "staff can see arms." She always knew that second part, and the distinction keeps the culprit an instrument rather than a person. **Tracy herself is blinded completely** — she reads aggregates two miles away and never needs an individual assignment. Which is why her suspect is at the clinic and not in her own building. ### The false lead, and why it's the good part Blocks of four in a four-arm trial: three cards down and the fourth has nowhere left to go. Anybody enrolling who bothers counting knows what's coming next, and nobody has to cheat for that to be true — it's subtraction. Once you know the next slot is the control arm, you can move the sicker patient to Thursday, *because you are trying to help them*. So Tracy's first answer is a person, and it is a completely reasonable answer. **The blocks are not a red herring.** The same arithmetic that frames a colleague is what explains the table: a run of referrals walking into a block of four takes whatever cards are left in it. She is right about the mechanism and wrong about the agent, which is the most useful mistake a detective can make. **And she proves the leak exists without using it** — pencil, envelope, no access required. The proof and the crime are the same act and she performs only the harmless half. > The Quiet Participant was *absence doesn't print*. > The Invisible Line was *staleness doesn't print*. > The Perfect Delivery was *sent was true, delivered was true, and both were useless*. > The Stacked Deck is **the table was right, the reading was wrong, and the difference was a sort order**. --- ## What the case argues The software neither prevented the imbalance nor found it — she found it, with a pencil. And the log does *not* exonerate anybody, because counting cards leaves no trace; what it kills is her one checkable lead, fast enough that she's still looking when the real answer turns up. *The alternative to an audit log isn't safety; it's a PI holding a suspicion she can neither confirm nor drop.* **Fair play:** the answer is physically on the desk in the opening image (a courier envelope from the second clinic) and stated in the first caption. Both must survive editing. --- ## Notes **Nobody is a villain, including the suspect.** The second-site enroller never does anything, and in the comic she is cleared completely rather than reluctantly. The only person altered is Tracy. **The narration is the metaphor; the drawing is the office.** She calls it a deck because that's how she thinks. Nothing on the page is ever drawn as a playing card, a cup or a stage — the art stays literal, or the page shows her doing the one thing it says she cannot. **Tracy is locked out, and the temptation is a switch rather than a button.** She can't peek; she could unblind the study. She doesn't, because the act would be *noticed* — not because of willpower. That the watchdog turns out to be hers is the closing reveal, so nothing before it may hint at it. **Don't let the audit read as surveillance.** It records that a thing was looked at, not that somebody was watched. Here it exists to clear people, and that is the only way it should ever be drawn — the moment our software is shown catching a colleague, this is a different product. --- ## Open - QUARTET vs. EQUIPOISE. - The radio version ends unsolved. That's deliberate and it's the best thing in the episode, but it's the first case in the set that doesn't close — worth a decision before it becomes the format by accident. - Whether *stopped asking who, started asking where* is a satisfying series signature against The Invisible Line's *stopped looking at who, started looking at when*, or one repeat too many.