The Stacked Deck

9 panels · 3 across × 3 down · LEAD: DR. TRACY GRANT

Row 1 — the trial

Night. Tracy alone with a printed Table 1 — four columns, one of them ringed in pencil.

Four arms. Two clinics. One randomized allocation list, like a shuffled deck of cards, drawn up back in the spring by a man two states over I've never met.

The same column, read again.

And one of those arms kept getting dealt people who were sicker to begin with. Not by a lot. Just enough to notice.

Lady Luck can do a lot of things, but she can't do the same thing six times in a row and then just whistle on the way out.

The allocation page on screen, refusing her. Her hands off the keys.

THIS STUDY CONCEALS ARM ASSIGNMENT.
ONLY THE STUDY STATISTICIAN CAN READ THE ALLOCATION ORDER.

First thing you check is the deal itself. A bad shuffle would explain everything and blame nobody.

Four cards face down on the table, and not a one of 'em I'm allowed to turn over.

Row 2 — the case, and the break

Tracy still in the chair. Past the open door, the corridor is dark.

Now, there's a switch down in the basement that'd bring the house lights up on the whole act. Every card, every name, six months of 'em.

Trouble is, the watchdog's down there too. And Fido's barking wakes up half the neighborhood.

Her hands. A pencil and the back of a parking memo.

Fine, I'll try it myself. Four suits, four arms. Here we go. Clubs... Diamonds... Hearts... and hey, now the fourth one's obvious, even dealt face down!

Which means somebody at that second clinic could deal that card to someone who needs help. And who would ever notice?

The monitor, the only light in the room.

ALLOCATION ORDER · QUARTET — SEALED
VIEWS 0 · DOWNLOADS 0 · LAST OPENED —

So who turned the whole deck face up? Now that's something I can easily find out!

Uh oh. Six months and not one bark out of Fido. Looks like I'm back to square one.

Row 3 — the answer

Light. The same table as panel 1, now drawn twice — one per clinic, side by side.

That whole theory was just one big misdirection. What if I look at this another way? Suppose I split these columns by clinic. And suddenly there it was, as obvious as an ace up a short sleeve!

—

Two clinics. One list. The second one's the specialty clinic. Sends us the sicker folks, and in big bunches! That's nobody's doing, that's just how it is.

And a big bunch can get sawed in half, but never down the middle.

Next morning. Connie in the doorway with two coffees.

Connie. Who put that watchdog down in the basement?

You did, Tracy. Back in March. Then you complained about it for a solid week.

Spend the whole night hunting a phantom magician. And in six months, the only hand that dog ever had to growl at was mine.

THE CASE OF THE STACKED DECK

Closed by asking where, not who.

What has to be true

This is the refusal: she opens the allocation order and the software says no. The cards are hers, not the artist's. She calls it a deck because that is how she thinks — nothing anywhere on this page is drawn as a playing card, a cup, or a stage. The narration is the metaphor; the drawing is the office.

1·2·7

It's a Table 1 — the baseline characteristics table every trial paper opens with, and the one whose job is to prove randomization worked. Columns are the four arms, rows are characteristics, each cell is one number about a whole group:

                       A           B           C           D
Randomized            41          40          41          40
Age, mean     58.2 (9.4)  57.6 (8.8)  58.9 (9.1)  57.4 (9.6)
Women, %              62          59          64          61
Severity       5.9 (2.1)   6.0 (2.0)   6.7 (2.2)   6.1 (1.9)

The ringed column is C, and the ringed number is that severity sitting most of a point above its neighbors — about a third of a standard deviation, which is small enough to live with once and impossible six times running — the whole case, in one cell. No names, no ID numbers, no row belonging to any one person. In panel 7 it becomes two of these side by side, one per clinic.

This is what lets her work at all: a number about forty people identifies nobody, so she can read it on a blinded study. She never sees the allocation order — who got which arm — and drawing a spread of face-up cards or a roster would show her doing the one thing the whole page says she cannot. The card talk lives in the captions only.

The switch is never drawn — it lives in dialogue, in a basement, which is doing the work of a setting you have to go out of your way to reach. The panel is the refusal, and it is not willpower — she declines because the act would be noticed. That's the whole product argument: a protection that depends on everyone behaving is not a protection. The caption never says audit, name, timestamp or role — a dog barking at two in the morning carries all of it. The dog is introduced here, catching her, and in panel 6 she goes and asks him about everybody else. Same log, both directions.

She is writing, not dealing. The caption names suits, so the pull will be to draw playing cards — don't. There is no deck in that office. Pencil on the back of a parking memo, and whatever she scratches down is hers: letters, tallies, four short columns. The suits are in her head, same rule as panels 3 and 7.

First purple anywhere. The screen names an absence, not a fault — the three zeroes are the evidence.

Nothing before this panel may hint the watchdog is hers — in panels 4 and 6 she speaks of him as somebody else's. Connie's line is the reveal the page has been holding.

What this page argues

The culprit is an unstratified list. One list feeding two clinics that were never the same clinic. The specialty site sends sicker people and sends them in bursts, those bursts land inside permuted blocks, and the arms come apart with nobody doing anything wrong. The fix — stratify by site — is real and supported.

The log does not exonerate anybody, and the page must not claim it does. Panel 5's leak works by counting assignments everyone can already see, and counting leaves no trace. What the log kills is her one checkable lead, which is why panel 6 ends at square one rather than in relief. The exoneration comes from the explanation in row 3: once two clinics on one list accounts for the table, nobody needs to have done anything. The log's value is speed, not proof.

What we don't claim. The software neither prevented the imbalance nor found it — she found it with a pencil. Fixed blocks stay predictable to anyone counting; our own default is arms.size * 2, harder to call and still fixed.

Nobody is a villain. The second-site enroller never does anything and is cleared completely. The only person altered is Tracy. The radio version has no log, so she never clears her, never stops suspecting, and never reaches the sort that would have explained it — same case, opposite ending.

Series rules live in the Casebook.