Shuffle eight cards the way Edward Thorp described in 1973 — cut the deck in half, then let a coin decide which card of each pair drops first — and after three shuffles every card is perfectly random: any card is equally likely to be in any position. The deck is not. Of the 40,320 possible orders of eight cards, only 4,096 can have appeared, and the deck sits at 0.90 on a scale where 0 is random and 1 is not shuffled at all. This case computed that exactly. How many shuffles the whole deck needs was described by Persi Diaconis as the “longest-standing open card shuffling problem.” For 32 years only trivial bounds existed. One of the results OpenAI released on October 6, 2026 claims the answer in order of growth — proportional to the number of halvings of the deck — across twelve manuscripts and 683 pages, with the main statement checked by machine. A week later this case found no specialist comment, no press, and no change to Wikipedia. The only public response is an anonymous draft that tightened the proof's constant from 1,600 to 24. The case is what silence means when the answer is checked, the question is famous, and nobody has looked.
Edward O. Thorp is best known for beating blackjack: his 1962 book Beat the Dealer showed that counting cards could overcome the house advantage.[6] In 1973, studying how imperfect shuffling affects the game of Faro, he described a shuffle simple enough to do by hand.[5] Cut the deck into two equal halves. Take the top card of each half and let a coin decide which of the two drops first. Repeat for every pair. The deck is now interleaved — and the question is how many times you must do this before every order of the deck is equally likely.
For most shuffles that question has a precise answer. In 1992 Dave Bayer and Persi Diaconis showed how many ordinary riffle shuffles a deck needs, in one of the few mathematical results to reach the front page of the New York Times.[7][8] Thorp’s shuffle resisted. Ben Morris, opening his 2005 paper on it: “The problem of determining its mixing time is, according to Persi Diaconis, the ‘longest-standing open card shuffling problem.’” He added that “the only known upper bounds are trivial ones … that have circulated in the folklore.”[8] Morris gave the first real bound that year, a power of 44, and cut it to a power of 3 by 2009.[8][9]
One of the results OpenAI released on October 6, 2026 claims the end of that story in order of growth. For a deck of 2 to the power d cards, the deck is random after a number of shuffles proportional to d — the number of times the deck can be halved — and no fewer than about 2d will do.[1] The main statement is checked by machine in the proof language Lean.[3] The upper bound uses 1,600d shuffles, and the paper is direct about it: “the constant 1600 is an explicit choice in the proof, not a proposed sharp constant.”[1] It is also asymptotic: it describes what happens as decks grow without limit, not a guarantee for any particular deck. The result family runs to twelve manuscripts and 683 pages, though the main proof stands alone in fourteen.[1][2][12]
The paper opens with the paradox at its centre: “A single card can become uniformly distributed long before a shuffled deck is close to uniform.”[1] This case computed it exactly for eight cards. After three Thorp shuffles, every card is perfectly random on its own. The deck is still 0.90 from random, on a scale where 0 is random and 1 is unshuffled, because only 4,096 of the 40,320 possible orders can have occurred. It takes five shuffles to get within a quarter of random, and ten to get within one hundredth. The companion page runs the same calculation in the browser.[13] Eight cards cannot test an asymptotic theorem; they show what the theorem is about.
Computed exactly for this case. At that point every individual card is already perfectly random; the deck is 0.90 from random on a 0–1 scale.[13]
The shuffle is not only a puzzle. In 2009 Morris, Phillip Rogaway and Till Stegers showed how to use it to encrypt small values — “say, a U.S. social-security number” — into numbers of the same shape, building “a practical and provably-secure blockcipher-based scheme for deterministically enciphering credit card numbers and the like.”[10] Guarantees of that kind rest on knowing how fast the shuffle mixes. The new result concerns the whole deck and carries a constant far too large to use directly, so it changes no system today; it settles which order of growth is right.
That makes the response so far unusual. This case found no comment from any specialist, no press coverage specific to the result, and no edit to the Wikipedia articles on Thorp or on shuffling since the release.[6] The one public trace is a repository of remarks on the release, self-described as “local draft, not yet published,” with author attribution left as a placeholder. Re-tuning the paper’s own argument, it claims 24d shuffles suffice instead of 1,600d — while noting that “a remark that re-selects parameters is only as correct as the manuscript it builds on.”[11] The paper itself has not been withdrawn or revised.[4]
Across this release, silence has taken different forms. For Turán’s brick factory (UC-330), it was deliberate: an editor removed the claim from Wikipedia, asking to wait for reliable sources. Here nobody added it. A famous problem, a machine-checked statement and a practical lineage have produced, a week on, one anonymous improvement to a constant. That is not evidence the proof is wrong or right. It is evidence of how much a field can absorb at once — and of how the result most clearly checked may also be the one least read.
Bayer and Diaconis answer the question precisely for the ordinary riffle shuffle.[7]
Solved ElsewhereMorris, Rogaway and Stegers build a provably secure scheme on the Thorp shuffle.[10]
In PracticeA draft repository tightens 1,600d to 24d. No other response found.[11]
The ResponseDescribed by the mathematician who beat blackjack, in a study of the game of Faro.[5][6]
Three make every card random; five bring the deck within a quarter of random. Computed exactly for this case.[13]
The problem of determining its mixing time is, according to Persi Diaconis, the ‘longest-standing open card shuffling problem.’ — Ben Morris, 2005 [8]
| Dimension | Evidence |
|---|---|
| Quality (D5) Origin · 80 | After three shuffles of eight cards every card is random, but only 4,096 of 40,320 orders can occur. The right growth rate was open for 53 years.[1][8][13]Each Card Random, the Deck Not |
| Regulatory (D4) L1 · 76 | The main statement is machine-checked. The constant 1,600 is “not a proposed sharp constant,” and the bound is asymptotic. No specialist review found.[1][3]Checked, Not Reviewed |
| Operational (D6) L1 · 72 | Twelve manuscripts in the family; the main proof stands alone in fourteen pages; companions give 32,800d and 512d.[1][2][12]683 Pages for One Question |
| Employee (D2) L2 · 64 | The only public response is a draft remark, author attribution a placeholder, claiming 24d. No specialist comment or Wikipedia edit found.[6][11]An Anonymous Draft, Nothing Else |
| Customer (D1) L2 · 58 | A 2009 scheme enciphers card and social-security numbers with the Thorp shuffle. This result changes no system until the constant shrinks.[10]Encryption Rests on Mixing |
The cascade originates in D5 (Quality): how quickly a simple, practical shuffle actually produces randomness, open in its right order of growth since 1973.[5][8] It propagates to D4 (Regulatory), where the main statement is machine-checked but the proof has not been reviewed and the constant is far from practical,[1][3] and to D6 (Operational), where the answer spans twelve manuscripts and 683 pages, and its improvement so far comes from an anonymous draft.[2][11] It lands in D2 (Employee), the specialists who have not responded,[8] and D1 (Customer), the cryptographers and engineers whose constructions rest on mixing bounds — for whom nothing changes until the constant does.[10] D3 (Revenue) is not scored.
This is the fifth result from the same release in this library. UC-328 recorded specialists reading, UC-329 a crowd building, UC-330 a field waiting and UC-331 a prediction arriving. This one records silence around the most clearly checked claim of the five. The honest limit: this case did not and cannot verify the proof; eight cards illustrate the question, not the theorem; and silence measured in public may hide private reading.
-- UC-332: Every Card Is Random. The Deck Isn't.: 6D Diagnostic Cascade
-- Thorp 1973 (JASA 68:842-847) shuffle: cut in halves, coin orders each pair. Diaconis via Morris 2005: 'longest-standing open card shuffling problem'; folklore polynomial bounds only. Morris 2005 O(d^44); Montenegro-Tetali O(d^29); Morris O(log^4 n), 2009 O(log^3 n). Morris-Rogaway-Stegers CRYPTO 2009: Thorp-shuffle encryption of card and social-security numbers. OpenAI release Oct 6 2026, result 238: 12 manuscripts, 683 pages; main paper (14 pp) mixing in Theta(d) shuffles, upper 1600d (asymptotic, constant not sharp), lower 2d-O(1); Lean formalizes the main statement. Companion: 32800d alone, 512d combined. Anonymous draft remarks repo (Oct 10): 24d. Not withdrawn or revised. No specialist comment, press, or Wikipedia edit found. Exact computation, 8 cards: each card uniform after 3 shuffles; 4,096 of 40,320 orders reachable; TV 0.898 at 3, 0.203 at 5, 0.008 at 10.
FORAGE every_card_random_deck_not
WHERE famous_open_problem_answered = true
AND main_statement_machine_checked = true
AND specialist_response_found = false
ACROSS D5, D4, D6, D2, D1
DEPTH 3
SURFACE every_card_random_deck_not
DIVE INTO checked_but_unread
WHEN proof_scale_pages > 500
AND constant_practical = false
TRACE silent_absorption_cascade
EMIT every_card_random_deck_not_signal
DRIFT every_card_random_deck_not
METHODOLOGY 88
PERFORMANCE 38
FETCH every_card_random_deck_not
THRESHOLD 1000
ON MONITOR CHIRP high 'An AI-written, machine-checked proof claims the right order of growth for the Thorp shuffle's mixing time, called the longest-standing open card-shuffling problem. Twelve manuscripts, 683 pages. A week later the only public response is an anonymous draft tightening the constant from 1,600 to 24. The proof is unreviewed.'
SURFACE analysis AS json
Runtime: @stratiqx/cal-runtime · Spec: cal.semanticintent.dev · DOI: 10.5281/zenodo.18905193
After three shuffles each of eight cards is perfectly random, yet nine in ten orders of the deck are still impossible. Randomness of the parts is not randomness of the whole.[1][13]
The result settles how the number of shuffles grows, not how many a real deck needs. The constant is an explicit choice, and the bound is asymptotic.[1]
The main statement is machine-checked. A week later no specialist has commented. A formal check removes one doubt; it does not create an audience.[3]
For the brick factory, silence was a decision to wait. Here it looks like absence. Both are recorded, because neither is a verdict on the proof.[11]
Primary sources are OpenAI’s main paper and its companion, the machine-check note and change history; Thorp’s 1973 paper; Morris’s 2005 and 2009 papers; and the 2009 cryptography paper. All were read directly. The eight-card figures were computed for this case by exact enumeration of all 40,320 orders. No specialist comment on this result was found in a search of the public record to October 14, 2026. This case does not claim the proof is correct. Source 13 is an interactive companion visual made for this case.
This case separates randomness of the parts from randomness of the whole — each card can pass while the deck fails. The same gap appears in audits, tests and dashboards that check components one at a time. Book a discovery call to find where yours are measuring the cards and not the deck.