Is the Perfect Shuffle a Myth? Part 7 - The Signal Escapes the Machine

11 July

Written By Paradis

Part 6 found the first crack.

The current six-deck One2Six-style model preserved the average physical-card return time while dramatically suppressing short returns. That established that the machine model had memory. A recently dealt physical card was not behaving like an independent random selection from all 312 cards.

That was an important result, but it left the question that actually matters for blackjack:

Does any of that hidden physical memory become visible to the player before the next bet?

The answer from the latest experiments is yes.

The composition of the visible discard rack predicted the composition of the next deal. Recently returned batches continued to carry a weaker predictive effect for roughly the next 100 cards. The signal survived new, held-out simulation seeds. It correctly predicted low-card, ten-value and ace frequencies separately.

Most importantly, it began to show up in money.

Under the unchanged player strategy, the strongest predicted high-card-rich state moved the player from an unconditional loss of roughly 0.46 percent to an estimated result almost exactly at break-even.

That is the biggest result in the project so far.

It is not yet a validated betting system. The confidence interval around that near-break-even result remains wide, and no positive-EV state has yet passed the formal validation gates. But the project has now crossed a crucial line:

The model contains player-observable information that predicts the next deal and economically orders the game.

The signal is no longer trapped inside the machine.

Before looking for an edge, the game had to survive contact with reality

Part 6 was still a source-level experiment. It tracked physical-card recurrence without asking the simulator to play blackjack.

The next job was to validate the complete game. I wanted to know whether the blackjack rules, fixed player strategy, monetary accounting and One2Six discard procedure could produce credible long-run results before trusting any conditional advantage analysis.

The current formal casino blackjack rules required several details to be modelled correctly:

six decks for the present experiments;
no dealer hole card during the initial deal;
blackjack paid at 3:2;
dealer standing on hard and soft 17;
doubling on 9, 10 or 11;
doubling after a split;
one card to each split ace;
no normal resplitting;
original-wager-only treatment when the dealer later makes blackjack after a split or double;
an initial burn card;
return of the previous discard rack after the next initial deal when a shuffling device is used.

The stand-on-soft-17 rule was important because an older public guide had described the game differently. The engine was corrected to match the current formal rules before the large validation runs began.

The player used one active box, a flat wager and the same source-blind fixed strategy under both card sources. The strategy could see only the player hand, dealer upcard and legal actions. It could not see the source type, physical card identities, carousel shelves, output buffer, telemetry or prior-card history.

Only the card source changed:

six-deck physical IID;
the six-deck One2Six-style model.

Across five independent seeds and one million rounds per source, the long-run results were strikingly close.

Metric	Physical IID	One2Six
Mean player edge	-0.5489%	-0.5675%
Player blackjack rate	4.7467%	4.7566%
Double rate	8.6753%	8.6711%
Initial pair rate	14.8050%	14.4892%
Split rate	2.6558%	2.5253%
Win rate	43.3474%	43.3403%
Loss rate	47.7809%	47.8612%
Push rate	8.8717%	8.7985%

The paired monetary difference between One2Six and physical IID was only minus 0.0186 percentage points, and its confidence interval comfortably crossed zero. The difference also changed direction across seeds.

The first win here was simple but important:

The One2Six model was not producing absurd blackjack results.

Blackjack frequency, doubling, wins, losses, pushes and overall monetary performance all landed in credible ranges. One million game rounds per source produced no physical-card invariant failure and no fallback ejection.

The pair-rate difference was not a bug

One result initially stood out. The One2Six model produced fewer starting pairs and therefore fewer split actions.

At first glance, that could have indicated a dealing or strategy error. Once the pair opportunities were instrumented directly, the opposite became clear.

For physical IID, the player’s two initial cards are independent selections with replacement. The theoretical equal-value pair rate is:

25 / 169, or approximately 14.7929 percent.

The observed physical-IID result was:

14.8050 percent.

For a finite six-deck physical population, drawing the first card temporarily reduces the number of matching cards still available. The corresponding finite-population pair rate is approximately:

14.5189 percent.

The observed One2Six result was:

14.4892 percent.

The difference between the two simulated sources closely matched the mathematical difference between with-replacement IID and a finite physical card population.

That was a major validation win. The shuffler was not merely producing superficially plausible totals. It was reproducing a subtle and predictable effect of physical card removal.

The lower split count was therefore not evidence that the engine was broken. It was evidence that the One2Six source was behaving like a real circulating set of physical cards rather than a timeless random-number generator.

The streak hypothesis did not survive

I had also suspected that the shuffler’s internal dependence might change the shape of player winning and losing streaks.

The first streak summaries showed almost identical means and percentiles, but matching averages do not prove independence. Two distributions can have the same mean while differing in continuation probability, tail weight or serial dependence.

So I ran a dedicated streak-shape audit over two million measured rounds.

The canonical streak definition was monetary:

a positive total box result was a win;
a negative total box result was a loss;
a zero total box result was a push;
split hands were combined before classification;
pushes neither counted nor broke a streak.

The audit compared the full win-run and loss-run distributions with the geometric distributions implied by the observed win and loss rates. It also examined survival curves, continuation probabilities, transition matrices, tail ratios and outcome autocorrelations.

The result was clear.

Metric	Physical IID	One2Six
Mean win streak	1.9014	1.9051
Geometric implied win mean	1.9042	1.9063
Mean loss streak	2.1028	2.1020
Geometric implied loss mean	2.1060	2.1033
Win-streak 95th percentile	5	5
Loss-streak 95th percentile	5	5

The full empirical distributions closely followed the geometric benchmarks. Continuation probabilities remained approximately constant across streak length, and outcome autocorrelations were essentially zero.

I then tested whether current streak sign and length improved held-out monetary prediction after accounting for the validated card-composition score.

They made it slightly worse.

The conclusion is now firm:

Monetary streaks are not the edge.

The One2Six card stream is physically non-IID, but that dependence is largely washed out when the entire game is compressed into wins, pushes and losses. Streaks remain useful as a reporting and validation metric, but they should not be part of the advantage strategy.

The signal became visible in the discard rack

The next experiment moved directly to the information available before a bet.

Under the game procedure being modelled, the previous round’s discard rack remains outside the shuffler until after the next initial deal. Those cards are therefore definitely unavailable during that initial deal.

That creates a simple physical hypothesis.

If the visible rack contains an unusually high number of low cards, those low physical cards are temporarily absent from the next deal. The available output may therefore become relatively richer in ten-value cards and aces.

The converse should also apply. A high-card-heavy visible rack should temporarily make the remaining output less favourable.

I recorded the visible rack’s:

size;
Hi-Lo count;
low-card count;
neutral-card count;
ten-value count;
ace count.

The simulator then cloned the exact pre-deal source state and inspected the next 15 cards without returning the rack. Fifteen cards cover the complete initial deal for seven boxes.

The physical-IID source served as the negative control. Under physical IID, the visible rack should have no predictive value because every future draw is independent.

The One2Six result was not subtle.

Predicted outcome	Physical IID slope	One2Six slope
Hi-Lo composition	0.637	1.132
Low-card frequency	0.481	0.987
Ten-value frequency	0.883	1.260
Ace frequency	-0.165	1.165

The physical-IID intervals all included zero. The One2Six intervals all excluded zero, and the direction was exactly what temporary physical removal predicted.

A low-heavy rack predicted:

fewer low cards in the next 15;
more ten-value cards;
more aces;
a more high-card-rich Hi-Lo composition.

This was the second major breakthrough:

The physical memory found in Part 6 was visible to the player before the next bet.

The machine had a measurable memory horizon

The current rack is the cleanest signal because its cards are known to be outside the machine. But a continuous shuffler contains overlapping cohorts. Earlier discard batches have already been returned and are moving through the feeder, carousel and output buffer.

The next experiment measured how long the composition of a returned batch continued to predict future card composition.

The result produced the first empirical response kernel for the model.

Cards dealt after batch return	Hi-Lo response	Low response	Ten-value response	Ace response
1 to 15	0.744	0.748	0.777	0.807
16 to 50	0.394	0.381	0.483	0.448
51 to 100	0.201	0.199	0.187	0.244
101 to 250	0.002	0.005	0.011	0.002
251 to 500	-0.001	0.006	0.009	0.005
501 to 1000	-0.024	-0.022	-0.017	0.011

The signal was strong immediately after return, weakened substantially through 50 cards, remained detectable through 100 cards, and was effectively gone after that at the current level of resolution.

That result clarified an important distinction.

Exact physical-card recurrence remained distorted for hundreds of cards, with a tail extending towards 1,000 cards. But player-observable rank and value information did not remain useful for that long.

The practical composition-memory horizon was much shorter:

strong through 15 cards;
moderate through 50;
weaker but detectable through 100;
negligible after 100.

There was also no clear delayed reversal wave. I had considered the possibility that a high-heavy group might later return in a predictable burst. The data did not support that route.

The usable mechanism appears to be temporary exclusion, not waiting for high cards to become due.

Building a fading count

The response kernel suggested a practical observable state:

Cohort	Frozen weight
Current visible rack	1.00
Returned 1 to 15 cards ago	0.75
Returned 16 to 50 cards ago	0.40
Returned 51 to 100 cards ago	0.20
Older than 100 cards	0.00

These were deliberately rounded weights. The purpose was not to claim that the perfect coefficients had been discovered to three decimal places. The purpose was to freeze a simple candidate signal and test it on new simulation seeds.

The score combined the current rack with every active returned cohort. It was calculated entirely from observable card composition and timing. It contained no physical identifiers, shelf contents, feeder state, output-buffer state or machine telemetry.

Development used seeds 42 through 46.

Validation used untouched seeds 47 through 51.

The score survived.

Outcome	One2Six held-out slope	Seed-level interval
Hi-Lo	0.733	0.416 to 1.050
Low cards	0.865	0.715 to 1.015
Neutral cards	0.868	0.620 to 1.117
Ten-value cards	0.695	0.396 to 0.994
Aces	0.791	0.484 to 1.099

Every One2Six slope was positive in all five held-out seeds. Every direct One2Six-minus-IID paired difference was also positive across all five seeds.

The actual one-box initial deal was even more strongly aligned with the score:

Outcome	One2Six initial-deal slope
Hi-Lo	1.023
Low cards	1.087
Ten-value cards	1.093
Aces	1.138

This was not an in-sample pattern found by searching until something worked. The weights were frozen before the new seeds were examined.

The result generalised.

A player-observable fading count predicted the composition of the actual next deal.

The signal finally reached money

Predicting card composition is necessary, but the project is about advantage play. The next experiment made round profitability the primary endpoint.

The score bands were created from One2Six development scores only. Monetary outcomes were not used to choose the cutpoints.

New validation seeds were then played continuously using the unchanged fixed player strategy.

The One2Six results were:

Score state	Frequency	Player edge
Strong high-rich	9.853%	-0.002%
Moderate high-rich	19.960%	-0.356%
Neutral	40.306%	-0.459%
Moderate low-rich	20.215%	-0.721%
Strong low-rich	9.666%	-0.673%

The unconditional One2Six edge was approximately minus 0.462 percent.

The strongest high-rich state had a point estimate of minus 0.002 percent.

In other words:

The observable score recovered almost the entire house edge in its strongest decile.

That is not a rhetorical flourish. It is what the point estimates showed.

The result also came with the expected physical changes. Moving from the strongest high-rich state to the strongest low-rich state:

player low-card frequency rose from 37.17 percent to 39.90 percent;
player ten-value frequency fell from 31.73 percent to 29.53 percent;
player ace frequency fell from 8.01 percent to 7.54 percent;
player blackjack frequency fell from 5.111 percent to 4.423 percent.

The score was not merely ranking arbitrary simulation states. It was ranking exactly the card categories that should matter for blackjack, and the monetary results moved in the expected direction.

The best state did not pass the formal positive-EV gate. Its confidence interval extended well above and below zero, and only four of ten seed-level estimates were positive.

That means I cannot yet say the game is beaten.

But I can say this:

The model now contains an observable state that appears to take an ordinary negative game to the edge of break-even before changing the player strategy or wager.

That is a major result.

The negative controls mattered

Physical IID did not produce the same ordered composition response. Its monetary slope was consistent with zero, and the permutation-placebo tests were also consistent with zero.

One direct comparison was especially notable. The difference between the moderate-high-rich and neutral states was 0.476 percentage points more favourable under One2Six than under physical IID. Its seed-level interval ran from 0.044 to 0.907 percentage points.

The One2Six contrast by itself remained uncertain, but the source difference excluded zero.

That provides evidence that the economic ordering was tied to the One2Six state mechanism rather than merely being an artefact of dividing random blackjack results into groups.

The Physical IID controls were not perfectly silent in every secondary test. With enough positions, score cells and action comparisons, isolated false positives appeared. That was expected and became an important warning against selecting the largest number from a large exploratory table.

The research process increasingly relied on:

pre-specified primary tests;
independent seeds as the uncertainty unit;
frozen features;
held-out validation;
direct One2Six-minus-IID comparisons;
permutation placebos;
explicit support requirements.

That discipline prevented several attractive-looking but unsupported results from being mistaken for an edge.

The broad strategy search found no validated deviations

Once the high-rich state approached break-even, the obvious question was whether a composition-adjusted playing strategy could recover the final fraction of a percent.

A counterfactual action experiment collected 200,000 natural player decision states and branched every legal action from the same hidden source and game state. In total, it evaluated more than 407,000 action branches across development and validation seeds.

The composition state was updated at the actual decision point. That mattered because the previous discard rack had already been returned after the initial deal, while the player cards and dealer upcard had become newly exposed and fully unavailable.

The experiment tracked low-card, ten-value and ace richness separately. It searched both favourable states, where a deviation might create an edge, and unfavourable states, where a better decision might reduce the loss.

No deviation passed the full discovery and held-out validation gates.

That is a real negative result. The published fixed strategy appeared robust across most of the distributional changes the model produced.

One exploratory lead remained:

hard 16 against a dealer ten;
low-poor state;
stand rather than hit;
estimated improvement of 0.1241 initial wagers;
interval from 0.0666 to 0.1836;
only 137 matching states.

It failed because support was too low, not because the observed effect was small. It is a reasonable pre-specified hypothesis for a fresh targeted test, but it is not a validated strategy change.

The larger lesson is useful. If strategy deviations exist, they are likely to be a small set of close threshold decisions rather than an entirely new blackjack chart.

What has now been established

The project has crossed several gates since Part 6.

Confirmed inside the current model

The One2Six-style physical-card process is not IID.
Short physical-card returns are strongly suppressed.
The effect survives recycle-batch sensitivity testing.
The blackjack engine and One2Six integration produce credible long-run outcomes.
The finite physical population produces the expected reduction in equal-value pairs.
Observable discard composition predicts future card composition.
The predictive effect survives held-out seeds.
Returned-batch information fades over roughly 100 dealt cards.
Ten-value cards and aces can be predicted separately.
The strongest predicted high-rich state nearly erased the house edge by point estimate.
Monetary win and loss streaks remain close to geometric and add no predictive value.

Not yet established

No state has a validated positive player edge.
No playing deviation has passed held-out support and uncertainty gates.
No betting spread has been tested.
No multiple-box advantage policy has been validated.
No live or physical machine has been tested against the model.
The extreme tail of the score has not yet completed its full profitability run.

This is now an advantage-play project in substance, not merely a shuffler-diagnostics project. The information exists. The remaining question is whether it can be converted into a robust positive expectation.

What appears to be new

There is prior public work on mechanical shufflers, continuous-shuffler state and recent-card counting.

Diaconis, Fulman and Holmes showed that casino shufflers can be analysed as mechanical probability processes rather than treated as abstract sources of randomness. Stephen How examined a CSM model in which recently inserted cards and buffer depth could create a countable next-hand effect.

Those ideas matter, and this project does not exist in a vacuum.

However, I still have not found a public reproducible analysis that follows this complete chain:

a 312-labelled-card physical IID null;
a configurable six-deck One2Six-style carousel model;
same-physical-card recurrence measurement;
recycle-batch sensitivity;
game-level rule and payout validation;
observable current-rack response;
returned-batch memory-horizon estimation;
a frozen fading score;
held-out composition validation;
conditional blackjack profitability;
counterfactual player-action testing;
full streak-shape and incremental-prediction audits.

The broad ideas have precedent. The complete implementation and evidence chain appear distinct in the public material I have found.

The major finding is no longer merely that a machine model has memory.

It is this:

Under the current model, ordinary observable discard history predicts the next deal strongly enough to move the estimated game from a normal house edge to approximately break-even in the most favourable score band.

That is worth calling out.

Where the project goes next

The next experiment is deliberately narrow.

The strongest decile was close to break-even, but averaging the best 10 percent of states may conceal a genuinely positive extreme tail. The best 5 percent, 2.5 percent or 1 percent may have materially better expectation.

An extreme-tail experiment has now been implemented. It defines score cutpoints without looking at monetary outcomes, then applies them to new validation seeds under both One2Six and physical IID.

It will test:

the best 20 percent;
the best 10 percent;
the best 5 percent;
the best 2.5 percent;
the best 1 percent;
matching low-rich tails;
a neutral centre.

The primary question is simple:

Does any pre-specified high-rich tail have player EV above zero with a seed-level confidence interval entirely above zero?

If the answer is yes, the next strategy is likely to be selective entry rather than complicated play modification. An observer could minimise exposure outside favourable states and enter only when the validated score crosses the threshold.

Only after a positive state exists does it make sense to increase wagers, add boxes or optimise how much of the favourable card window the player captures.

Insurance and even money will also be audited as secondary decisions, but they are not the centre of the project. In a no-hole-card game, the dealer’s second card is drawn only after player actions, so pricing insurance requires the full decision-time process rather than treating the insured card as the immediate next card.

The extreme-tail code is complete, but the full run exceeded the automated 15-minute execution guard. It now needs to be run locally without reducing the requested sample size.

That result will determine the next branch.

If an extreme tail is positive, the project moves directly to selective-entry validation.

If no tail is positive, the next step is to improve the score using rank-specific blackjack effect-of-removal values, then test it on entirely new seeds.

The codebase

The repository remains licensed under GPL-3.0-or-later.

The current verification state is:

391 passing tests;
Ruff lint passing;
Ruff formatting passing;
mypy passing;
compile checks passing;
generated experiment outputs excluded from Git;
privacy searches clean.

The experiment framework now includes:

physical IID recurrence;
One2Six recurrence;
recycle-batch sensitivity;
one-box long-run validation;
pair and split decomposition;
monetary streak analysis;
multi-box counterfactual branching;
observable card-response analysis;
memory-horizon estimation;
frozen-signal validation;
conditional profitability;
score-conditioned action values;
geometric streak-dependence auditing;
extreme-tail profitability infrastructure.