Notes on Played Pages Beat New Cards by 41%
A claim like the one in the title must be handled with care. "Played pages beat new cards by 41%" is not a universal law of blackjack or a finding from a single casino's floor data; it is a specific, reproducible result from a controlled simulation of a particular rule set. Specifically, in a six-deck shoe game with dealer standing on soft 17, double after split allowed, and no surrender, a strategy that selects its action based on the composition of the played cards (the visible discards and the current hand) outperformed a strategy that assumes a fresh, full shoe by 41% in terms of expected value per hand, measured over a sample of 10 million simulated rounds. This article unpacks what that stat means, why the gap is so large, and where the methodology strains against real-world constraints.
The Experimental Setup: What Was Actually Compared
The simulation was not testing card counting in the traditional sense—no running count, no true count conversion, no bet spreading. Instead, it compared two decision engines operating on the same fixed betting unit (flat betting, 1 unit per hand) and the same basic strategy thresholds. The "new cards" baseline used standard basic strategy derived from a full shoe's initial composition: the classic hit/stand/double/split indices for a six-deck game, assuming the shoe is always at its starting density of tens and aces. The "played pages" engine, by contrast, recalculated the optimal action for every decision point using the exact remaining deck composition after accounting for every card seen—both the player's own hand and all other hands at the table, plus the burn card.
The 41% figure emerged from comparing cumulative expected value (EV) per 100 hands. The baseline engine lost at a rate of 0.42 units per 100 hands (the house edge for that rule set under basic strategy). The composition-aware engine lost at a rate of 0.25 units per 100 hands. The difference—0.17 units—represents a 41% reduction in the house edge, not a 41% increase in win rate. This distinction is critical: the "played pages" engine did not flip the game into a player advantage; it merely chipped away at the casino's margin. The title's percentage, while technically accurate, is a reduction in losses, not a gain in winnings.
Why Composition Awareness Matters More Than You Think
The Non-Linearity of Deck Composition
The conventional wisdom among recreational players is that card counting matters only at extreme deck penetrations (e.g., the last 20% of the shoe). The simulation found that the benefit of composition awareness is not linear but convex. Early in the shoe, the played cards' information is negligible—the remaining deck is statistically near-identical to a full shoe. But by the third round of a six-deck game, the cumulative effect of removing a few low cards or a single ace shifts the optimal strategy on marginal hands (e.g., 12 vs. dealer 4, or 16 vs. dealer 10) with measurable frequency.
The 41% gap is not driven by dramatic plays like splitting tens or doubling on soft 21. It is driven by hundreds of micro-decisions per shoe where the correct play differs from basic strategy by one card's worth of probability. For instance, consider a hand of 12 against a dealer's 4. Basic strategy says hit. But if the played pages show that the remaining deck is rich in tens (because many small cards have been exposed), the expected value of standing on 12 rises above the EV of hitting. The simulation recorded 1,847 such decision points across the 10 million hands where the composition-aware engine chose a different action than the baseline, and in 1,623 of those cases (87.8%), the composition-aware choice was correct.
The Marginal Card Effect
Here is the numerical anchor that matters for practitioners: the simulation showed that the marginal information value of each additional played card decays according to a power law. The first 10 cards seen after the shuffle reduce the house edge by 0.09 units per 100 hands. The next 10 cards reduce it by only 0.04 units. The marginal value crosses zero at approximately 78 cards seen—after that point, knowing the exact composition of the remaining deck provides no additional EV benefit over knowing just the running count. This is a counterintuitive result because it suggests that a player who tracks every single card (as the "played pages" engine does) gains nothing over a player who merely tracks the ratio of tens to non-tens after roughly half a shoe has been dealt.
The Methodology's Structural Assumptions
Perfect Memory and Zero Error Rate
The simulation granted the "played pages" engine perfect recall of every card seen, including the dealer's hole card when it was revealed. In practice, human players misremember or fail to see cards from adjacent hands, and the error rate compounds. A sensitivity analysis within the same study showed that a 2% error rate in card recall (i.e., misremembering one card out of every 50 seen) erases 31% of the advantage, dropping the edge reduction from 41% to 28%. At a 5% error rate, the advantage is statistically indistinguishable from zero. This is the primary reason why the academic literature on card counting has long favored simple counting systems over full composition tracking: the human error term dominates the theoretical gain.
The Dealer's Penetration Point
The simulation ran with a fixed penetration point: the dealer cut card was placed at 75% of the shoe (i.e., 234 of 312 cards dealt before reshuffle). This is a generous penetration for the player; many casinos use 70% or even 65%. The 41% figure is highly sensitive to this parameter. At 65% penetration, the composition-aware engine's advantage over the baseline drops to 22%. At 85% penetration (rare, but found in some high-limit rooms), it rises to 58%. The relationship is roughly linear in the penetration range of 60–90%, meaning that the title's headline number is a midpoint estimate, not a worst-case or best-case scenario.
Comparison to Existing Literature
The result aligns with, but does not replicate, the work of Peter Griffin in The Theory of Blackjack (1979), who calculated that perfect composition knowledge yields a theoretical player edge of about 0.15% over basic strategy. Griffin's figure was derived analytically for a single-deck game. The 41% reduction in house edge found here corresponds to a 0.17-unit improvement per 100 hands—which, on a $100 average bet, amounts to $17 per 100 hands. That is a larger absolute improvement than Griffin's prediction, but the difference is explained by the rule set: the simulation allowed late surrender, which Griffin's model did not, and surrender decisions are among the most composition-sensitive plays in blackjack.
More recent work by Michael Shackleford (the "Wizard of Odds") on the effect of card removal on specific hands suggests that the simulation's findings are plausible but at the upper bound of what one would expect. Shackleford's published strategy variations for single-deck games show that composition-dependent plays (e.g., hitting 16 vs. 10 when the deck is rich in fives) occur roughly once per 40 hands. The simulation found a similar frequency—once per 43 hands—which lends external validity to the methodology.
The Open Question: Does This Matter for the Modern Player?
The practical implication of the 41% figure is not that players should abandon basic strategy for full composition tracking. The error-rate sensitivity analysis alone disqualifies that approach for all but the most disciplined savants. The real takeaway is that the marginal value of attention in blackjack has a sharply diminishing return curve. A player who spends the mental effort to track the ratio of aces to non-aces (a crude but low-error proxy for composition) captures perhaps 60–70% of the theoretical maximum benefit. A player who tracks every card captures the remaining 30–40% but at a tenfold increase in cognitive load.
This raises a question that the simulation cannot answer: if the house edge is reduced by 41% under perfect information, and by roughly 25% under a realistically achievable high-card tracking system, then why does the casino industry not see a wave of professional composition players emptying their high-limit rooms? The answer, presumably, lies in the gap between simulation and deployment—the 2% error rate, the heat from surveillance, and the fact that a 0.17-unit improvement per 100 hands, at $100 a hand, yields $17 per hour before tips and variance. That is a wage, not a fortune. The 41% is real, but it is a number that describes a laboratory condition, not a livelihood. The next useful study would measure the deployment cost: how much time and error does a human need to achieve even half of this theoretical edge, and at what point does the effort exceed the return? That number does not yet exist, and it may be the only one that matters.