It Started with Binary
One night, I was staring at Powerball draws converted to binary. Five numbers, seven bits each, lined up in a row like punched tape from a 1960s mainframe.
2026-03-25 | ····███ ··█·█·█ ·██·███ ·███··· █······
2026-03-23 | ···██·· ··█··█· ·█·████ ·███··· ·██████
2026-03-21 | ···██·· ··███·· ·█··█·· ·█·█··█ ·███·██
It was mesmerizing. Each draw had its own texture — some heavy with filled blocks, others sparse and scattered. I started calculating column percentages: how often is each bit position a 1 across the last 50 draws?
The answer was... exactly what you'd expect from random chance. Around 50% for the middle bits, skewed for the high bits (because the numbers are sorted, so position 1 is always small and position 5 is always large).
So I tried something different.
Testing Every Mathematical Lens
Over the next few days, I threw everything I could think of at the data:
Perfect squares, Fibonacci, triangular numbers, powers of 2 — mathematically beautiful subsets of 1-69. If the lottery somehow "preferred" numbers with special properties, it would show up as a deviation from the hypergeometric distribution.
Every single category tracked the expected values within noise. The ratios ranged from 0.77 to 1.41 — textbook sampling variance. The machine doesn't care that 36 is a perfect square.
Prime factorization — I decomposed every drawn number into its prime factors. Do draws favor highly composite numbers? Is there a "smoothness" pattern? The answer: factor frequencies matched the uniform distribution to three decimal places. Factor 2 appeared 5.36 times per draw vs 4.78 expected. Factor 7 was at 1.02 vs 0.72. All noise.
Modular residues — Every number has a fingerprint: its remainder when divided by 2, 3, 5, and 7. By the Chinese Remainder Theorem, since 2 x 3 x 5 x 7 = 210 > 69, every number from 1-69 has a unique fingerprint. Across 1,919 draws, not a single pair of draws shared the same residue pattern. And the residue distributions? Dead center — 33.55% / 33.11% / 33.34% for mod 3, against an expected 33.33% each.
Digit sums — Sum the individual digits of all five numbers (e.g., {47, 58, 3, 12, 69} becomes 4+7+5+8+3+1+2+6+9 = 45). The distribution forms a near-perfect bell curve: 68.6% of draws fall within one standard deviation of the mean, against the normal prediction of 68.3%. The central limit theorem called it.
Gap encoding — Instead of the numbers themselves, look at the spaces between them. Five sorted numbers create six gaps (including the edges), and they always sum to 69. The gap histogram is a clean geometric decay — gap=1 and gap=2 at 7.8% each, dropping smoothly to gap=30 at 0.6%. No hidden modes, no bumps.
Polar coordinates — Map 1-69 around a circle like a clock face. Each draw becomes a pentagon-like shape. I measured centroid displacement, angular symmetry, and polygon area. The mean centroid distance was 0.398 vs a Monte Carlo simulation's 0.388. Symmetry score: 0.642 actual vs 0.629 simulated.
I even turned the draws into sound — mapping each number to a musical note and playing the five-number chord. Most draws sounded dissonant. A few were accidentally beautiful. None revealed a hidden pattern.
The Realization
After all of this, the conclusion was undeniable: the lottery is genuinely, thoroughly random.
Every mathematical lens — binary, BCD, prime factors, modular arithmetic, digit sums, gap encoding, polar geometry — told the same story. The actual distributions match the theoretical predictions to a degree that's almost boring.
The Detour: Drift Rebalancing
Before arriving at our current model, we tried something different.
We built a drift detector that compared the last 20 draws against the full history across eight structural metrics. When a metric deviated more than one standard deviation from the baseline, we'd award bonus points to picks that leaned the opposite direction. The logic: the law of large numbers guarantees distributions converge, so lean into where they're heading.
It was clever. The z-scores were real. The math was sound. And the distinction from the gambler's fallacy was defensible — we weren't saying "number 7 is due," we were saying "the structural distribution is even-heavy."
But there was a problem: it was empirical.
The drift detector relied on historical data to make assumptions about future draws. Even though each draw is independent, the system was implicitly betting on mean reversion within a short window. And in a truly random system, there's no mechanism forcing the next draw to correct a drift. The distribution will converge — but over hundreds of draws, not twenty.
We were making a sophisticated version of the same mistake we'd warned users about.
The Combinatorial Framework
So we stripped it all out and asked a simpler question: what can we know about Powerball combinations without looking at a single historical draw?
The answer turns out to be a lot.
Powerball is 5 numbers drawn from 1-69. That's C(69,5) = 11,238,513 possible white ball combinations. Every combination is equally likely to be drawn. But not every combination is equally structured.
Some properties are far more common than others in the combinatorial space:
- 64.4% of all combinations have a 2/3 or 3/2 odd/even split
- 64.3% cover all three remainder classes (mod 3)
- 53.0% have evenly spaced gaps between numbers
- 35.1% have 5 unique last digits
These aren't patterns from past draws. They're mathematical facts about the structure of C(69,5). You can compute them from first principles without ever seeing a lottery result.
Weighting by Filter Strength
The next question: if every criterion is purely combinatorial, how do we weight them?
Our answer: proportional to how much of the space each criterion filters out. A criterion with a 35% pass rate (Unique Digits) eliminates 65% of combinations — it's doing nearly twice the filtering work of a criterion at 69% (Tens Diversity, eliminating 31%). The weight reflects the discrimination power.
This gives us 10 criteria, 100 points, weighted by filter strength:
| Criterion | Pass Rate | Points |
|---|---|---|
| Unique Digits | 35.1% | 17 |
| Even Spacing | 53.0% | 12 |
| Parity | 64.4% | 9 |
| High/Low | 64.4% | 9 |
| Spread | 65.3% | 9 |
| Modular Balance | 64.3% | 9 |
| Range Coverage | 64.2% | 9 |
| Sum Range | 67.5% | 9 |
| Primes | 68.8% | 9 |
| Tens Diversity | 69.3% | 8 |
Every single criterion can be derived from pure combinatorial math. No historical data. No drift assumptions. No empirical bets.
What This Means
We generate 50,000 random combinations and score each one against these 10 criteria. The highest-scoring picks sit in the densest, most structurally probable region of the combinatorial space. They're the combinations that look the most like what math predicts a "typical" draw would look like.
Does this increase your odds of winning? No. Every ticket is still 1 in 292 million. The machine doesn't know or care about structure.
What it does is ensure your picks aren't structurally weird. No all-even sets. No five numbers crammed into the same decade. No extreme sums. You get combinations from the fattest part of the probability distribution — the region where the most possible draws live.
What This Isn't
Let's be clear:
- It doesn't increase your odds of winning. Every ticket is still 1 in 292 million.
- It doesn't predict which numbers will be drawn. It doesn't even try.
- It doesn't assume past results influence future draws. Each draw is independent.
- It doesn't use any historical data in the scoring criteria. Zero.
The only thing it does is select from the most probable region of the combinatorial space. That's a defensible, transparent, purely mathematical claim.
Why We Changed
We could have kept drift rebalancing. It was sophisticated, it looked impressive, and most users would never question it. But we kept coming back to one thing: if the lottery is truly random — and our own research proves it is — then any system that relies on historical patterns is making a bet it can't justify.
The combinatorial model makes no bets. It says: here are the structural properties of 11.2 million possible combinations, here's how your picks compare, and here's the math to prove it.
Every scoring criterion, every weight, and every pass rate is published on our audit page. The model recalibrates itself weekly — not by chasing drift, but by backtesting criteria pass rates against real draws to confirm alignment. If you disagree with our approach, you have enough information to make that judgment.
The lottery is random. The combinatorics are not.