The Itch After a Cleanup
We just shipped scoring model v7.0. The headline changes: we recalibrated everything to the current 5/69 era, rebuilt every weight from pure combinatorial math, and dropped a criterion — Range Coverage — because it was too correlated with the ones around it. (The full story is on the audit page.)
Cutting a criterion leaves an obvious itch. If one of our nine was redundant, is the reverse also true — is there some structural pattern in a Powerball combination that we're not scoring, that we should be? A genuine tenth axis we've been blind to?
So we went looking. Thirty candidate patterns, three million simulated combinations, one correlation matrix. This is what came back — including the part where the search talked us out of adding anything at all.
First, What "Discovery" Can and Cannot Mean
This is the trap, and it's worth being explicit about before a single number appears.
Discovery cannot mean finding a pattern that predicts winners. We've proven, repeatedly and publicly, that the current-era Powerball draw is indistinguishable from uniform random — every one of our criteria sits inside its expected combinatorial band. There is no signal to find. Any "pattern" that appeared to predict outcomes would be noise wearing a costume, and we'd be lying to you if we shipped it.
So what's left? Something narrower but real:
A new criterion is worth having only if it measures a structural property of a combination that is (1) discriminating — it filters out a meaningful chunk of the possibility space — and (2) independent — it isn't just a repackaging of a criterion we already score.
That's it. A new criterion doesn't make your numbers more likely to hit. It makes the "this combination is structurally typical" judgment a little richer, along an axis the current model ignores. The bar is: does this tell us something genuinely new about the shape of a combination?
The Method
Every criterion's behavior is a fixed mathematical fact about the 11,238,513 possible white-ball combinations, so we measure it directly:
- Simulate 3,000,000 random 5/69 combinations with a seeded generator (so the run is reproducible).
- Score each one against the current 9 criteria — using the exact code from our live scoring engine, so the comparison is honest, not a re-implementation that might drift.
- Score each one against 30 candidate patterns (listed in full below).
- For every candidate, compute two numbers: its pass rate (filter strength) and the φ-correlation (phi coefficient) between it and each of the 9 existing criteria. The largest absolute correlation tells us how much the candidate overlaps with something we already do.
The bar for "keep": a pass rate in a useful range and |φ| < 0.15 against every existing criterion. Below 0.15, you've found a new axis. Above ~0.30, you've reinvented one you already have.
The Sanity Check: The Method Catches Itself
Before trusting any "new" result, the sweep has to correctly identify the old ones. It does:
- Range Coverage (the criterion we just dropped) comes back at φ +0.41 with Sum Range — exactly the redundancy that got it cut from v7.0.
- Average Gap correlates +1.00 with Spread. Perfect. Of course it does: average gap is literally the spread divided by four. If the method hadn't flagged that, the whole exercise would be worthless.
- A broadened Primes test (1–3 primes instead of 1–2) lands at +0.76 with our existing Primes criterion.
Good. The ruler measures itself correctly. Now we can trust what it says about the rest.
All 30 Candidates
Sorted from most-independent to most-redundant. "Max │φ│" is the strongest correlation to any of the nine existing criteria; the criterion it's closest to is named.
| # | Candidate pattern | Pass rate | Max │φ│ vs current 9 | Closest existing | Verdict |
|---|---|---|---|---|---|
| 1 | Sum Even (parity of the total) | 50.0% | 0.02 | Parity | ★ new axis |
| 2 | Mod-7 All Distinct (5 distinct residues mod 7) | 17.3% | 0.05 | Unique Digits | ★ new axis (strict) |
| 3 | Multiples of 5 ≤ 1 | 76.4% | 0.05 | Unique Digits | ★ new axis (mild) |
| 4 | Perfect Square Present | 47.0% | 0.06 | Even Spacing | ★ new axis |
| 5 | Triangular Number Present | 59.2% | 0.07 | Even Spacing | ★ new axis |
| 6 | All Gaps Distinct (4 distinct gaps) | 79.1% | 0.12 | Tens Diversity | ★ new axis (mild) |
| 7 | Fibonacci Number Present | 51.4% | 0.15 | Even Spacing | borderline |
| 8 | Mod-4 Diversity ≥ 3 | 84.4% | 0.17 | Parity | some overlap |
| 9 | Last-Digit Sum within 1σ | 65.9% | 0.18 | Unique Digits | some overlap |
| 10 | Exactly One Consecutive Pair | 24.2% | 0.24 | Tens Diversity | some overlap |
| 11 | Digit Sum within 1σ | 68.3% | 0.27 | Sum Range | some overlap |
| 12 | Gap Variance within 1σ | 82.9% | 0.28 | Even Spacing | some overlap |
| 13 | Single-Digit Number Present (≤9) | 51.4% | 0.30 | Even Spacing | overlap |
| 14 | No Consecutive Pair | 73.5% | 0.30 | Tens Diversity | redundant |
| 15 | Tens Groups = 5 (all different) | 17.3% | 0.30 | Tens Diversity | redundant |
| 16 | High Tail Present (≥60) | 55.5% | 0.31 | Even Spacing | redundant |
| 17 | ≥4 Distinct Last Digits | 85.0% | 0.31 | Unique Digits | redundant |
| 18 | First Number ≤ 15 | 71.9% | 0.32 | Even Spacing | redundant |
| 19 | Last Number ≥ 55 | 71.8% | 0.32 | Even Spacing | redundant |
| 20 | Quintile Spread ≥ 4 | 45.7% | 0.40 | Tens Diversity | redundant |
| 21 | Range Coverage (all 3 thirds) | 64.3% | 0.41 | Sum Range | redundant (the dropped one) |
| 22 | Mod-5 Diversity ≥ 4 | 45.7% | 0.45 | Unique Digits | redundant |
| 23 | Balanced Thirds (≤2 per third) | 39.3% | 0.46 | Sum Range | redundant |
| 24 | Mean Absolute Deviation within 1σ | 67.1% | 0.50 | Spread | redundant |
| 25 | No 3-in-a-row Same Parity | 52.8% | 0.53 | Parity | redundant |
| 26 | Sum within a tight 0.5σ | 37.1% | 0.53 | Sum Range | redundant |
| 27 | Median Centered within 1σ | 63.8% | 0.54 | Sum Range | redundant |
| 28 | No Tens Triple (≤2 per tens group) | 87.0% | 0.58 | Tens Diversity | redundant |
| 29 | Primes 1–3 (broadened) | 79.3% | 0.76 | Primes | redundant |
| 30 | Average Gap within 1σ | 65.3% | 1.00 | Spread | redundant (= spread ÷ 4) |
What Actually Survived
Six candidates cleared the bar. They split cleanly into two groups, and that split is the whole story.
The principled ones. A handful measure real structure our model genuinely ignores:
- Sum Even — whether the total of the five balls is even or odd. It's a 50/50 coin flip, and it correlates with essentially nothing (φ 0.02). Counterintuitively, this is independent of our Parity criterion: Parity counts how many individual balls are odd; this is about the parity of their sum. Two different things.
- Mod-7 All Distinct — all five numbers landing in different residue classes mod 7. Only 17% of combinations pass, which makes it a strict filter (in the same spirit as Unique Digits at 35%), and it's almost perfectly independent of everything we score. Of all thirty, this is the most genuinely surprising: a hard, completely orthogonal constraint hiding in plain sight.
- All Gaps Distinct — the four gaps between sorted numbers are all different values. Independent, though a mild filter at 79%.
The numerology ones. The other survivors — Perfect Square Present, Triangular Number Present, and (just outside the bar) Fibonacci Present — are mathematically independent and statistically well-behaved. And we are not going to use them.
Here's why, and it's a line we won't cross: there is no structural reason a "typical" Powerball draw should contain a triangular number. These patterns are independent precisely because they're arbitrary — they slice the number line along sequences that have nothing to do with how a combination is shaped. Scoring a combination higher because it happens to contain 36 (a square and a triangular number, as it happens) would be numerology dressed up as math. It would test well in this exact analysis and still make the model worse, because it would reward picks for a reason we can't honestly defend. The whole point of v7.0 was that every criterion be derivable from meaningful combinatorial reasoning. A Fibonacci bonus fails that test on day one.
That's the uncomfortable, useful lesson of this sweep: independence is necessary but not sufficient. A pattern can be a genuinely new axis and still be a bad criterion.
The Result: A Tenth Criterion That Wasn't
We didn't add anything. And that turned out to be the most reassuring possible outcome.
Look back at the table. Every "obvious" extension — anchors, tails, medians, tighter sum bands, broader primes, no-consecutive rules — is redundant. It just re-measures Sum Range or Spread or Tens Diversity from a slightly different angle. There was no discriminating structure hiding between our nine criteria, waiting to be scooped up. The handful of truly new axes are either marginal filters or arbitrary number-set trivia we won't touch on principle.
In other words: the model is well-spanned. The nine criteria cover the independent structural dimensions of a combination, and the cleanup that produced v7.0 left it lean rather than thin. If we ever do add a tenth, the only two candidates worth the debate are Mod-7 All Distinct (strict and independent) and Sum Even (clean and independent) — and even those would be additions for completeness, not because the model is missing something it needs.
As always, the honest footnote: none of this changes anyone's odds. Every combination is equally likely, every draw. What a sweep like this protects is something else — that when we tell you a pick is structurally sound, the criteria behind that word are real, independent, and defensible, and that we'll reject a good-looking pattern the moment it can't earn its place. We tested thirty. We kept zero. That's the system working.
You can see the nine that survive — every weight, every pass rate, recalibrated weekly — on the scoring audit page.