The Prime Knows Where It Is Because It Knows Where It Isn't

The shape of a guess

Photo by Cooper Hofmann on Unsplash

I'm a designer who codes. No math degree. No college degree, actually. I think in shapes — component APIs, state graphs, the geometry of how systems fit together. When I encounter a hard problem, my first move isn't to look up the literature. It's to rotate the thing until I can see a shape I recognize.

I want to be upfront about this: I didn't know what the Cramér model was when I started. I didn't know what a singular series was, or an Euler product, or iterative proportional fitting. I learned each of these things when I needed them, because the problem pushed me there. The names came after the shapes.

I honestly think that helped. If I'd started by reading the standard approaches to prime gaps, I'd have followed the standard path. I would have known what "should" work and what "can't" work, and I would have skipped the geometric intuitions that turned out to be the whole game. Not knowing the rules meant I couldn't be constrained by them.

There's an old military meme about missile guidance: "The missile knows where it is because it knows where it isn't. By subtracting where it is from where it isn't, or where it isn't from where it is, it obtains a difference, or deviation."

I kept laughing about this during the investigation, because it turns out that's literally how prime numbers work. A prime is a number that isn't divisible by anything that came before it. You find it by mapping everything it isn't. The prime knows where it is because it knows where it isn't.

So when I started poking at prime numbers, I didn't start with number theory. I started with a feeling: the primes, viewed the right way, should have a frequency. Not the primes themselves — the gaps. The stuff between them. The negative space.

Here's the intuition that started everything:

Normalize the primes as a straight line, and instead take the shape of what is not a prime. What sits in between has a particular smell, and it feels like a frequency.

I know how that sounds. But I also know that when I get that feeling — that there's a shape hiding in the projection — I'm usually not wrong about the shape existing. I'm just wrong about which layer it's on.

The specific guess was geometric. Picture a helicoid — a spiral surface, like a parking garage ramp. If you trace the edge, you're going around and around, but also climbing. Each revolution resets your angular position to zero. That's a sawtooth wave: climb, reset, climb, reset.

If you map the primes onto that surface and project down, the angular coordinate wraps modularly. Each revolution is a modular cycle. And the gap between where one prime lands and where the next one lands — that gap is a step along the edge.

The helicoid gave me the spatial reasoning to see the problem as modular. Tracing the edge, watching it reset, was what led me to fizzbuzz — to thinking about overlapping modular rhythms, each prime's multiples sweeping around at their own rate. The sawtooth wasn't the answer, but it was the shape that made me chase the modulo. And chasing the modulo turned out to be the entire game.

I was right that wrapping onto a periodic surface would reveal something. I was wrong about which layer it would be.

Layer 1: The chirp that wasn't prime-specific

I built the helicoid. I computed the wrapped-phase spectrum. And there it was: a beautiful, structured spectral signature. Peaks at specific frequencies. A clear signal rising out of what should have been noise.

For about thirty minutes, I thought I'd found something.

Then I tested it against a null model — what would you see if the primes were replaced by a purely random sequence with the same density? The answer: the exact same spectrum. Correlation 0.99+ with the analytic prediction from just the prime number theorem's logarithmic density.

The spectral signature was the density of primes projected through my transform. Not the primes themselves. The prime number theorem says primes thin out like 1/ln(x), and that slow thinning, when you wrap it onto a spiral, creates a deterministic chirp. I wasn't seeing prime structure. I was seeing the measurement apparatus.

If not this, then what?

The helicoid caught the density. So: subtract the density. Whatever remains is the prime-specific content.

Layer 2: The dechirped residual — something is actually there

After removing the chirp, I looked at the residual: the normalized gap rn=gn/ln⁡(pn)−1rn​=gn​/ln(pn​)−1. If primes were perfectly random (the Cramér model), this should be white noise.

It wasn't.

The consecutive gaps showed a lag-1 autocorrelation of about -0.057. Small, but real. After a small gap, the next gap tends to be larger. After a large gap, smaller. The primes compensate.

I threw three different surrogate tests at it:

• Shuffle the gaps randomly → correlation vanishes. The ordering matters.

• Shuffle within blocks → still vanishes. It's not a trend.

• IAAFT surrogates (preserve the power spectrum, destroy phase) → correlation is fully reproduced.

That third test was the key. IAAFT surrogates preserve all second-order statistics. The fact that they reproduced the correlation exactly meant there was no higher-order structure hiding underneath. The autocorrelation function was the complete description of the prime-specific content in the gap sequence.

So the question sharpened: what produces this specific anticorrelation?

The dead ends: what didn't work and why it mattered

Before I get to what did work, I want to talk about what didn't. Because the dead ends shaped the path more than the successes did.

The hazard model. My first attempt at modeling the gap transition law was a survival-analysis approach: given that the current gap is gg, what's the "hazard" of the next gap being hh? I built it from the admissibility ratio S3/S2S3​/S2​ as a conditional hazard multiplier. It produced the wrong sign of correlation — positive instead of negative. The model said "after a small gap, expect another small gap." The data said the opposite. I had to throw it out completely.

This was important. It told me the anticorrelation isn't a direct consequence of the sieve structure in the way I initially imagined. The sieve creates admissibility patterns, but the sign of the correlation depends on how you normalize and condition. A plausible-looking model can get the sign wrong. That killed my confidence in purely top-down reasoning and pushed me toward empirical fitting with honest holdout evaluation.

The constrained correlation optimizer. Later, I tried optimizing the transition matrix to directly target a correlation constraint — "find the joint distribution that achieves lag-1 = -0.047 while minimizing KL divergence." The optimizer was wildly unstable. Most targets either collapsed to the IPF baseline (the optimizer gave up) or blew up the row KL to 0.8+ (the optimizer destroyed the conditional distributions to hit the correlation target). I abandoned direct correlation-constrained optimization entirely and switched to NLL-only fits with correlation as an evaluation metric.

This was the right call. It taught me that correlation is a consequence of getting the conditional law right, not a target you can chase directly. Optimizing for the shadow instead of the object.

Gap repulsion. There's a principle in complex systems where nearby states push apart — similar values become rare, like the system is minimizing surprise by spreading things out. I tested whether consecutive prime gaps do this. They don't. Gaps of 6 followed by gaps of 6 happen all the time. There's no repulsion. The primes aren't trying to spread their gaps out. They just land where the sieve lets them.

The row-wise penalized fit. I tried fitting each row of the transition matrix independently with heavy regularization — 1,600 parameters, one per cell, shrunk toward the IPF baseline. This was supposed to give an upper bound on how much correlation the model class could achieve. Instead, the excessive shrinkage produced a false upper bound that was lower than what my 10-feature projected fit achieved. I mistook a fitting pathology for a structural ceiling. Once I realized this, I stopped using it as a reference point.

Each of these failures narrowed the search space. The hazard model said "the sign isn't trivial." The constrained optimizer said "don't chase the shadow." The GUE test said "no eigenvalue structure." The row-wise fit said "don't trust ceilings from overparameterized models." By the time I found what worked, I knew a lot about what the answer couldn't be.

Layer 3: Fizzbuzz on an infinite orchestra

Here's where the intuition about frequencies turned out to be right, just aimed at the wrong layer.

Think of each small prime as a metronome. The "3" metronome blocks every third number from being prime. The "5" metronome blocks every fifth. The "7" every seventh. These metronomes tick at different rates, and where their ticks overlap, the density of composites is higher — which pushes the primes into specific gap patterns.

This is the sieve of Eratosthenes, but I wasn't thinking of it that way. I was thinking of it as an interference pattern — overlapping rhythms creating dead zones.

When I went looking for the math, I found a formula that did exactly this — the "singular series," apparently from the 1920s. It computes, for each pair of consecutive gaps (g,h)(g,h), how the overlapping modular exclusion schedules of all small primes make that pair more or less likely. It was the fizzbuzz pattern, written as a product over primes. I couldn't tell you who came up with it if you asked me at a dinner party. I'm the kind of person who once walked a table of Seahawks players to their seats without realizing they were famous. Names don't stick. The formula stuck.

I built a transition matrix from S3S3​, calibrated it to match the observed gap frequencies using a technique called iterative proportional fitting (IPF — it adjusts a table to match known row and column totals), and asked: how much of the -0.057 anticorrelation does this explain?

About 30%.

Real, but not enough. The sieve tells you which pairs are admissible (not blocked by any small prime) but it doesn't tell you the right probabilities among the admissible pairs. Something else was modulating the static sieve pattern.

If not this, then what?

Layer 4: The row-local tilt — a correction that resisted explanation

On top of the sieve prior, I fit a small correction: for each conditioning gap gg, how does the distribution over the next gap hh tilt relative to what the sieve predicts?

The answer was a smooth, row-dependent mean reversion. After small gaps, the conditional distribution shifts toward larger next gaps. After large gaps, toward smaller. Eight parameters (one per gap-size bin) captured 83% of the observed anticorrelation, and with a few geometric features (near-diagonal repulsion, banded structure), I hit 93.6%.

This held up under cross-validation. Block bootstrap gave 93.8% ± 1.3%. It was real structure, not overfitting.

But I couldn't explain it mechanistically. I tried replacing the fitted coefficients with per-prime collision features — classifying each gap pair by its residue structure under small primes. The collision features added nothing. The sieve prior already contained all the per-prime information. The row-local tilt was something else.

If not this, then what?

Layer 5: The moiré — interference of the interference

Here's where I had a second geometric intuition that paid off.

That formula is what's called an "Euler product" — it multiplies each prime's contribution independently, like computing fizz and buzz separately and combining them. But here's the thing: when two grids overlap at slightly different angles, you see a moiré pattern — a beat frequency that neither grid has on its own. The fizzbuzz formula can't see that. It computes each prime's blocking pattern in isolation.

So I built features to capture the cross-terms — cos⁡(2πg/p1)⋅cos⁡(2πh/p2)cos(2πg/p1​)⋅cos(2πh/p2​) for pairs of primes. These are the Fourier modes at the beat frequencies where two primes' rhythms interfere.

I tested it. And it worked.

Pairwise cross-terms for primes up to 23, added to the row-local basis, reached 98% lag-1 recovery in-sample and 97.9% out-of-sample. The spectral signature confirmed it: power at beat frequencies (5,11), (3,13), (11,13) was absorbed by the moiré features by 50-96%.

Triple interactions (three-prime compound beats) added another 1.3 percentage points, reaching 99.2% out of sample. Quadruple interactions added nothing — the cluster expansion had converged by third order.

The sieve is a progressively deeper Tetris board.

Each order adds less than the previous. The remaining 0.8% turned out to be the row-local tilt varying with scale — the anticorrelation gets slightly stronger as you move to larger primes — plus noise from sparse large-gap cells.

This was satisfying. But something still felt wrong. Why was the "row-local tilt" — the biggest single correction at +64 percentage points — so resistant to a clean arithmetic explanation? Why did it take 10 features to describe what should have been a simple effect?

If not this, then what?

The feature-chasing trap

At 99.2%, I went looking for the last 0.8%. I tested fourth-order (quadruple) interactions — they added nothing. Finer row-local binning (16, 24, 32, 40 bins instead of 8) — it got worse, not better. Row × moiré interaction terms — nothing. Every new feature I tried either didn't help or actively hurt.

Then I tried making the row-local coefficients depend on scale — a continuous θ(g)+θ′(g)⋅ln⁡(x)θ(g)+θ′(g)⋅ln(x) parameterization. The optimizer produced coefficients with ratios of 400-7000× between base and scale terms. Numerically degenerate. The parameterization was wrong, not the idea.

When I fit each scale band independently, the model class achieved 95-99% at every band. The model was sufficient everywhere. The 0.8% was just the aggregate fit averaging over scale-dependent tilt.

This was the moment I should have stopped adding features. The model wasn't underparameterized. The coordinates were wrong. But I didn't see that yet. I was still thinking "what feature am I missing?" instead of "why does my current frame need so many features?"

The prediction reality check

Before going further, I asked the obvious question: can this model actually predict primes?

The answer was humbling. The full model — 218 features, 99.2% lag-1 recovery — predicted the correct next gap 18% of the time. The marginal model (ignoring the current gap entirely) got 16.5%. Knowing the current gap, with the best transition matrix in the world, bought me 1.5 percentage points of mode accuracy.

The conditional entropy told the story: 3.65 bits of uncertainty per step. That's ~12 equally likely outcomes. Even a perfect one-step Markov model leaves you choosing from a dozen plausible next gaps. The model assigns calibrated probabilities — it knows the shape of the uncertainty — but it can't collapse it.

I had mapped 99.2% of the transition law's serial dependence. That serial dependence was worth 0.34 bits out of ~4 bits of total uncertainty. I'd captured the structure with extraordinary precision, and the structure barely mattered for prediction.

This was the deflation point that made the next insight possible. If all that structure — the sieve, the moiré, the scale correction — only bought 0.34 bits, maybe the structure wasn't fundamental. Maybe it was a projection of something simpler.

Layer 6: The keyhole that kept moving

At this point I'd been modeling the transition law T(h∣g)T(h∣g) — the probability of the next gap given the current gap — in raw gap coordinates. Gap of 6 followed by gap of 12. Gap of 2 followed by gap of 10.

But I noticed something in the Tetris framing. When I conditioned on the board state — which position in the modular cycle the current prime occupies — the same gap of 6 produced completely different next-gap predictions depending on which board you were on:

Board state p ≡ 11 (mod 210): after gap 6, mode = gap 2 (29% probability)
Board state p ≡ 83 (mod 210): after gap 6, mode = gap 8 (27% probability)
Board state p ≡ 113 (mod 210): after gap 6, mode = gap 6 (15% probability)

Same conditioning gap. Completely different predictions. The "law" I'd been modeling wasn't one law — it was many laws averaged together, and the averaging was creating all the apparent complexity.

When I stepped back and looked at the pattern, it clicked: the keyhole isn't moving. My coordinates are.

Raw gap space is a camera-attached coordinate system. The wall (the actual transition law) is stationary, but I was watching it through a moving lens. Every time I added features — row-local bins, geometric bands, moiré cross-terms — I was adding stabilizers to the camera. What I needed was to bolt the coordinate system to the wall.

Layer 7: The comoving frame

The fix was to stop thinking in "gap of 12" and start thinking in "the 3rd legal opening."

At each board state, the small primes block certain candidates. If p≡11(mod30)p≡11(mod30), then p+2p+2 is divisible by... well, it depends on the specific state. But the point is: some gaps are blocked (the next number is guaranteed composite) and some are legal (might be prime). The slot rank — "this is the kk-th legal gap at this board state" — is the wall-attached coordinate.

I built the transition matrix in slot coordinates: T(slotj∣sloti)T(slotj​∣sloti​). One matrix, shared across all 48 board states at W=210W=210.

The result: one universal slot matrix matched 48 per-state raw matrices almost exactly. Mode accuracy 28.7% vs 28.8%. NLL 2.044 vs 2.031. And the universal matrix beat the per-state matrices out of sample, because one well-estimated matrix generalizes better than 48 noisy ones.

The collapse ratio — how much state variation the coordinate change absorbed — was 0.099. 90% of the apparent board-state dependence was a coordinate artifact.

And the universal slot matrix itself was almost trivially simple. Every row was approximately the same. The "which slot am I at" question contributed essentially zero information about "which slot am I going to."

So the full matrix reduced to a single vector: the marginal probability of each slot rank.

And that vector was geometric.

The punchline: one number

P(next prime at slot j)∝rjP(next prime at slot j)∝rj

One parameter. r≈0.705r≈0.705 at W=210W=210.

I tested this across three wheels:

At every wheel, the one-parameter geometric law matched the full transition matrix to within ΔNLL ≤ 0.004. The per-row decay rates were nearly constant. The holdout confirmed at every level.

And then: where does rr come from?

A geometric waiting time with parameter r=1−qr=1−q is what you get when each legal opening has the same independent probability qq of being the next prime. If the primes behave like a Bernoulli process on wheel-legal candidates, the hazard rate should be:

q(x,W)=W/φ(W)ln⁡xq(x,W)=lnxW/φ(W)​

This is just the prime number theorem: primes have density 1/ln⁡x1/lnx among all integers, and among integers coprime to WW, the density is scaled by W/φ(W)W/φ(W).

I fit qpred=(W/φ(W))/ln⁡xqpred​=(W/φ(W))/lnx against the observed qemp=1−rqemp​=1−r across four wheels and six scale bands. Twenty-four measurements.

Correlation: 1.0000. Relative RMSE: 0.3%. Best scaling factor: 0.9987.

When I looked up what this meant, I found that it's called the Cramér random model — apparently the simplest probabilistic model of primes, proposed in the 1930s. The idea is: treat each number as independently "prime or not" with probability 1/ln⁡x1/lnx, and most of the statistics work out. What I'd done was arrive at that same model from the opposite direction — not by assuming independence and computing consequences, but by empirically stripping away layers of apparent structure until independence was all that remained.

What I actually found

I didn't find a new law of primes. I found the coordinates in which the oldest simple model of primes is exactly right for the observable I was studying.

Every layer of complexity I built along the way — the sieve admissibility tables, the row-local mean reversion, the pairwise and triple moiré corrections, the scale-dependent tilt — was a projection artifact. They were real patterns in the data, but they were the distorted image of a simple object viewed through the wrong lens.

The fizzbuzz intuition was correct: primes do emerge from overlapping modular exclusion patterns. But those patterns define the coordinate system (which candidates are legal), not the law (what happens among them). Once you factor out the coordinates, the observable collapses to the wheel-filtered renewal law: each legal candidate, with hazard (W/φ(W))/ln⁡x(W/φ(W))/lnx, well-modeled as an effective Bernoulli process.

And the earlier layers — the row-local tilt, the moiré cross-terms, the triple interactions — were not mistakes. They were correct decompositions of what this simpler law looks like when you project it into raw gap coordinates. In raw space, those structures are real and measurable. The coordinate discovery doesn't invalidate them; it explains why they were needed. I didn't go down blind alleys. I reverse-engineered the projection map.

Scorecard: what the intuition got right and wrong

I started this investigation with a collection of geometric intuitions. Some were vague. Some were wrong. All of them pointed at something real, just not always at the right layer.

The pattern: every directional intuition (there's a frequency, there's orthogonal structure, it's geometric) was right. Every specific mechanism intuition (it's eigenvalues, it's the hazard, the helicoid is the right surface) was wrong. The intuition was consistently correct about the category of the answer and consistently wrong about the instance.

That's probably the most useful thing I learned about how geometric intuition works on hard problems. It's a compass, not a map.

What this is and isn't

This is not a breakthrough in number theory. The model I converged on has been known for nearly a century. The idea that primes look approximately independent on sieve-filtered candidates is apparently old knowledge in the field. A number theorist would have arrived here much faster, and would know exactly which corrections matter and why.

What I think is interesting isn't the destination. It's that someone with no formal training in the field, armed with geometric intuition and a willingness to be wrong, can empirically reconstruct a known result from the other direction — and that the reconstruction path reveals things about the result that the standard derivation doesn't emphasize. I didn't start from "assume independence" and compute consequences. I started from raw observations, built increasingly sophisticated models of apparent structure, and watched all that structure dissolve once I found the right coordinates. The fizzbuzz decomposition, the moiré analysis, the cluster expansion — those were the work of dissolving, not the endpoint.

Every instinct I followed — "it's a frequency," "it's an interference pattern," "there's something orthogonal" — turned out to point at real mathematical objects. The frequency was the chirp. The interference was the sieve. The orthogonal thing was the board state. But none of them were the final answer. They were each one layer of the lens.

The abstraction-first framework I use in my day job — "is this a real quotient or just a re-encoding?" — turned out to be exactly the right question. Each layer of the model was a faithful re-encoding in the wrong coordinates. The coordinate change to slot rank was the actual quotient. And once I had the quotient, the law became trivial.

And the helicoid — the thing that "failed" in Layer 1 — was actually the spatial reasoning that set up the entire arc. The sawtooth wrapping, the modular reset, the idea of tracing an edge that comes back around: that's not just a metaphor. It's what made me think in terms of modular cycles. Modular cycles led to fizzbuzz. Fizzbuzz led to the sieve. The sieve led to the board state. The board state led to slot rank. And slot rank was the coordinate system where the law collapsed to one parameter.

The helicoid didn't give me the answer. It gave me the shape of the question.

The 0.3% residual

There is something left. After subtracting the predicted hazard from the observed one, a residual of about 0.3% remains. It shows a sign structure: at small wheels, the observed hazard is slightly less than predicted; at large wheels, slightly more. Apparently there are known corrections to the simple independence model — things with names I'd have to look up again. That residual is where any genuinely beyond-simple prime structure would live.

I haven't touched that residual yet. It's a different problem — one that starts where this investigation ends.

What I was actually doing

I want to be honest about something. I didn't set out to solve a problem in number theory. I set out to follow my own thinking process to a full conclusion and watch what happened.

I work across design systems, agent tooling, complexity theory — domains that don't obviously connect but that share a common move: find the level of abstraction where the problem simplifies. I wanted to take that move and push it as far as it would go on something hard enough to resist. Primes were the test case. The question was never really "what generates primes?" The question was: can I find the right question to ask?

Each layer of the investigation was a round of that meta-process:

• The helicoid was a frame of reference change — project to a new surface, see what emerges.

• The chirp subtraction was a level of abstraction change — separate the density from the structure.

• The sieve was a relation to other domains — connecting fizzbuzz to number theory.

• The moiré was a cross-pollination — interference patterns from optics applied to modular arithmetic.

• The Tetris framing was an analogy — game state and legal moves applied to prime gaps.

• The slot coordinate was the actual quotient — the point where the abstraction stopped being a re-encoding and became a simplification.

What I learned isn't "primes follow a geometric law." What I learned is that the process works. Each time I changed the frame — the reference, the abstraction level, the analogical connection — it unlocked the next chunk. Not because the new frame was "right," but because it pushed the problem space into the next round. Most frames were wrong. All of them were useful.

The dead ends mattered as much as the hits. The hazard model giving the wrong sign told me the answer wasn't where I expected. The feature trap at 99.2% told me I was optimizing in the wrong space. The GUE falsification closed a door I might have wasted weeks on. Each failure was a constraint that narrowed the search.

And the endpoint — Cramér's model in the right coordinates — is almost anticlimactic. The answer was known. What wasn't known (to me) was the path. The path is the thing.

When your model keeps needing more features to explain the same thing, you might be in the wrong coordinates. Every feature I added — row-local bins, geometric bands, Fourier cross-terms — was compensating for a coordinate mismatch. The model wasn't underparameterized. The space was wrong.

If you can see the shape of what something isn't, you can often find the frame where what it is becomes simple. I started by mapping the negative space. When I found the coordinates that factored out all that negative space, what remained was almost boring.

The right question is worth more than the right answer. I could have looked up Cramér's model on day one. I wouldn't have understood why it's the right model, what it's the right model of, or what it doesn't explain. The two days of wrong turns gave me that understanding in a way that reading a textbook never would have.

I think there's something to be said for approaching hard problems without knowing what "should" work. Not because ignorance is a virtue — I had to learn a huge amount along the way, and a trained number theorist would have seen things I missed. But because not knowing the standard path meant I couldn't follow it. Every step had to come from "what shape am I looking at?" instead of "what technique applies here?" And that led me through the problem in an order that made the coordinate discovery feel inevitable rather than clever.

I'm not a mathematician. I'm a perpetual student who likes finding hard things to work on. The primes were hard enough to resist every wrong frame I threw at them, and specific enough to reward the right one when I finally found it.

To be precise about what this does and doesn't cover: this is about one specific observable — consecutive prime gaps on the support {2, ..., 80} — tested on primes up to 8 million, with wheels through 2310. It is not a claim about all prime behavior, and a number theorist would rightly point out that the known corrections to the simple independence model (which I haven't studied yet) are exactly the things that make prime gaps genuinely interesting at a deeper level.

The next problem is not "find a better main model." The main model is done for this observable. The next problem is: characterize the 0.3% residual after subtracting the wheel-filtered renewal law. That's where the genuinely prime-specific corrections live — the stuff that makes primes not quite independent, the fine structure that the simple model can't see. That's a different investigation, and it starts where this one ends.

The prime knows where it is because it knows where it isn't. I spent two days mapping the isn't. When I found the coordinates where the isn't was factored out, the is was almost boring.

Almost.

There's still that 0.3%.