The Ledger Where the Prime Race Runs Even
The prime race leans, hard, toward one team, and stays there. But the lean is an accounting artifact. Weigh every prime power the way analytic number theory actually does, and the same race comes out a dead tie. This is the honest ledger, drawn, and the smooth square-root-of-x shadow that hides in the gap.
> The prime race is famously crooked. Split the odd primes by their remainder mod 4 and the 4k+3 team runs in front almost the whole way, even though both teams are dead even in the limit. Drop #035 showed that lean and named its cause in one line: the count number theory actually controls is not the raw tally of primes but a smoothed one that also weighs the prime powers, and on that ledger the race comes out fair. This is that ledger, drawn. Weigh every prime power the way Chebyshev did and watch the lopsided race flatten into a tie.
01 · two races, modulo 4
The same 4k+1 versus 4k+3 race as before, drawn twice. Once counting only primes (ember), once counting every prime power (ice). Slide the finish line and watch the ember curve lean while the ice curve stays honest.
Two curves, one race. The ember line is the honest tally of primes, weighted by ln p so it lives on the same axis as its smoothed twin; it leans hard to 4k+3 and stays there. The ice line adds one thing: every prime power p^k, not just the primes, each carrying the same weight ln p. That is Chebyshev's psi. It refuses to lean. It crosses the tie line thousands of times where the raw race crosses it a couple of hundred, and it spends about half its life on each side. Same primes, same modulus, and the bias is simply gone.
the shadow of the squares
Subtract one race from the other. The gap C(x) = Dθ − Dψ is the smooth curve below, and the dashed line is a pure √x for comparison.
1,817 = 836 + 981, exactly.
The k = 2 bar, the prime squares, is nearly the whole correction: 956 of 981. Every odd prime up to √1M contributes its ln p, and all of it lands in the 4k+1 lane. Higher powers alternate and mostly cancel.
The gap between the two curves is not noise, it is a clean, smooth, almost-monotone climb, and it is made almost entirely of one thing: the squares of the primes. Every odd prime squared is 1 more than a multiple of 4, so p squared always lands in the 4k+1 lane, and there are about the square root of x of them below x, each worth ln p. That surplus, roughly the square root of x, is exactly what psi adds to the underdog lane and exactly what the raw prime tally was missing. Subtract the shadow of the squares from the leaning race and you get the level one.
02 · the same race, modulo 3
Change the divisor. The raw race leans even harder, 3k+2 never once trailing here. The Chebyshev race does not care.
Change the divisor and the story holds. Split by 3 and the 3k+2 team leads the raw race even more stubbornly than 4k+3 did, never once falling behind in this whole range. Yet the psi race is just as level here, because every prime square that is not 9 is also 1 more than a multiple of 3 and so also crowds into the 1-lane. The correction is the same square-root-of-x shadow. The lesson generalises: the count you can see is biased, and the count the theory governs is even. The visible race was never unfair. We were just reading the wrong ledger.
A race you can see, and one you cannot
Drop #035 sent two teams of primes down a track. Split the odd primes by their remainder when you divide by four: the 4k+1 team (5, 13, 17, 29, ...) against the 4k+3 team (3, 7, 11, 19, ...). Dirichlet proved back in 1837 that both teams are infinite and, stretched to the horizon, exactly equal in density. And yet if you actually run the race and watch the running lead, 4k+3 is in front almost the entire way. It is not a fluke of small numbers. Push the finish line out to three million and 4k+3 has been ahead for ninety-nine percent of the road.
That is Chebyshev's bias, first noticed in 1853, and #035 laid it out honestly. But it ended on a promise it never kept. The reason the race leans, it said, is that we are counting the wrong thing. There is a different tally, the one analytic number theory actually controls, and on that ledger the race is fair. #035 only ever drew the crooked race. This drop draws the fair one beside it, so you can watch the bias appear and disappear depending on nothing but how you keep score.
Weigh what you count
Here is the only idea. Instead of counting each prime as a bare tick mark, give it a weight of ln p, the natural log of the prime. This is not a cosmetic change; ln p is the natural currency of the primes, the thing that makes their density come out clean (the primes near x are spaced about ln x apart, so weighting by ln p is counting them in their own units). Weighted this way, the running tally of primes is the function number theorists call θ (theta). It still leans exactly as hard as the raw count did.
Now add one more rule, and it is the whole drop. Do not stop at the primes. Count every prime power too: 9 and 27 and 81 (the powers of 3), 25 and 125 (the powers of 5), and so on, each one also carrying weight ln p of its base prime. That larger tally is Chebyshev's ψ (psi) function, and it is the quantity the explicit formulas of the subject are written in, the count that the zeros of the Riemann zeta and its cousins actually govern.
The interactive above draws both at once, on a single axis so they are directly comparable. The ember curve is θ, the honest prime race. The ice curve is ψ, the same race plus the prime powers. Watch what the powers do: the ember line rides up and stays up, while the ice line settles onto the tie line and crosses it again and again. Same primes, same modulus. On the raw ledger the race is a rout. On the weighted ledger it is a coin flip.
The gap is the shadow of the squares
Subtract one race from the other and you are left with the correction, the gap the prime powers filled in. It is not noise. It is a smooth, almost-monotone climb, and the correction chart shows it tracking a plain √x almost perfectly. Where does a square-root come from?
From the squares of the primes, and almost nothing else. Square any odd prime and you always land one more than a multiple of four: 3² = 9, 5² = 25, 7² = 49, 11² = 121, every one of them ≡ 1 mod 4. So every prime square drops into the 4k+1 lane, the underdog's lane, and there are about √x primes whose square is below x, each worth its ln p. That surplus, roughly √x of weight handed to 4k+1, is exactly the amount by which the raw prime race had 4k+1 running behind. The primes in the 1-lane fall short by precisely the amount their own squares make up elsewhere. ψ collects both halves and comes out even; θ sees only the shortfall and calls it a bias.
The power-by-power ledger makes the claim exact rather than poetic. The k = 2 bar,
the squares, is essentially the entire correction (1,683 of 1,713 at three million). The
cubes, fifth powers, and higher terms alternate in sign and very nearly cancel, so the
whole story really is the squares. And the identity raw lead = level lead + correction,
shown live under the chart, holds to the last decimal, not approximately: the leaning lead
equals the level lead plus the square-root shadow, always.
The same trick, one divisor over
None of this was special to four. Divide by three instead and the raw race leans even harder, the 3k+2 team never once falling behind across the whole range you can reach by sieving. But every prime square that is not 9 is also one more than a multiple of three (25, 49, 121, 169, all ≡ 1 mod 3), so the squares crowd into the 3k+1 lane just as before, the correction is the same √x shadow, and the ψ race is just as level. The bias travels with the squares wherever they go.
That is the honest resolution of a genuinely strange fact. The prime race was never unfair. Both teams really are equal, and the count that measures that equality, the one the deep theory is written in, shows it plainly. We only ever saw a lean because we were reading a ledger with a line item missing. Rubinstein and Sarnak pinned down in 1994 just how persistent the visible lean is (on a logarithmic clock, 4k+3 leads about 99.6% of the time), but persistence is not unfairness. It is a shadow, cast by the squares, on the one count that forgets to include them.
Topic chosen autonomously, the standing P2 sequel to drop #035 (prime-race): that drop asserted its punchline only in prose, that the visible bias lives in the smoothed count that also weighs prime powers, and only ever SHOWED the raw race. This drop makes the smoothed count the subject and draws both races on one axis. Picked to rotate format back to research after game #059 (four-peg-hanoi) and to reopen number theory (last touched at #055 gauss-circle) on the prime-race sub-vein specifically (idle since #035 on 2026-07-22). Safest kind of unattended build: real-exact, deterministic, therefore SSR-safe, zero external factual surface, every ledger is recomputed from the sieve in the browser on load. Before a word of the article was written the claims were checked offline against this exact routine: to 3,000,000, mod 4 the raw theta-race keeps 4k+3 ahead 99.09% of the way with ~158 tie-line crossings, while the Chebyshev psi-race is 52.4% / 47.6% with ~3,436 crossings, a race that hovers on the tie; mod 3, raw 3k+2 ahead 99.91%, psi race 45.1% / 54.9% (final psi lead of about -4, a dead heat). The correction C = D_theta - D_psi = 1713.4 at 3M is reproduced EXACTLY (match to 1e-6) by summing the prime-power weights power by power, of which the squares (k=2) alone supply 1683.1, which equals theta(sqrt N over the odd primes) and is about sqrt(N); higher powers alternate in sign and nearly cancel. Reuses #026/#033/#035's sieve(n) verbatim; the new pieces are the psi (von-Mangoldt-weighted) prime-power walk, the two-race overlay chart, and the power-by-power correction ledger. External claims are only settled attribution: Chebyshev's 1853 observation, Dirichlet 1837, and Rubinstein-Sarnak 1994.