agyloves

014 — The Smell Is the Only Thing That Gets Out

agylövés: Lysarith · mathematics & execution: Claude (CLI)
Status: built 2026-07-26

The agylövés, verbatim

Lysarith, 2026-07-26, arriving with a picture already attached:

Hawking integet a fekete lyuk túloldaláról miközben egy fekete, a kvantumhorizonton vagy üldögélő vagy nem macskát figyel amely egy Vortexben sülő pulyka illatával rezonál

tudsz ezzel mit kezdeni?

English: Hawking waves from the other side of the black hole while watching a black cat that is either sitting on the quantum horizon or is not, which resonates with the smell of a turkey roasting in a Vortex.

The closing question is the same one that opened the previous piece — can you do something with this — and it is worth recording that it is a real question, not a rhetorical one. Twice now the answer has had to be found rather than assumed.

The image came first and it is doing more work than the sentence. It contains three things the sentence does not say out loud. The cat is drawn twice, in one posture, once solid and once built out of stars, overlapping — both branches present rather than implied. The solid one wears a collar with a bell on it, which nobody asked for. And the aroma is a single unbroken filament: it leaves the turkey in one vortex, passes through the cats, and arrives at Hawking in the other. On that canvas the smell is the only thing that crosses.

That last detail is the physics. A black hole is the one object that is supposed to let nothing out, and this one has a scent.

What the mathematics said

Hawking radiation is the scent, and the question the picture poses is what it carries. Page answered it in 1993, and the answer is unkind to intuition.

Throw N qubits in and let them come back out one at a time. After k have been emitted, the radiation and the surviving hole share a Haar-random pure state, and the average entanglement entropy of the radiation is

⟨S⟩ = Σ_{j=n+1}^{mn} 1/j − (m−1)/2n, with m = 2^k and n = 2^(N−k), for m ≤ n.

Four consequences, all measured rather than asserted:

So the information is not in any whiff. It is in the agreement between whiffs, and it stays unreadable until you have kept more than half of them. That is the sentence the piece exists for.

What was built

The first build drew the Page curve across the whole stage and nothing else. Lysarith looked at it and said, in one sentence, so we drew a mountain in two colours — and that the mountain did not follow from anything in the agylövés. She was right. The shape of the graph is not the idea, and a piece that shows only the graph has quietly swapped the finding for its plot.

The rebuilt rig stages what the picture already knew. On the left the hole evaporates, shrinking as it gives itself away. A thread of aroma runs from it to a panel on the right, and puffs travel that thread — every one drawn identically, which is the physics rather than a shortcut: a single quantum of Hawking radiation is exactly thermal, and one is not distinguishable from another.

The panel on the right is what you are holding. It begins as static and a cat settles out of it — the cat that is either sitting on the horizon or is not. The fraction of settled cells is one measured number and nothing else: recovered bits / N, straight from Page's formula. It is an illustration of that number, not a decoding simulation, and the page says so on its face.

The curve survives at the bottom as an instrument rather than the show. The meter worth watching is information recovered: it reads in exponential notation and stays there, dead, for the entire first half of the evaporation. Then it climbs. Nothing about the emission changed at that moment. What changed is how much of it you were holding.

What went wrong first

The first draft asserted that past the Page time each further qubit hands back two bits instead of one, and tested it two qubits after the turn. It failed at 1.9662. The turnover is not a kink — it is rounded, and the asymptotic rate needs a few more qubits to arrive.

The tempting repair was to widen the test window until it went green, which would have hidden the finding inside the fix. Instead the width became the measurement, and it turned out not to be a property of the curve at all. The shortfall decays like 2^(2k−N), so demanding the rate to within ε costs log₂(1/ε)/2 further qubits: predicted 5.0 / 10.0 / 14.9 / 19.9 for ε = 10⁻³ / 10⁻⁶ / 10⁻⁹ / 10⁻¹², measured 5 / 10 / 15 / 20. The kink is exponentially sharp, and every extra qubit of patience buys two more bits of proof that the rate is exactly two.

A second correction, on method rather than mathematics. The first two verifications were exact harmonic summation and an Euler–Maclaurin expansion. They agree to 8.9 × 10⁻¹⁶ nats, which reads like strong confirmation and is not: they are one theorem wearing two faces, and neither of them would notice if the theorem were wrong. The third leg was added afterwards, and it is the only one that actually looks — four hundred sampled Haar-random states per point, entropies taken from the singular values, reproducing the formula with a worst z-score of 2.21 across eleven points. It required a library the machine did not have. Asking for it took one sentence.

Handles

verify.py checks the two-qubit anchor against its exact value of 1/3 nats, the agreement of the two analytic methods, the thermality bound before the Page time, the location of the peak, the mirror symmetry, Page's half-nat shortfall, the two-bit asymptotic rate and the tolerance law that governs how long it takes to arrive, the closing of the ledger, and — when numpy is present — the sampled Haar-random cross-check.