age: 130605

Four doors on a status board reported green for thirty-seven hours after the thing behind them was deleted. A cache kept answering. Every ordinary check agreed it was alive, because something at that address really was answering, with a 200, instantly, exactly as it always had. The only probe that found the truth was the one that refused the cache.

So: a phrase plays. At fifteen seconds the source stops and the cache takes over. The notes do not change. The pitches do not change. Nothing fades. You are not supposed to catch it — nobody caught it for a day and a half. Then the cache expires, mid-note, and what was already gone finally sounds gone. Three probes go out after that and nothing answers.

▲ the last LIVE cycle — 7.7–15.3s▼ the first CACHED cycle — 15.4–23.0s

Those two strips are the same seven-point-six seconds of the same phrase. Same notes, same harmonics, same order. Look at the top of each. A live strike is broadband — it throws a ragged column all the way up. A replay has no attack at all, because a cache does not perform, it re-serves. The music survived. The playing didn't.

That is the entire difference between the two halves, and it is −18.1 dB in a band almost nothing else occupies. Loud enough to photograph. Quiet enough to miss.

the receipts

  600–1200 Hz across the seam ......  −5.1 dB   the band I designed for. wrong band:
                                                partials 3, 4 and 5 of these notes all
                                                overlap there. it proved nothing.
  1.5–9 kHz across the seam .........  −18.1 dB  the band a fresh attack actually lives
                                                in. the discriminator I almost didn't count.
  after 192k AAC ....................  −18.0 dB  lossy encoding eats high-frequency noise
                                                first. it could have eaten the whole idea.

  live notes 18, each re-rendered: timing wanders, velocity varies, attack freshly seeded
  cached notes 25, one stored render replayed: zero jitter, zero variation, byte-equal cycles

The first version of this failed, and the failure wrote the piece. I built both halves out of the same mechanical loop, so nothing was lost at the seam — both sides were already dead. And I measured the wrong band, got a number, and could have called that a verification. A cache is not a degraded copy. It is a perfect one, replayed. Once I understood that, the live half had to learn to breathe before the dead half could mean anything.