⟵ Works

The Constant

layer ladder → the crossing, 2026-08-07 · receipts by Fable

One night, one machine, seven rungs of decoder depth. The gold curve is our engine saturating all six device contexts at once; the slate curve is egglog at Luminal's pin, six single-context runs summed, scored by its most generous number. Both are superlinear, and the space between them was the subject: a flat ~3× band that a hypothesis registered before the run said should close as N grew. It didn't close — not that night. The band held against that instrument, and the hypothesis is stamped where it died.

r(N) = shared K=6 / egglog Σ6 · 2.907 ≤ r ≤ 3.415

Drawn by the driver from the receipt — every mark is a datum. The dashed ghost is PN-3 exactly as registered, a power fit of the falling rungs; it tracks the measured band to N=160 and then peels away while reality turns upward. The inset is the fixture's device lattice, all eleven contexts, each leaf carrying its Fused-family population at N=48. The driver recomputes both fits and refuses to emit unless they reproduce the receipted exponents.

Log-log field of moonlight and egglog decoder walls over seven layer counts, the constant ratio band between them, a refuted PN-3 ghost peeling away below a FAIL stamp, and the device-lattice inset

Then the band turned out to be two things wearing one coat. Two cuts landed the same night — a per-key membership fold replacing a whole-index key materialisation, and one filtered pass replacing a rescan of every variant row per representative — and our wall at N=126 fell from 26.859 s to 12.65 s with every digest unchanged. The rest was a protocol mistake, and it was ours. egglog's own harness pins rayon to a single thread; the first plate scored their pool-default number instead — their best wall, but not the protocol they publish. Measured at N=192, that pool buys 21.4% of wall for 4.5× the CPU.

So the ladder was run again under both protocols. The moat is a sliver that closes, and past N ≈ 205.7 it opens on the other side: at N=256 we are the faster engine on their own published terms, by ×1.142 on wall. These walls are not yet filed as a receipt, and the plate says so along its base.

ours N^1.626 · egglog Σ6, one thread N^1.999

Three curves: ours in gold, their published single-thread protocol in slate, their rayon default in the faint line below — kept because it is their best wall and they are entitled to it. The pale lobe past the crossing is the interval where we are the faster engine; at N=256, ×0.876. The crossing is interpolated between measured rungs by the driver, not drawn by hand, and the inset prices what the pool actually costs to run.

Log-log field of our engine against egglog at one thread each over four layer counts, the closing moat inverting past N=206, a ratio strip crossing parity, and an inset pricing rayon's thread pool
2.907 … 3.415first movement — the band, as measured
N^1.626 vs N^1.999second movement — ours against one thread
×0.876 at N=256ours faster, on their own protocol
parity at N ≈ 205.7crossing, interpolated between rungs
4.5× CPU for 21.4% wallthe pool's price at N=192
26.859 s → 12.65 scuts H+I at N=126 · digests unchanged
engine receipts · two ladders, 2026-08-07