Evidence
What we have measured and how. Every record states its method once, checks that every engine returns the same result before timing, and lists where we lose.
All measurements
-
What attention costs
At the same shape and the same parameter count, reading a 1,984-token prompt takes 182 ms with softmax attention and 7.08 ms with holographic attention. Each further token takes 181 µs against 3.60 µs, and the serve cache holds 2,253 KB against 3.40 KB. Softmax reading grows with the square of the length (exponent 1.9) and holographic reading grows in step with it (1.0).
-
What a gate bound does
The gate never admitted more downloads than its bound: in every recorded run the most in flight equalled the bound. The time to finish 24 queued jobs fell from 1,113 ms at a bound of 1 to 139 ms at 8, then rose again to 410 ms when all 24 ran at once, because the simulated service slows when it is crowded.
-
HoloDb benchmarks
On a 10,000,000-invoice ledger joined to 200,000 customers, the dashboard rollups read in microseconds in the embedded engine, and every result is identical across the engines. The queries where HoloDb is slower are listed beside the wins.
-
Detail against components read
One 16 by 16 chunk stored as a single hologram decodes 98.4 percent of its cells correctly at the full budget of 8,192 components, 131,072 bytes and 17.0 ms. Reading fewer components is cheaper and fuzzier: 67.1 percent at 1,568 components, and 19.5 percent at 24, which takes 384 bytes and 0.09 ms.
-
Exact recall and refusal
At equal storage the exact key view recovered 64 items with no errors where holographic reduced representations held 95 percent or better only up to 4. Past capacity the exact store refused every time, while the baseline answered wrongly every time.
-
Pleroma costs and checks
A superposed kernel applies in 0.17 to 0.44 ms on a 10,464-cell store whether it holds 1 or 10,368 elements, and the repair and corruption checks recovered or detected every case tried.
-
Generated sentences pass the gate
Of 2,000 sentences requested in one run, 1,966 different ones came back. Every one passed the package's validity gate on a fresh parse, and every word in every one was observed in the corpus.