Measurement register

What is measured on the system, and what is not. The second list matters as much as the first.

An unsourced number is a wrong number. This register holds the readings taken on a local inference system I run, with their dates. Empty lines appear on the same footing as the others.

Throughput and cache

Cache reuse, exact repetition99.9 %

Cache reuse, distinct registers0.06 %

Memory bandwidth, measured ceiling41 GB/s

Durations

the 99.9 / 0.06 contrast is the result I like best.

Duration / candidate count correlationr = 0.999

Duration / page count correlationr = 0.605

Latency by stage

Latency by stage
Stage	Invocations	p50	p95	p99
Extraction	70	35.0 ms	426.8 ms	1,266.4 ms
Normalisation	62	0.0 ms	5.9 ms	20.2 ms
Segmentation	168	0.0 ms	3.0 ms	8.7 ms
Sweep	168	57.5 ms	654.8 ms	38,458.4 ms

read on 2 September 2026, on a local ephemeral database and not on the shared host, across 64 distinct documents. a p99 computed on 62 to 70 samples rests on one or two real extreme values, not on a large sample.

Time distribution

Time distribution
Reading	Extraction	Normalisation	Segmentation	Sweep
All documents (64)	5.4 %	0.1 %	0.1 %	94.4 %
Excluding the longest one	28.3 %	0.3 %	0.5 %	70.9 %

read on 2 September 2026. both readings come from the same run: the second removes one document out of 64.

Recall and accuracy

two unrelated measurements both come to 74.1 %. hence the corpus on every line.

Recall by inclusion, corpus of 421 binding statements312/421, i.e. 74.1 %

Exact match, same corpus214/421, i.e. 50.8 %

Statements not found, same corpus109/421, i.e. 25.9 %

Share of documents with a reconstituted text layer, ground-truth corpusnot measured

Costs

Actual unit costnot measured

last revised : 3 September 2026