Skip to the result

ResearchTechnical note 5

One million requests. Zero errors.

An AI system is judged in production by two things: how often it is wrong and how long it takes. METRON answered a million requests with no error, in a median of 17.6 millionths of a second each, on a single CPU core.

Measured
6 November 2026
Requests
1,000,000
Hardware
One CPU core
Record
114

At a glance

0

errors in 1,000,000 requests, 256,000,000 bytes checked one by one.

17.6µs

median time to answer a request, with no GPU.

240bytes

is the size of the whole artefact that answered: 60 laws.

In plain words

We asked the system a million questions in a row and checked every answer. All were right. Each answer took about 18 millionths of a second, on one core of an ordinary processor, and the thing doing the answering is smaller than this paragraph.

For engineers

60 sealed laws in a 240-byte artefact, served on one CPU core with no GPU visible. 1,000,000 requests in batch plus 100,000 one at a time; every output byte compared with the reference. Latency: p50 17.6 µs, p99 25.8 µs, p99.9 34.1 µs.

Why it matters

Inference is where AI spends its money. A component that answers in microseconds on commodity CPU, fits in a few hundred bytes and does not err changes the unit cost of every request it takes over.

MTRMETRON is AIDON’s exact co-processor. It learns laws from data, proves them and runs them, alongside a neural model or on its own.

The bottleneck

Fast and cheap, or accurate. Rarely both.

Serving a neural model means a GPU, a queue and a latency measured in milliseconds, for an answer that is right most of the time. Every improvement in accuracy tends to cost more compute per request.

For the part of the work that has an exact answer, that trade is unnecessary.

Result

Every answer checked. None wrong.

Time to answer one request, in microseconds

Median17.6 µs
99th percentile25.8 µs
99.9th percentile34.1 µs

Scale: 0 to 40 µs. One microsecond is a millionth of a second.

Fig. 1 Response time over 100,000 requests answered one at a time, on a single CPU core. 999 requests in 1,000 are answered in under 34 µs.

Table 1 The endurance run.
MeasureValue
Requests answered1,000,000
Wrong requests0
Bytes checked, wrong bytes256,000,000, 0
Time per request, in batch0.98 µs
Median, first tenth and last tenth of the run17.57 µs, 17.58 µs
Size of the whole artefact240 bytes, 60 laws
HardwareOne CPU core, no GPU

Method

How the run was done

  1. Requests1,000,000 in a row
  2. METRON60 laws, 240 bytes, one CPU core
  3. Byte-by-byte checkevery output against the reference
  1. The artefact is the sealed one. Its fingerprint is checked before the run starts.

  2. Requests are served in batch and one at a time, so that both throughput and per-request latency are measured.

  3. Every output byte is compared with the reference. A single wrong byte would count as a wrong request.

Controls

What was checked

The check can fail

With each law swapped for another, errors were detected in 60 of 60 blocks. The comparison is not blind.

Sealed artefact

The file served is the sealed one: its fingerprint matched before the run.

No drift

The median in the first tenth of the run and in the last tenth: 17.57 and 17.58 µs.

No GPU

The run was confined to a single CPU core, with no GPU visible to the process.

Open questions

What this opens

  • Service under concurrent load, across many cores.
  • The same measurement on a dedicated chip instead of a general-purpose CPU.

Fields of application

Where it fits

  • High-volume inference, where cost per request decides the margin.
  • Real-time control and edge devices, with no GPU and a hard deadline.
  • Any pipeline step that today calls a large model for something that has an exact answer.

Record

How this result is kept

Record
114
Run
6 November 2026
Sample
1,000,000 requests in batch; 100,000 one at a time
Measurement file
SHA-256 b386aec2a2c3c5418041a1ef15867c1c7d60c428e7457b1ed051221fa9af5ad9

Every figure on this page was read by a program from the raw measurement file, not typed. The fingerprint identifies that file: when it is disclosed, anyone can check that it is the same one.

Contact

Talk to the founders

Get in touch. We would be glad to share our ideas and walk you through the results, the proof and the measurements.