Skip to the result

ResearchTechnical note 4

From a neural network to exact laws

A trained neural network is an approximation nobody can read. Given only the inputs and outputs of a network that is right 99.575 % of the time, METRON writes out the eight exact laws behind it, and proves each one.

Measured
14 November 2026
Network
871k parameters
Laws extracted
8 of 8
Record
119

At a glance

8

exact laws extracted from the network’s behaviour, each proved by exhaustive enumeration.

100%

accuracy of the extracted laws, from a network that reaches 99.575 %.

7.6s

to extract all eight laws.

In plain words

Imagine a student who has memorised thousands of worked examples and gets almost every exercise right, but cannot tell you the rule. METRON watches the student answer, works out the rule, writes it down in one line and checks that the rule is never wrong.

For engineers

Black-box setting: only the network’s inputs and outputs are used, 8,192 samples. METRON returns one closed-form Boolean law per output bit; each is verified against the ground truth on every input of its support. Baseline: a decision tree on the same data.

Why it matters

Trained networks are expensive to run and impossible to certify. Turning one into a few exact lines of logic changes both: it can be audited, proved, and executed at a fraction of the cost.

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

A network that is 99.6 % right cannot tell you which answers are the wrong ones.

Millions of parameters approximate a function; none of them states it. The network works, mostly, and nobody can say exactly what it computes or when it will fail.

For a product, that is a ceiling. Something that is almost always right cannot be certified, and it has to be executed at full cost on every request.

Result

The network approximates. The laws are exact.

Exact laws recovered, out of 8

METRON8 / 8
Decision tree0 / 8
Scrambled network0 / 8

Correct outputs

Extracted laws100 %
Neural network99.575 %

Fig. 1 Left: how many of the eight output laws each method recovers exactly from the network’s behaviour. Right: the laws that come out are exact, where the network they came from is not.

Table 1 The network and what was extracted from it.
MeasureValue
Network accuracy, by output byte99.575 %
Laws extracted and proved exact8 of 8
Accuracy of the extracted laws against the truth100 %
Agreement between the laws and the network99.92 to 99.93 %
Decision-tree baseline, accuracy99.57 to 99.87 %, none exact
Time to extract the eight laws7.6 s

Where the laws and the network disagree, it is the network that is wrong. The extraction also shows on which inputs it errs.

Method

How a network becomes laws

  1. Neural network871k parameters, right 99.6 % of the time
  2. METRONobserves inputs and outputs only
  3. Eight exact lawseach proved on every input
  1. The network is queried, like any black box. Its weights are never read.

  2. METRON searches for the law that explains each output bit, among laws that can be written down exactly.

  3. Each law is proved. It is compared with the truth on every possible input it depends on. Only laws that pass are kept.

Controls

What was checked

Exhaustive proof

Each of the eight laws was compared with the true function on every input of its support. There is no sampling in this step.

Scrambled network

With the network’s answers shuffled, METRON accepted no law at all. It does not find structure where there is none.

Standard baseline

A decision tree trained on the same data recovered none of the eight laws exactly.

Ground truth

Run directly on the true function, METRON recovers the same eight laws.

Open questions

What this opens

  • Extraction from public, pre-trained networks, beyond a network we trained ourselves.
  • Functions with more inputs per law, and networks of larger size.

Fields of application

Where it fits

  • Certification of machine-learning components in regulated products.
  • Compressing a trained network into logic that runs on small, cheap hardware.
  • Auditing: seeing exactly where a deployed network gives the wrong answer.

Record

How this result is kept

Record
119
Run
14 November 2026, 3 min 31 s
Sample
8,192 samples to extract; verification on every input of each law
Measurement file
SHA-256 2326799ff6e6b610b5a33d5a375e210a8d21198e0745c16d60e09d0c15c5897d

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.