Skip to the result

ResearchTechnical note 6

900 new laws, each learned exactly

Learning is where AI spends its time, its data and its energy. Given 900 laws it had never seen, METRON learned every one of them exactly, each from 192 bytes of data, in about half a millisecond.

Measured
6 November 2026
New laws
900
Families
3
Record
113

At a glance

900

new laws learned with no wrong output, 300 in each of three families.

0.54ms

median time to learn one law.

192bytes

of observed data are enough to learn a law.

In plain words

We handed the system 900 different rules it had never met, one at a time, each hidden behind a short stream of data. It worked out every rule, and it did so faster than a camera shutter. There was no rule it got partly right.

For engineers

900 fresh laws drawn from three families of bitwise functions, generated with new seeds. Each is discovered from 192 observed bytes and then scored; success means 100 % of outputs. Lower bound of the success rate at 95 %: 99.59 %. Control: scoring each law against another law’s data.

Why it matters

Training is the largest cost in AI and it is paid again for every new task. Here a new task costs a fraction of a millisecond and a few hundred bytes, and the result is exact instead of approximate.

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

Every new task means new training.

A neural network learns a task from large amounts of data, over many passes, and ends close to right. A second task means doing it again. Whether it will succeed on the next one is known only after paying for the attempt.

An earlier result showed our system learning 90 laws out of 90. The question here was whether that held on ten times as many, all new.

Result

Nine hundred attempts. Nine hundred exact.

Laws learned at 100 %, by family

Family A300 / 300
Family B300 / 300
Family C300 / 300
Control: wrong law paired0 / 900

Time to learn one law, in milliseconds

02755500.40.50.60.70.80.91.0+

Fig. 1 Left: 900 laws the system had never seen, 300 in each of three families, all learned without a single wrong output. Right: how long each of the 900 took. The median is 0.54 ms.

Table 1 Learning 900 new laws.
MeasureValue
New laws learned at 100 %900 of 900
Lower bound of the success rate, 95 % confidence99.59 %
Data observed to learn one law192 bytes
Time to learn one law, median0.54 ms
Time to learn one law, slowest of the 9002.45 ms
Earlier 90 laws, reproduced90 of 90

Method

How a law is learned

  1. Data stream192 bytes produced by a hidden law
  2. METRONfinds the law that produced them
  3. Checkevery output must match
  1. A new law is drawn and used to generate a short stream of data. The system sees the data, never the law.

  2. The system discovers the law that produced the data.

  3. The law is scored. It counts only if every output is right.

Controls

What was checked

Wrong pairing

Scoring each learned law against a different law’s data gives 0 of 900 at 100 %, with a median agreement of 0.59 %. A law only fits its own data.

Reproduction

The earlier 90 laws were run again first and matched the sealed result law by law.

Statistical bound

With 900 successes in 900, the true success rate is above 99.59 % at 95 % confidence.

Fresh laws

The 900 were generated with new seeds; any that repeated an earlier law were discarded before the run.

Open questions

What this opens

  • More families of laws, and laws with more inputs.
  • Learning from data with noise and missing values.

Fields of application

Where it fits

  • Systems that must adapt in the field, without a training cluster behind them.
  • Protocol, signal and logic analysis, where the rule behind the data is the product.
  • Hardware that learns its own function after it has been manufactured.

Record

How this result is kept

Record
113
Run
6 November 2026
Sample
900 new laws, 300 per family; 90 earlier laws reproduced
Measurement file
SHA-256 5382367ad0a3adf33f429f55967fe1cf734cdb60d54061afab1f6c533051b7aa

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.