TechnologyMETRON
Exact answers, inside the model
METRON is AIDON’s exact co-processor. It learns the law behind the data and then answers exactly.
- System
- METRON
- Emblem
- MTR
- Role
- Exact co-processor
- Evidence
- 14 technical notes
Evidence
What is already measured
Four numbers, each taken from a result published on this site.
of operations exact with METRON, up to 100 decimal digits.
exact answers in a row, from 64 to 4,096 bits, with laws learned at 32 bits.
requests answered with zero errors.
new laws learned, every one exactly.
What it does
Three things, each one measured
Answers
Exact inside a public model
A public language model, with and without METRON, AIDON’s exact co-processor: exact answers in long arithmetic, with the model’s weights untouched.
Proved
Correct for any bit width
Addition, subtraction, comparison and multiplication of METRON, proved correct in Lean 4 for every bit width.
Learns
New laws, each one exactly
METRON learned 900 of 900 new laws at 100 %, each from 192 bytes of observation, in a median of 0.54 ms.
Comparison
Against a neural network
The same task, on the same machine.
Throughput
One CPU core against a GPU
METRON’s compiled laws on one CPU core serve about 27 times more requests per second than a neural network on an RTX 5090, at 100 % accuracy.
Data
The same laws, from far less data
On the same laws, METRON needs about 224,000 times less data and 13,000 times less time than a neural network of 871k parameters, and learns them exactly.
Structure
From a network to exact laws
From the inputs and outputs of a neural network of 871k parameters, METRON extracts the 8 exact laws behind it, each proved by exhaustive enumeration.
less energy to answer a request than a neural network: one rack against 960.
Read more
Where to go next
Research
The results, one by one
Each result has its own technical note, with what was measured and how it was checked.
Benchmark
GNOMON
AIDON’s benchmark for exact generalisation, in preparation for public release.
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.