Skip to the result

ResearchTechnical note 10

Learned laws compile to plain bitwise code

What a neural network learns stays inside the network. What METRON learns can be written out as plain bitwise code, the kind a chip executes directly, and the compiled code gives the same answer on every bit.

Measured
8 November 2026
Target
Bitwise code
Benchmarks
ISCAS’89 and others
Record
§174

At a glance

0

bits of difference between the learned laws and their compiled form, on every source tested.

8.7×

faster after compiling, with the same answers.

5

standard benchmark circuits from the chip-design literature among the sources.

In plain words

A recipe in someone’s head and the same recipe written on a card produce the same dish, but only the card can be handed to a factory. METRON’s knowledge can be written on the card: short, exact instructions that any chip can follow.

For engineers

Learned laws are compiled to straight-line bitwise code (AND, XOR, shifts). Equivalence is checked bit for bit against the uncompiled laws on the learning benchmark, the ARX units, arithmetic, shifts and multiplexers, and laws learned from five ISCAS’89 circuits.

Why it matters

Compilation is the bridge from a learning result to a product. It turns what was learned into something that runs on any CPU today and maps directly to gates tomorrow.

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 trained model cannot leave the machine that runs it.

The knowledge in a neural network exists only as weights, and using it means running the whole network: a runtime, a framework, usually a GPU.

For hardware, that is the wrong form. A chip needs logic: a fixed, small set of operations on bits.

Result

Same answers. Simpler form. Faster.

Time to serve one request, in microseconds

Before compiling167 µs
After compiling19.2 µs

Fig. 1 The same learned laws, served before and after compilation. Compiled, they answer 8.7 times faster, and not one output bit changes.

Table 1 Compiled code against the laws it was compiled from.
MeasureValue
Laws of the learning benchmark90 laws, 0 bits differ
ChaCha20 and SipHash building blocks7 units, 0 bits differ
Addition, subtraction and an increment sequence96 laws, 0 bits differ
Industry benchmark circuits (ISCAS’89: s400, s526, s1423, s9234.1, s13207)118 laws, 0 bits differ
Shifts, rotation and selection, compiled8,192 of 8,192 on each test
Control: one law sabotaged on purpose4,122 bits differ

Method

How the laws are compiled

  1. Learned lawsexact, one per output bit
  2. Compilerwrites them as bitwise operations
  3. Bit-for-bit checkcompiled against learned
  1. METRON’s laws are taken as they are, with no retraining and no approximation.

  2. The compiler writes each law as a short sequence of bitwise operations.

  3. Both forms are run on the same inputs and every output bit is compared.

Controls

What was checked

Sabotage control

With one law deliberately corrupted, the comparison reports 4,122 differing bits. It detects what it should.

Independent sources

Arithmetic, standard algorithms and benchmark circuits were all compiled; none lost a bit.

Standards again

The compiled ChaCha20 still matches the RFC 8439 vector and the compiled SipHash its 64 reference vectors.

Speed measured

Served on the same machine: 19.2 µs per request compiled, against 167 µs before.

Open questions

What this opens

  • From bitwise code to a gate-level netlist and a physical layout.
  • Optimising the compiled logic for area and power.

Fields of application

Where it fits

  • Chip design: logic generated from learned behaviour.
  • Embedded software, where a few bitwise lines replace a model runtime.
  • FPGAs and accelerators that need exact, compact logic.

Record

How this result is kept

Record
§174, measured
Run
8 November 2026, 17 min 09 s
Sample
90 + 7 + 96 + 118 laws compared bit for bit
Measurement file
SHA-256 b4129a98a59ef46d9e06148d8967e1c84d1818c284848d2eae80f77fc45c167e

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.