Skip to the result

AIDON Technologies · Research

Exact arithmetic, as an organ of the model

We wired METRON, a small, formally proven co-processor, into a public language model. Alone, the model fails long arithmetic. With METRON, the answers are exact.

Each result on this page is sealed. “Sealed · #118” is the entry in AIDON’s results ledger where the figures are recorded together with their raw measurement file.

Multiplication · two 50-digit numbers

92553211483448741871128839965279029188482722789736 times 73136452087870836049438311436517966110515469013597

Model alone

6763000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000

Model with METRON

6769013517237825776429320195069256703244383430126070134031947301923359217851160864356462007956040392

A question from the measured test of November 14, 2026. The digits the model got wrong are dimmed.

Technical notes

Fourteen results, each with its own note

Each note states the bottleneck, shows the measurement and its controls, and says where the result applies.

  1. 1Exact arithmetic in a public modelA public model fails long arithmetic. With METRON attached, and its weights untouched, it is exact up to 100 digits.Exact arithmetic50 exactSealed#118
  2. 2Proved correct for any bit widthAddition, subtraction, comparison and multiplication are proved correct in Lean 4 for every bit width.ProofAny wSealed#117
  3. 3From 32 to 4,096 bits without relearningLaws learned at 32 bits answer exactly up to 4,096 bits, with nothing relearned.Exact arithmetic56,256 exactSealed#116
  4. 4From a neural network to exact lawsFrom the behaviour of a neural network, METRON extracts the exact laws behind it and proves each one.Learning8 exactSealed#119
  5. 5One million requests without an errorOne million requests answered with no error, at a median of 17.6 µs each, on one CPU core.Performance0 errorsSealed#114
  6. 6900 new laws learned at 100 %Nine hundred laws the system had never seen, each learned exactly, in about half a millisecond.Learning900 exactSealed#113
  7. 7ChaCha20 and SipHash, learned exactlyThe building blocks of two standard algorithms, learned exactly from examples and matched against the official test vectors.Exact arithmetic128 exactMeasured§170
  8. 832-bit multiplication, learned exactlyA 32-bit multiplier assembled from 32 learned steps, exact on every case tried, including an exhaustive one.Exact arithmetic65,536 exactMeasured§178
  9. 9Comparison, shifts and selectionComparison, shifts, rotation and selection: the logic of a processor, learned from examples and exact.Exact arithmetic8,192 exactMeasured§172
  10. 10Learned laws, compiled without losing a bitWhat METRON learns compiles to plain bitwise code that a chip can run, with every output bit unchanged.Performance0 bits lostMeasured§174
  11. 11A hidden function among 128 inputsNonlinear functions hidden among up to 128 inputs, recovered exactly, with noisy labels and no wrong law accepted.Learning510 exactMeasured§191
  12. 12Exact laws from noisy dataWith one training answer in twenty deliberately wrong, every law is still learned exactly.Learning162 exactMeasured§185
  13. 13One CPU core against a GPUOne CPU core serves many times more requests per second than a neural network on a high-end GPU, with exact answers.Performance27×Measured§195
  14. 14224,000 times less data than a neural networkThe same laws learned from five orders of magnitude less data than a neural network needs, and learned exactly.Learning224,000×Measured§153

Result

A public model, with and without METRON

Qwen3-1.7B-Base, weights untouched. Same model, same questions. The only difference is METRON.

Model alone

Addition and subtraction, from 50 digits0%
Multiplication, from 6 digits0%

Model with METRON

Every operation, up to 100 decimal digits50/50
Hexadecimal, up to 128 digits (512 bits). Multiplication at 32 hex digits: 49 / 5050/50
Sealed · #118Qwen3-1.7B-Base+  −  ×  <

Read the technical note

Pre-computed on November 14, 2026. Wrong digits are underlined.

1M

requests answered with zero errors.

17.6µs

median time to answer a request.

900

new laws learned, every one exactly.

Proof

Proved correct for any bit width

Addition, subtraction, comparison and multiplication modulo 2w are proved correct in Lean 4 for every width w. The proof has no sorry and uses only the standard axioms.

Sealed · #117Lean 4No sorry

Read the technical note

56,256

exact answers in a row, from 64 to 4,096 bits, with laws learned at 32 bits.

From network to code

From a neural network to exact laws

Given a neural network of 871k parameters that is right 99.575 % of the time, the engine extracts the 8 exact laws behind it. Each law is proved by exhaustive enumeration.

Sealed · #1198 / 8 lawsExhaustive proof

Read the technical note

A hall of 960 data-center racks; in front of it, a single rack with blue lights.

Energy

Less energy, measured

The same task, on the same machine, against a neural network.

≈960×

less energy to answer a request than a neural network: one rack against 960.

≈4×106

times less energy to learn a law than the neural network.

See the comparison

The rack counts draw measured ratios; they are not an inventory of machines.

The AKME S1 module: a black package with a machined frame, a photonic die with a waveguide of blue light, gold bond wires and a lid engraved AKME S1.
Concept visualization

Photonic hardware

AIDON AKME S1

Our own chip, in development

AKME S1, the Akme Shadow One, is an integrated optical memory. Data is held as pulses of light in a silicon-carbide ring and read by an electronic chip bonded on top of it.

We design the photonics, the electronics and the models side by side, and we verify the chip before any silicon is made. The image is a concept visualization, not the layout of the chip.

16

read channels on the CMOS chip, built on an open 130 nm process.

50MHz

digital block, closed across all 18 process corners with no violations.

3.3GHz

optical receiver bandwidth, at 0.94 mW per channel.

3.96M

reads through the full chain with zero errors.

Pre-silicon simulation and verification, before and after layout; full chain with an ideal converter. Not yet fabricated.

Explore AKME S1

Design

Light as memory

A ring of silicon carbide keeps pulses of light circulating. Each pulse is a stored bit. A 16-channel CMOS chip sits on top and reads them.

Closed in simulation

Verified before silicon

The photonic layout is frozen. The digital block closes timing at 50 MHz in every process corner, the receiver reaches 3.30 GHz, and the full read chain ran 3.96 million reads without an error.

Next

Towards first silicon

The fabrication package and the characterisation of the first chips. Until then, every statement here is a result of design and simulation.

Portfolio

What the results add up to

Fourteen measured results describe one technology: systems that learn exact laws, prove them, and run them at a fraction of the cost.

Bottlenecks solved

  • Language models that predict numbers instead of computing them.
  • Correctness that could only be sampled, now proved for every width.
  • Failure on inputs longer than the training data.
  • Trained networks that nobody can read or certify.
  • Inference that needs a GPU for an answer that is only probably right.
  • A new round of training for every new task.

Frontiers opened

  • Exact co-processors placed inside a model’s own computation.
  • Formal proof carried from the mathematical model down to the circuit.
  • Exact laws extracted from public, pre-trained networks.
  • The same architecture on dedicated chips, designed side by side with the models.

Where it fits

  • AI agents for finance, science and engineering, where a figure has to be right.
  • Chip design and verification.
  • Cryptography and wide-integer infrastructure.
  • Real-time, embedded and safety-critical systems.
  • Certification and audit of machine-learning components.

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.