Skip to the result

ResearchTechnical note 1

Exact arithmetic inside a public language model

A public language model answers long arithmetic by predicting digits, and gets it wrong. Connected to METRON, AIDON’s exact co-processor, the same model, with its weights untouched, returns the exact answer.

Measured
14 November 2026
Model
Qwen3-1.7B-Base
Operations
+  −  ×  <
Record
118

At a glance

50

exact answers in every operation, up to 100 decimal digits, with METRON.

0%

for the model alone in addition and subtraction from 50 digits, and in multiplication from 6.

0

weights changed. The public model is used exactly as published.

In plain words

Ask a language model to multiply two long numbers and it writes something that looks like a number. It is usually wrong. We gave the same model a small exact calculator of our own, called METRON. With it, every answer in the test was right.

For engineers

Qwen3-1.7B-Base, weights frozen, in two arms: alone, and with METRON attached by declared routing. Four operations, nine operand lengths from 1 to 100 decimal digits, 50 questions per cell, checked against an independent reference.

Why it matters

Exactness becomes a component that can be added to a model that already exists. No retraining, no larger model, no new data: the part that must not be wrong is handed to hardware-grade logic.

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

Language models predict digits. They do not compute.

A language model produces a number the way it produces a sentence: one likely token after another. For short numbers that is often enough. As the numbers grow, the chance that every digit is right falls to zero.

In this test the public model alone answered 33 of 50 additions correctly at 20 digits and none at 50. In multiplication it was at zero from 6 digits.

Result

Same model, same questions. One difference.

  • With METRON
  • Model alone
+Addition
025501234692050100
−Subtraction
025501234692050100
×Multiplication
025501234692050100
<Comparison
025501234692050100With METRONModel alone

Decimal digits per operand

Fig. 1 Exact answers out of 50 questions, by number of decimal digits per operand. The model and the questions are the same in both lines; the only difference is METRON.

Table 1 Exact answers out of 50, decimal operands.
DigitsAdditionSubtractionMultiplicationComparison
AloneWithAloneWithAloneWithAloneWith
15050505050504950
25050505049504750
3505050506504450
4505050501504450
6475050500503750
9465048500503050
20335022500502950
500500500503250
1000500500502550

In hexadecimal, up to 128 digits (512 bits), the model with METRON answered 50 of 50 in every operation, except multiplication at 32 digits: 49 of 50.

Method

How METRON is connected

  1. Language modelreads the question and writes the expression
  2. METRONcomputes it bit by bit with frozen laws
  3. Answercarries the exact result
  1. The model reads the question and writes the expression to be computed, as it would write any text.

  2. The expression goes to METRON, which computes it bit by bit with laws learned once at 32 bits and then frozen.

  3. The exact result returns to the answer. The model’s weights are never changed.

Controls

What was checked

Independent reference

METRON’s answer matched an independent reference computation in every cell of the test.

Scrambled laws

With its laws shuffled, METRON scored 0 in every cell from 2 digits up. The result comes from the laws, not from the wiring.

False triggers

0 in 2,000 lines of ordinary text. METRON only acts when there is something to compute.

Untouched weights

The model file had the same SHA-256 before and after the run.

Open questions

What this opens

  • Placing METRON inside the model’s own computation, so that no routing step is needed.
  • Extending the same exactness to other operations a language model cannot be trusted with.

Fields of application

Where it fits

  • Financial and accounting agents, where a figure has to be right, not likely.
  • Scientific and engineering assistants that work with long numbers.
  • Embedded and safety-critical systems, where an answer cannot be approximate.

Record

How this result is kept

Record
118
Run
14 November 2026, 1 h 50 min
Sample
50 questions per cell; 4 operations; 9 lengths from 1 to 100 decimal digits
Measurement file
SHA-256 c3f278f9fe4b4461f114d4e80dd7e2c58576e1c920d5c7a146656cf5f3eb51dc

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.