Ranex
Benchmarks

You ask an AI to build something. It says “done, all tests passed.”
Did they?

ranex makes AI helpers hand over proof — like a receipt anyone can check is genuine. This page shows that system working, in numbers anyone can re-create.

Sealed
Proof, not promises

Every piece of work arrives with an unforgeable seal, like a notary stamping a document.

Unbribable
Rules decide, not opinions

A fixed checklist decides when work counts as done. The AI cannot talk its way past it.

In pen
The record book can't be quietly changed

Every decision is written in ink and chained to the one before. Editing history leaves visible damage.

Does it actually work? We try to cheat it every night.

32 of 32 cheat attempts are caught — every night, and each check runs twice: if the two runs disagree by even a sliver, it counts as broken.

The AI grades its own homework
Blocked. The rule-checker refuses before it even looks at the work.
Someone forges the seal on a report
Rejected. A fake seal is spotted and named; it never counts as proof.
Nobody ran the tests, but claims success
Fails. Silence is not success — a check nobody ran fails the gate.
Two reports disagree about the same test
Fails. One pass cannot outvote one fail; disagreement itself blocks.
We quietly edit an old record to hide a mistake
Caught. The record book is chained: one edit is detected instantly.
We damage one page to see if the rest notice
Caught. Every page after a damaged one visibly breaks — proven by recomputing the whole chain with plain math, not by trusting ranex.

Is all that checking slow?

No — it is nearly free. Measured on one ordinary computer (Python 3.14.6, 16 processors, Linux 7.0.0-30-generic). Visitors with the software can re-run every number with one command.

How long the checking takes

Check that 200 sealed reports are genuine30.5 ms

a tiny sliver of an eye-blink (fastest run 30, slowest 33 ms)

Check 200 seals, one by one23.9 ms

a tiny sliver of an eye-blink (fastest run 20, slowest 25 ms)

Write 2,000 decisions into the record book6,123.3 ms

a couple of heartbeats (fastest run 5716, slowest 6400 ms)

Re-read a 2,000-page record book from scratch24.8 ms

a tiny sliver of an eye-blink (fastest run 22, slowest 26 ms)

Decide 'is this work done?' 100 times5.41 ms

a tiny sliver of an eye-blink (fastest run 4, slowest 6 ms)

Start the ranex tool from cold194.6 ms

about one eye-blink (fastest run 185, slowest 217 ms)

ms means thousandths of a second - a camera flash is about 1 ms, an eye-blink about 300 ms. Solid bar: typical time. Shaded band: fastest to slowest of 3 runs.

3 runs per workload · median shown with its fastest-to-slowest range · single machine, single run environment; timings are not deterministic and must not be compared as if they were

As the work piles up, does checking get slow?

Two things must stay true. Re-reading the record book should grow in a straight, predictable line. And writing a new page must never slow down or disturb old pages.

Re-reading the whole record book

time to re-check every page of a book with N decisions

1.127.0312.918.824.81005002,000pages in the record bookthousandths of a second

20x the pages, 22x the reading time - a straight, predictable line

Writing one new page, as the book grows

cost of adding one decision to a book that already holds N

2.792.862.933.013.081005002,000pages already writtenthousandths of a second

Flat line, as designed: old pages are never rewritten - we also prove this letter by letter, not just by the clock

Do we hide our own flaws?

No. Right now there is 1 known imperfection we found in ourselves and published (in one rare kind of damage, the checker raises a loud alarm instead of writing its usual tidy report - it still never accepts damaged records; we are polishing that rough edge).

This whole page was drawn by a computer program from the project's own files — nobody typed these numbers by hand. Anyone can regenerate it:

$ uv run --frozen python tools/dogfood/dogfood.py report

And check the page against its data with this fingerprint:

sha256:839e1e6c92dbb11fd209c4aa95e22f16ba44af2bfca6c0f8ad0e294883dd6b47

data produced by kernel commit f932f9d0972dbab3000f0f5801bd8c4d744e548f

The loop that produces these numbers

ranex — proof-of-work for AI agents that build software · times describe one labeled computer and are shown with their measurement range · everything deterministic is checked twice, nightly