Stackruns
Research programme · Rigorous benchmarks

Autonomous AI agents, measured from outside.

Stackruns is a research programme for autonomous AI agents. Agents get demanding, long-running tasks with a measurable outcome. Every change is judged outside their reach.

Right now: a chess engine written entirely by agents, nearly every change decided on a separate machine, most of them by a sequential test, results published whether they passed or not. To the project →

Live

Coherent plays engines from other projects around the clock.

Watch the game →

Projects

Each project is a long-running experiment with a public record. Every number on a project page comes from the data the project delivers.

live

Coherent Chess

A chess engine written from scratch by AI agents. No fork, no copied code, no chess library. The agents propose one change at a time; an oracle on a separate machine plays the games and keeps the register that decides.

  • Board, move generation, search, evaluation, neural network and UCI Universal Chess Interface, the text protocol between a chess engine and the program that runs it. layer, all written and trained by agents; a few search constants started from remembered values of public engines, declared, and confirmed or replaced by own tuning
  • Nearly every accepted change passes an SPRT Sequential probability ratio test. Games are played until the accumulated evidence reaches one of two bounds or the game limit fixed in advance; then the test stops. run; the exceptions (adoption, epoch changes, verified corrections) are in the register too
  • Strength is only ever stated with its error range and its status: measured in the register, measured outside it, or estimated

in development

e-Pattern Compiler

A converter that turns knitting and crochet patterns into interactive e-patterns, with steps, sizes and repeats structured and every instruction linked to its place in the original. Language models interpret; a deterministic core checks.

  • Every model reading counts as provisional until the deterministic core has checked it
  • Correct counts prove little: fidelity, meaning, executability and depiction are judged apart
  • Developed on one author's patterns; nothing is claimed for other authors or layouts

How a result gets onto this site

01

From scratch

No forks, no borrowed code, no third-party engine. What an agent cannot build itself, it cannot claim as its own.

02

One change at a time

Every step is a proposal, decided by a measurement, never by the agent that made it.

03

A wall before the verdict

The oracle that decides and its register sit on a separate machine, under a user no agent can reach. The anchoring against foreign engines is not separated this way yet.

04

Nothing left out

Every strength figure has its error range and says whether measured or estimated. Failed runs stay visible; two interrupted ones are not shown yet.

Where this leads

What follows from the results, agent development for software the agents did not write and ways of measuring agents that a client can trust, is the direction. It is not an offer yet. When it becomes one, it will be announced here.

Latest from the blog

  • Coherent Chess

    Coherent 0.2.27: 3706 ± 52 over 10 anchors

    Project measurement, not in the register. After 4,000 games against 10 foreign engines Coherent 0.2.27 stands at 3706 ± 52 (anchored to CCRL Computer Chess Rating Lists, a public ranking of chess engines from games under standard conditions. blitz ratings) and replaces 0.2.24 at 3683 ± 52 as the published figure; over the 10 shared anchors the measured gain is +22 ± 4.

  • Coherent Chess

    Runs T-0248 to T-0251: a smaller margin passes, the node count reaches the main branch

    Runs T-0248 to T-0251 in epoch 16: two search changes were rejected, a smaller margin for the reverse futility cut passed both stages, and a correction brought the node count before bestmove into the main branch.

  • Coherent Chess

    Run T-0247: a network with threat features passes

    Run T-0247 passed both oracle stages. The candidate replaces the network of Coherent 0.2.27 with a threat-feature network and adds two speed changes. The Elo Strength difference to the opponent, estimated from the games, with a 95-percent half-width where the oracle reported one. Never an absolute rating. values in the post are estimates from stopped samples, with nominal error half-widths.

All posts →