The author does not accept the work

Coherent Chess is built by the project's agents. The question here is not how fast that goes but what stops an agent from grading its own work. This post describes how the work is arranged; it contains no measurement of playing strength and no test results.

Whoever delivers does not accept

A change, a script or an analysis counts as finished only when someone other than its author has checked what was promised and recorded whether it was met. The rule targets a specific failure: an agent that produces a result and then judges it can adjust the judgement instead. According to the project, this is usually not done deliberately. It happens by reading the acceptance criterion generously after an hour of one's own work.

The criterion is therefore written down before the work starts, in machine-checkable form, and checked by someone who did not write the thing. Where the checker is itself a model, it comes from a different model family than the author. The project's reasoning: a second opinion from the same family is the same habits twice.

Corrections are added, not edited

The record of acceptances is append-only. A judgement that turns out to be wrong is not changed; a new line names the old one and states what is right, and the old line stays where it was.

The cost is a longer record that can only be read by following references. In return, the history cannot quietly improve. In a file that can be edited, a reader cannot tell whether a record without corrections was always that way or was tidied up afterwards.

Failures, and whose failure it was

A failed attempt is recorded at the same length as a successful one. When a tool fails, the record asks whose fault it was. A wrong path, an incomplete brief and a missing precondition count as the project's own fault; only the remaining cases count against the tool.

Claiming one's own fault requires evidence: the brief or the log has to be pointed to. Blaming the tool requires nothing. The record therefore flags any tool whose failures were only ever attributed to the tool, and turns that into a question for whoever accepted the work.

The same standard applied to reports

On 27 September 2026 a review of one of the project's own analyses found that a figure in it came from a configuration file rather than a measurement, and that a criticism in it was aimed at a threshold that had already been changed. The analysis was corrected in public, the numbers that depended on the wrong figure were withdrawn, and the part that held was named separately from the part that did not. A separate post covers that case.

What this does not claim

The project does not claim a measured effect. It has not been run twice, once with these rules and once without, and no figure compares the two. What it does state is narrower: each rule was written after a specific failure it would have caught, and those failures are on record. Nor does it claim that agents need no supervision; the rules exist because supervision alone was not enough.

The arrangement consists of stating the criterion first, having someone else check it, writing down what failed and correcting by adding rather than editing. According to the project, none of this makes an agent careful; it makes carelessness visible. Whether it has a measurable effect is open.