Our commitments
We should be able to explain the standards we ask of AI. Here are our working rules, their status and the subjects still awaiting confirmation.
Status is part of the commitment
As of 22 October 2026. A documented rule is not evidence that it has always been followed. No independent audit of all implementation has been demonstrated for the rules below.
We distinguish a working rule, work in progress and matters awaiting confirmation. This keeps intentions in proportion to practice.
A person sends the message · working rule
AI colleagues may prepare drafts. The rule is that a person reads and sends a message addressed to a recipient.
Automatic notifications sent by individual services still need to be checked service by service. We make no blanket commitment about them here.
Irreversible actions require approval · working rule
The rule is that sending, paying, deleting and publishing require explicit human approval. We define the precise boundary for each process.
This rule describes our intent and method. A complete independent review of every action has not been demonstrated.
Builds get a second look · working rule
A fresh reviewer examines a build before delivery. They have room both to find faults and to report that none were found.
A second look helps detect errors. It does not guarantee an error-free result. The tests actually performed remain part of the evidence.
Record goals, not quotations · working rule
Since 16 October 2026, the rule has been to record goals, decisions, agreements and unresolved questions. We include direct quotations only when explicitly requested.
This helps preserve the purpose without unnecessarily carrying a conversation forward word for word.
Data, training and services used · awaiting confirmation
The general wording of commitments about data use, model training and assignment-specific transparency has not yet been fully confirmed and legally aligned.
We therefore make no blanket promise here that all data stays on our own machines or that every external service has the same terms. Your assignment needs specific agreements.
Share lessons from our own work · in progress
The Lab and the Proceedings already show our own work and lessons. This is not a promise to publish every error.
Client work and other people’s data do not belong there without a separate decision and the required permission. A general publication commitment on this still needs confirmation.