Open K.A.T.E. · Workflow

K.A.T.E. Brain Lab

Our own recurrent language experiments, including gated memory, trained locally from random weights. No external AI service, pretrained language weights or third-party machine-learning library.

Training result

Loading…

The answer-focused pilot starts from our own gated reading checkpoint. It compares full-text and answer-only training on 50 original dialogues, using 8 validation and 12 newly phrased development-test prompts. The older reading pilot used 45 public-domain stories. The validation set selects the checkpoint; the test set is scored after selection. This tiny development corpus does not establish general intelligence.

Try the experimental model

Generated replies are currently mostly garbled. This experiment cannot execute tools or control accounts. The regular app uses verified command handlers and, when configured, the separately approved local Qwen conversation model.

Sign in to K.A.T.E. in this browser before generating. Prompts are sent only to this computer’s local service and are not saved by Brain Lab.

Research and data sources

We used these as references for original code and training-data review.

Old fiction supplies language patterns, not modern facts or assistant instructions. Development test results are not proof of general intelligence. The existing dialogue holdout has been used before.

General-purpose capability and language matrix

Knowledge library and storage plan