Code & data

Contemporary AI is trained on observational data — behaviour recorded under conditions where the subject was, or assumed they were, watched. Nine decades of behavioural science establish that observation distorts behaviour. This paper proposes a complementary class of training data: declared intent paired with verified outcome, gathered where the declarant cannot be identified.

The repository is a Python package: the frozen dataclasses for the intent pair, four recipes (declare, classify, match, verify), and a JSONL file of example records.

Those example records are synthetic illustrations authored to demonstrate the schema. They are not real declared data — by design, real declared-intent data does not exist yet and must be produced.

This work is offered in the spirit of open science. Feedback, critique of the schema, and discussion of the production network are all welcome — the repository takes issues and pull requests.