Andy Dixon
67a825a995
Twenty-six pages covering concepts, architecture, every subsystem, an interface tour with screenshots, the full API route table generated from openapi.json, the CLI, security and threat model, operations, ADRs, development history and a glossary. Built into the dixon.cx page skeleton, unlinked and noindex.
2.2 KiB
Predictions give a mind a way to be wrong in a measurable way. A prediction is an explicit expectation with criteria a resolver can check; its outcome feeds back into the beliefs it came from and into a calibration record that says how well confidence has tracked reality.
Making predictions
Predictions are created by the executive when a belief with sufficient confidence implies something checkable about the future, by operators through the interface or noemactl prediction, and by experiments as expectations. Each has a statement, a subject, a confidence, a time window, machine-checkable criteria (an event of a type whose summary contains a term, a belief crossing a confidence, a metric crossing a threshold), the belief it derives from and its origin event.
Resolution
A scheduled resolver evaluates open predictions once a minute. Criteria met inside the window resolve to confirmed; a window that closes without the criteria resolves to expired (treated as refuted for calibration unless the criteria say otherwise); explicit contrary evidence resolves to refuted. Operators can resolve manually with a reason. Every resolution is a prediction_resolved event; a wrong one is also a prediction_error event with the size of the error.
Feedback
A prediction error adds evidence against the belief the prediction derived from, weighted by the size of the error, and lowers confidence through the ordinary belief arithmetic. A confirmed prediction adds support. Errors also move affect (uncertainty up, confidence down) and drives (uncertainty reduction up), and give curiosity a question. The demonstration scenario in the plan, "prediction error lowers confidence and produces a curiosity goal", is exactly this chain and is covered by tests.
Calibration
The calibration page bins resolved predictions by stated confidence and shows the observed frequency in each bin, the Brier score and the counts. A well-calibrated mind's 70 % predictions come true about 70 % of the time. The page exists so that an operator can judge whether a mind's confidence means anything, and so that experiments that change attention or personality can be compared on it.
{{include docs/predictions.md}}