Carl: a source-grounded family agent

An agent working with family information needs to be able to explain where its conclusions came from. Carl is a local agent application built around source-grounded records, reasoning and human decisions, with a financial-planning mission that depends on the completeness of the available facts.

CTO to Human: useful reasoning has visible limits

The public repository identifies incomplete whole-product acceptance and partial financial coverage. That matters more than a convincing demonstration. Missing information should remain missing information; a model should not turn an incomplete record into apparent certainty.

The project distinguishes a proposed plan or draft from an external financial or vendor action. Those actions remain human-led. This hub describes the architecture without reproducing private family records, financial exports or personal conclusions.

Engineer to Engineer: domain ownership and provenance

Carl is a Java21 application on the KofTwentyTwo Agent Foundation, with QQQ administration for its domain records and native workflows. Its documented design keeps services and reasoning in the application while shared infrastructure belongs in the foundation. Data-readiness, evaluation and acceptance documents provide separate places to record what has actually been qualified.

Controlled synthetic fixtures support regression testing; they are not a substitute for a working product or verified real-data integration. The public documentation also distinguishes the locally running package from newer source and separately qualified dependencies. That separation makes acceptance traceable to an actual artifact rather than a repository’s ambition.

The Carl repository records the current scope and limitations. Future journal entries will explain source boundaries, evaluation and acceptance as those pieces are verified, without exposing owner data.