Thoughtware

Software has historically been built by translating human intent into explicit procedures. Its behaviour has largely depended on rules, algorithms, and deterministic logic specified in advance by human designers and engineers. Modern foundation models introduce a different capability: software can now perform forms of cognitive work at runtime, including interpretation, reasoning, evaluation, recommendation, and judgment. In this paper, judgment refers to the selection or construction of an appropriate conclusion or action where the correct result cannot be fully determined in advance by a fixed procedure over the available inputs.

This shift has produced a rapidly expanding implementation vocabulary. Terms such as large language model application, retrieval-augmented generation, agent, copilot, and agentic system describe important technologies, interaction patterns, and architectural mechanisms, but they do not adequately name the broader class of software that emerges when cognitive capability becomes a structural part of the system. This paper proposes Thoughtware as that category: cognitive software in which capabilities such as interpretation, reasoning, judgment, and evaluation are deliberately organised as first-class elements of the architecture.

The paper develops a conceptual framework for reasoning about such systems beyond specific models, vendors, and orchestration techniques. It examines how judgment can be bounded and located within software, how cognitive and deterministic responsibilities can coexist, how cognitive behaviour can be composed into larger systems, and how systems whose outputs cannot always be verified deterministically can instead be evaluated, governed, and improved.

The resulting theory provides a vocabulary, architectural model, set of primitives, and engineering principles for software whose behaviour depends materially on machine judgment.