The Theory Machine: A Minimal Operational Epistemology for AI and Science
Mathine: Theory Card Compiler Machine
Link: https://doi.org/10.5281/zenodo.18869899
Many AI and scientific workflows still treat “theory” as a narrative artifact: a coherent explanation gets promoted into policy, products, or papers without explicit falsifiers, replayable evidence, stable measurement stamps, or verification-budget accounting. This paper draws a clean line: coherence is not closure, and authority is not verification.
The core proposal is the Theory Machine (THM), a minimal operational epistemology that compiles hypotheses into Theory Cards before promotion occurs. Each Theory Card is scope-stamped, falsifier-first, receipt-carrying, and probe-ready: it states what the hypothesis claims, where it applies, what it predicts, what would falsify it, and which receipts and low-cost probes are required for replayable evaluation.
That compilation step is the point. It converts narrative explanations into operational objects whose promotability can be gated under a fail-closed discipline—HOLD/YELLOW/GREEN—so promotion happens only when independent replay is feasible under declared budgets, not when an explanation merely sounds consistent or wins a benchmark.
THM is positioned as a subsystem inside an Evolution Machine loop: Solvers generate candidate hypotheses, THM compiles them into Theory Cards, Mathines execute probes and enforce promotion gates, and Learners extract reusable failure motifs and repair templates. The result is a workflow where disagreement becomes localizable (to scope, falsifiers, receipts, and probes) instead of becoming an argument about persuasion.
The contribution is intentionally scoped and defensively framed: the paper demonstrates THM as an executable compiler and replay harness, specifies an empirical validation roadmap, and defines success as reaching GREEN under independent replay across multiple seeds—explicitly avoiding world-level truth claims or domain-performance claims prior to those runs.
Mathine — Evidence In, Trust Out 