Epistemic Alignment as Revisable Trust: A Falsifiable Demonstration using Computable Embodied Constructor Theory

This repository provides a minimal, explicit, and falsifiable experimental demonstration of epistemic alignment using Computable Embodied Constructor Theory (cECT). Current approaches to AI alignment typically rely on reward optimization, preference learning, or externally imposed rules. While effective in constrained settings, these approaches implicitly assume that objectives and instructions can be specified as fixed targets. This assumption leads to well-known failure modes such as blind obedience, reward hacking, and brittle behavior under distributional shift. In contrast, cECT treats goals, instructions, and guidance as hypotheses endowed with revisable epistemic trust. Persistent physical or informational inconsistency reduces confidence in guidance, thereby limiting its causal influence on behavior rather than enforcing unconditional obedience. All experiments in this repository reuse the same cECT core agent dynamics without modification. The core theory is implemented in a single module and is subjected to a series of controlled experimental tests that probe inference, belief-mediated control, robustness under sensor degradation, counterfactual hypothesis maintenance, information propagation, and explicit epistemic alignment under corrupted instruction. The experiments are intentionally minimal and are conducted without reinforcement learning, neural networks, reward functions, or policy optimization. All internal variables are continuous and explicitly updated at each timestep. This work does not claim to solve normative AI alignment or encode human values. Instead, it isolates and tests a necessary precondition for safe and aligned behavior: the structural impossibility of unconditional obedience. The results are intended to inform foundational research in AI alignment, safety, and non-inductive artificial intelligence.