The Coherence Engine: Why Morals Cool Under Reflection, and What It Would Take to Call a Machine Human-Like

Bernard Williams observed that reflection can destroy moral knowledge: a person can subject a firm conviction to honest scrutiny and come out holding something thinner and colder without ever having been shown an error. This book treats that observation as a clue about cognitive architecture and builds a research program around it. The proposal is that a lived moral orientation is carried by a typed, weighted graph in which connections compete for a bounded activation budget. Pressure-testing a framework for rigor admits qualifiers into that competition, so felt force dilutes without anything being deleted, while direct perceptual routes explain why force returns in the particular case and practiced compilation explains how qualified structures can become second nature. The central structural notion is coherence failure: a datum that cannot be connected within a representation's current degrees of freedom without substantial reorganization. Five documents develop the idea in sequence. Paper I states the human-side model in full, with its moral argument and 75 labeled equations. Paper II turns the model into a falsifiable cross-domain research program with preregistered tests, rivals, and stopping rules. Paper III closes ten deliberately exogenous capacities into a bounded agent architecture with an exact nesting proof. Paper IV defines an Integrated Architecture Profile and Finite Human-Like Functional Sufficiency, a five-tier evaluation discipline that prices the words integrated and human-like. Paper V, a separately versioned extension profile, supplies bounded candidate contracts for social, relational, and developmental capacities, together with a candidate reference witness. The program claims no empirical validation. What exists is a defined, internally disciplined, falsifiable candidate, and the book states exactly which tests would show it wrong.