Epistemic Fusion or Epistemic Tunnel: A Phenomenology-Anchored, Global-Literature-Constrained, Problem-First Architecture of Human-AI Collaboration (Epistemic Note v8.1 — strong-claim / legitimate-defeater edition)

Abstract

Series. Standalone epistemic note of the Human–AI Readout programme; member of the Zenodo library When AI Expands Human Potential (hub DOI 10.5281/zenodo.22308201). Builds on the Bounded Knower sequence I–IV, Experience Is the Human LoRA, The Language Bridge, Operational Linguistic Wisdom, The Dialogue as the Ground of Enlightenment, and the Readout Genesis / Readout Condition base; methodology: glosa (DOI 10.5281/zenodo.22310837). Version 8.1 (2026-09-05) supersedes v7.1 (Genesis-aligned) in this record; the earlier PDF stays available under its own version DOI. Status. K0+ conceptual architecture with cross-domain empirical, meta-analytic, and theoretical constraints. Strong claims are retained by design. This edition separates phenomenological explananda, constitutive necessities, structural proposals, diagnostics, and empirical predictions; each is paired with a legitimate defeater. Calling a proposition "too strong" is not itself a scientific defeater. The burden for empirical causal claims remains with the author, while objections to constitutive or phenomenological claims must target the relevant derivation, counterexample, phenomenon, scope, or rival explanation. Core move. The central research problem is not simply whether AI improves an answer while it is present. The stronger problem is whether Human-AI collaboration can enlarge the human's corrigible space of inquiry, preserve independent resistance, avoid self-reinforcing semantic tunnels, and return durable epistemic capability to the human after the AI is removed. The evidential strategy is cross-domain convergence rather than dependence on one causal pathway: clinical workflow, explanation and reliance, creativity and homogenization, verification cost, self-confidence, Human-AI feedback, learning, sycophancy, warmth, and task ecology instantiate structurally homologous failures of collapse. No single study is claimed to reproduce the whole architecture. Their convergence constrains what a general architecture must keep distinct. Attached note (2026-09-05): Uplift Note — 25 proposals from the Epistemic Fusion papers to glosa: what was wrong, what was fixed, what stands. Record of the glosa uplift meeting that read this paper: 25 proposals across five lenses; the first round kept 4 because proposers cited fabricated claim ids; under the founder ruling "raise the work to the claim" every proposal was re-anchored to file:line citations, paired with its legitimate defeater and a failing control, and re-judged (13 build-now / 8 build-with-revision / 4 still-open). Independently checked before attachment (three off-by-N line citations corrected; r2 fixes the HU-5 line reference). Tier Dr; human owner Yaoharee Lahtee; AI drafted as assistant. Methodology: glosa v0.3.0 (DOI 10.5281/zenodo.22310837).