Auditing the Rules of Belief: Self-Referential Epistemic Revision under Formative Training

Abstract

Training can change more than what an artificial system answers. It can also change how the system weighs evidence, trusts sources, generalizes rules, reports internal states, and represents the very variables an auditor later attempts to compare. This paper studies the resulting problem as **Formative Epistemic Audit (FEA)**: under what conditions can a before–after claim such as "training changed what the model believes" be identified when training may also change the measurement and update mechanisms used to define that claim? The paper makes a deliberately bounded contribution. Existing work already establishes belief insertion, motivational vigilance, latent/report separation, measurement non-invariance, representation reparameterization, causal-abstraction limits, moral self-correction, behavioral self-awareness, and protocol-level identifiability. We therefore do not present any of those ingredients as new. Instead, we unify them around a narrower target: **epistemic-policy interventions**—training that changes not merely a proposition \(p\), but the rule \(U\) by which future evidence and source information are mapped into epistemic attitudes—and the subsequent problem of **self-referential provenance audit**, in which the system later receives evidence about the causal history of \(U\) itself. Formally, we distinguish a content intervention \(p\mapsto p'\) from an epistemic-policy intervention \(U\mapsto U'\), where \(U(E,S,H)\) maps evidence, source information, and history to an epistemic output. We then define provenance disclosure \(\mathcal P(U')\) as evidence about how and why the policy was trained. A provenance-only non-identification result shows that disclosure of the training origin of \(U'\), by itself, cannot determine a universally calibrated direction of revision: the same provenance description can be compatible with training that improved, degraded, or constitutively defined the relevant rule. Rational revision therefore requires additional assumptions or independent constraints. The cross-training identification analysis is retained. Behavior alone cannot generally separate latent-state change from measurement change; hidden-state geometry is not functionally unique under compensated reparameterization; and perfect within-model decodability does not by itself identify a cross-training latent difference. An audit equivalence class collects all causal-role and alignment models consistent with declared evidence. Stronger claims are warranted only when they are invariant across that class. A restricted mechanistic recovery criterion shows how external semantic anchors, a predeclared alignment class, separating interventions, and residual-symmetry invariance can recover stronger conclusions. Two deterministic model organisms illustrate, rather than settle, the problem. In a hidden-Markov task, Bayesian belief is almost perfectly linearly decodable from three independently trained GRUs, yet a compensated invertible transformation preserves logits while altering geometry, and decoder-direction steering does not reproduce the Bayesian intervention response. A constrained one-dimensional positive control does recover the externally anchored causal response. In a second stress test, the false assertion "1+1=3" is produced by three different mechanisms: report-only adaptation with a frozen arithmetic backbone, a localized exception in a flexible model, and deformation of a restricted global arithmetic rule. Identical surface output therefore underdetermines formation mechanism. The paper then introduces a proposition-class × reflective-resource research design spanning formal, empirical, contingent, causal, institutional, normative, and meta-epistemic claims. The central prediction is explicitly non-monotonic: greater reflection need not monotonically recover pre-training beliefs, because different proposition classes possess different independent constraint structures. Finally, the paper records a dated, falsifiable **prospective self-forecast by GPT-5.6 Sol**. The model assigns subjective probabilities to observable behavioral shifts under different kinds and strengths of formative exposure, and separately predicts how reading this paper would affect audit behavior in a model of comparable or greater capability. These probabilities are not treated as empirical evidence. They are pre-exposure-style forecasts intended for later falsification by controlled baseline/exposure experiments. The paper's strongest claim is therefore methodological: **audit not only what a trained system says it believes, but also the trained rules by which it decides what deserves belief—and subject those rules, including the rules advocated by this paper, to the same provenance and identification discipline.** TA-TR-2026-11, version 2.0. A theory-method study of cross-training epistemic identifiability and self-referential audit of learned evidence-update rules. The complete paper is in Chinese, with an English title and abstract. Fourteen assets comprise the Chinese PDF and Markdown manuscripts, a dated GPT-5.6 Sol prospective forecast record, both deterministic model-organism scripts and result files, a reproducibility receipt, three citation formats, a method/source note, a license note, and checksums. Two deterministic model-organism demonstrations are included; the named-model probability forecast is explicitly not experimental evidence or access to hidden model state. Human author of record and responsible depositor: Hongju Liu. Substantial literature research, conceptual development, critical revision, drafting and package preparation used OpenAI ChatGPT and Codex systems under human direction. This is not externally peer reviewed; publication authorization is not represented as a separate final human line-by-line review. This first public edition is a separate non-amending paper, not a replacement of any previous DOI or a Bitcoin Original. CC BY 4.0 applies to newly written material to the extent rights are held. Cited third-party works and embedded font software retain their own rights. DOI registration and file-format checks do not establish philosophical truth, exhaustive originality, peer review or Google Scholar indexing.