Superseded — 28 August 2026. This note recorded where the project stood on 3 July 2026 and is kept as a dated record rather than updated. Everything quantitative and every version number in it has since moved. Study-level ARsem is 76.67% and the gap about 72 points, not 78.33% and 74; the composite index is 0.596, not 0.605. The reason for the change, and the quality-control failure that produced it, are in What Counts as a Verdict. The Coding Manual is at v2.1, not v2.0. The paper is at v6, and the three v4 PDFs linked at the foot of this note are no longer the current draft.

Two formulations have also narrowed. The central hypothesis is no longer that language models lack metacognition but that they lack epistemic self-classification — a retreat forced by interpretability results and set out in What Counts as Introspection. And the eleven D1 categories are illustrated below as “historical, statistical-probabilistic, metaphysical, causal, normative”: causal is D4, an attributive dimension, and statistical-probabilistic is not a category in the taxonomy at all.

For the current state of the project, see Project status.

Origins and theoretical framework

This project grew out of my longstanding interest in epistemology and misinformation, which produced the book Le Fake News e il Marketing del Vero (2018). The central hypothesis is that language models generate outputs lacking metacognition — a condition Loru et al. (2025, PNAS) and Quattrociocchi et al. (2025) call epistemia — and that structured external intervention, which I call post-cognition, is required to reconstruct the ontological commitments implicit in those outputs. The guiding principle is: ontology precedes epistemology. The type of claim must be established before any truth evaluation.

The taxonomy

The framework is a seven-dimensional taxonomy (D1–D7). The primary dimension D1 classifies the content type of the claim into eleven categories (historical, statistical-probabilistic, metaphysical, causal, normative, etc.). Dimensions D2–D7 are attributive. A Pre-Step 0 — drawing on Austin (1962) — verifies the illocutionary act before applying the framework: if the claim is not an assertion, validation is suspended. A final Epistemic Responsibility Check assesses the epistemic responsibility of the source. The taxonomy is at version 2.2; the Coding Manual is at version 2.0.

The corpus and empirical validation

The horizontal corpus contains 31 claims demonstrating taxonomic coverage. The vertical validation produced 180 API runs (6 claims × 30 replications each), yielding the principal quantitative finding: a 74-point gap between Lexical Agreement Rate (ARlex = 4.44%) and Semantic Agreement Rate (ARsem = 78.33%); mean pairwise Jaccard similarity J̄ = 0.340; composite index IC = 0.605. This demonstrates that surface-level lexical variance masks high semantic convergence.

Current state of the paper (v4)

Eight sections drafted (Introduction through Conclusion), a 56-entry bibliography, and six appendices (taxonomy, Coding Manual, system prompts, pre-registration, horizontal corpus, dimensional redundancy matrix, OSF materials). EXP-1 (multi-model control, §6.6) and EXP-2 (uplift experiment, §6.7) remain to be executed, along with the OSF deposit (Appendix G).

Submission target

arXiv (cs.AI / cs.CL) as an immediate preprint; Synthese and Episteme as journal targets at 6–12 months. arXiv endorsement is currently being sought.

Attachments

The current working draft (v4), split into three PDFs:

WR·FM

Code assigned retrospectively on 23 August 2026, when the notation was introduced: a declared reconstruction, not a real-time record.

The formal apparatus of this note was built in dialogue with a language model and verified line by line. The verification is the part that counts: no step was accepted on the strength of the output’s apparent authority.