How do you check what a machine claims, after it has claimed it? A research programme on the post-cognitive validation of language model outputs, built from inside the content industry — where the outputs are already shipping.
Draft v6 · writing & submission Last change to the programme: July 28, 2026
Language models produce assertions without metacognition. The programme builds post-cognition: an external, auditable procedure that first fixes what kind of claim is on the table — historical, statistical, causal, normative, metaphysical — and only then asks whether it is true. Ontology precedes epistemology. The apparatus is a seven-dimensional taxonomy with a coding manual, tested on a horizontal corpus of 31 claims and a vertical run of 180 model replications.
Three claims that can be shown to be wrong
Disagreement between model runs is lexical, not epistemic. Across 180 runs, lexical agreement was 4.44% and semantic agreement 78.33% — a 74-point gap.
Wrong if the two rates converge on a larger and more varied corpus, or if the semantic rate collapses under a stricter annotation protocol.
Classifying the type of a claim before evaluating it changes the verdict, and improves agreement between coders.
Wrong if coders who skip the classification step reach the same verdicts, at the same agreement, as coders who apply it.
A model can report on its own internal states and still not classify the epistemic type of its own claims. That is why the check has to come from outside.
Wrong if a model, with no external scaffolding, assigns claim types to its own outputs at the agreement level of trained human coders.
Working paper, v6 — PDF → Back matter — PDF → Coding manual, v2.1 — PDF →
The changelog of the programme. Dated, in-progress notes — including the ones where the framework had to be corrected.
A new interpretability result shows that language models can report on concepts held in their own internal states. It did not refute the framework. It forced a flat premise — no metacognition — to be replaced by a narrow one: no epistemic self-classification. The narrow one is what the argument actually needed.
Recent notes
What Counts as a Verdict A spot-validation that failed at 77.8% did not turn up an arithmetic error. It turned up a hidden ontological decision inside the extractor — the same category collapse the framework is… Coding Manual v2.1 — The framework audits itself The protocol catches up with the study it governs: six claims codified, AR-lex and AR-sem formally defined, quality control turned into a documentable record. And the record’s most…Reading the Content Economy from Europe — one essay a month
Europe’s extension on high-risk AI systems makes validation no less urgent. The Fable 5 case showed, in forty-eight hours, what happens when an organization doesn’t know
Why the right learned the postmodern lesson better than its teachers — and what comes after
Why the term explains almost nothing, why that matters politically, and why Karp is right about the premise and wrong about everything else