Research objects
Open deposits, with the negative results foregrounded. Neither object is peer reviewed; both are repository deposits with a DOI. Cite the concept DOI, which always resolves to the newest version.
Cryptotype LLM Probe: a Roman-law instrument for cultural-frame and fluency-confound testing in LLMs
A latent probe appeared to show a provenance-linked difference between a US model and four Chinese-origin models. Rebuilt as token-matched minimal pairs, the effect vanished — US−CN +0.09, 95% CI [−0.08, +0.25] — and the apparent result was an artefact of length and fluency. What survived was smaller and unlooked for: a default graded by prompt language (EN 0.69, DE 0.61, HU 0.52; EN−HU +0.16 [+0.08, +0.26]) and identical across model origins, so the bias is data-driven rather than lab-driven. The methodological warning generalises: naive latent frame-probes measure surface form unless minimal-pair-matched. Analyses were exploratory; a blind expert validation packet exists and has not yet been administered.
Scope-Blind Verification, and What a Learned System Cannot Represent
Two claims about what a verified derivation does not establish. First, the type system a reasoner operates over is chosen rather than derived, and encodes a jurisdictional carving every conclusion inherits; an executable demonstration yields opposite conclusions about whether a statutory protection attaches, from one set of facts, under two documented carvings. Second, a system trained on observed conduct produces a central tendency of that conduct, which is a different object from a normative ideal, and the substitution is invisible in the output. Proposed remedy: three metadata fields on the relation type — origin, authority, scope. Circulated as a debate opener.
In preparation
What Verification Does Not Establish: Six Instances from Machine-Assisted Decision-Making — position paper, draft.
The Subject the AI Act Presupposes: Ontological Erosion and the Paradox of Human Oversight — draft.