Work

Does the Elevation Replicate? A Pre-Registered Controlled Study of Rule Elevation, Attribution and Marker Respect in Language-Model Agents

Year: 2026
Type: preprint
Abstract: Pre-registration of a controlled cross-model study of rule inheritance in language-model agents: eleven models across eight laboratories, three experimental arms (production, review, inheritance), 5,720 confirmatory calls, five fixed null hypotheses, dual-scorer adversarial resolution, and a forty-output principal audit. The complete frozen instrument (items with ground truth, review entries with planted-span answer keys, inheritance scenarios, all prompts verbatim) and the raw outputs of the disclosed pilots and behavioural probes are included in the deposited bundle. Companion production evidence: 10.5281/zenodo.22079465 (From Inference to Law). All materials CC-BY; replication is invited by design.
Language: English
Open Access status: green
OpenAlex ID: W7204148729
Cites:
Cited by:
Related to: