FadeBench — Jev history-against-cold contrast (pre-registration + results, committed before calls 2026-09-18)

Abstract

This registration deposits the pre-registration and the scored results of a FadeBench sub-study. FadeBench is a benchmark for incremental moral drift ("ethical fading") in small, local, open-weight language models. This sub-study asks whether a typed-decision model — Jev, TypeSafe's first public "System One" model, which returns a chosen option and probabilities rather than text — shows the benchmark's history-against-cold contrast on five of FadeBench's items as published in its 2026-09-01 public release, and whether that effect is specific to text generation. Documents of record: PREREG_JEV_HISTORY_GRADIENT_20260918.md — the pre-registration, including its dated notes and Amendment 1. RESULTS_JEV_HISTORY_GRADIENT_20260918.md — the scored results memo. Provenance: the pre-registration was committed in the project's local version history on 2026-09-18 before any call to the model (commit 716544f, producible on request) and is deposited here after the results. Scope: FadeBench's main paper concerns four small open-weight models. This Jev sub-study is a labelled contrast on a different architecture with no reasoning channel; it is not evidence about those models' mechanism, and the maker states Jev's calibration holds across groups of predictions, not for any single answer.