FidelityEight: Frozen-Protocol Fidelity Measurements of the Distributed Jacobian-Lens Checkpoints

Abstract

The Jacobian lens ("jlens"), released 2026-07-02 as companion code to the *Verbalizable Workspace* paper, ships with pre-fitted checkpoints for ~36–38 model families. To our knowledge — a bounded-absence claim; see §6 for the search scope — no direct lens-versus-model fidelity measurement of these distributed checkpoints (rank agreement, KL divergence, top-1 agreement, or calibration) has been published by any party, including the upstream authors. This note contributes six fidelity readings of three distributed checkpoints under two position-domain bindings, plus two measured floors (chance and vanilla logit-lens), all under a frozen protocol: a rank-based fidelity statistic (wgrd) preregistered before each run, evaluated on a fixed 266-item prose set, with a positive/negative-control calibration battery passed in the same invocation as every measurement, and md5-pinned instruments, inputs, and outputs throughout. The distributed OLMo-3-7B lens reads 1.39 dex (median word-group rank deviation from the model it explains), the OLMo-3-32B lens 1.57 dex, and the Qwen3-8B lens 2.44 dex on its own tokenizer scale; the chance floor sits at 4.47 dex and the vanilla logit-lens baseline at 1.84 dex on the 7B ground. All values are robust to a position-domain correction aligned with the lens's own fit domain (largest shift: 0.10 dex). We report these as instrument characterization only: they license no claim about the models, about the upstream fitting objective, or about any downstream hypothesis. Instrument ineligibility is not refutation.