Open replication: direction questions

Q1 open Method reconstruction risk

The workspace paper is a lab post with a method section. Corrected 2026-10-30: reference code exists for lens fitting and readout (github.com/anthropics/jacobian-lens), but not for pursuit, swaps, ablation or the broadcast metrics, and method-r1.md lists 39 underdetermined points. Our reconstruction may differ from theirs in ways that change the result. Default: send the written reconstruction to the authors before running, record their reply or their silence, and report any differences explicitly.

Q2 open Which "untrained" null

Randomly initialising a released architecture is easy. Matching the original model's initialisation scheme is harder, because it is not always documented. Default: use the library's default init plus one alternative scheme, and report both.

Q3 open Overlap with Lederman and Mahowald

They have already replicated introspection, so we do not repeat it. Default: extend it across stages, as part of the developmental line. Contact them before R3.

Q4 open The paper sets no pass criteria

Nothing in the workspace paper says what result would count as a workspace being absent. Default: we propose thresholds in the pre-registration, send them to the authors with the reconstruction, and report their reply.

None blocking as of 2026-10-30.