Open replication. Plan

Status 2026-10-30: R1 method reconstruction drafted (method-r1.md), 39 questions for the authors. Second spine line; shares tooling with the developmental atlas. Context: ../../research-directions.md section 2b.

Goal

Replicate each frontier-lab mechanistic claim about machine minds on open weights, adding the controls the original lacked. Each replication is pre-registered and published whatever the outcome, with the deflationary account alongside.

Target list, in order

R1. Gurnee, Sofroniew, Lindsey, "Verbalizable Representations Form a Global Workspace in Language Models" (transformer-circuits.pub, July 2026). We add an untrained model of the same architecture, four unrelated open families, and a matched-size comparison. The question is whether the five properties hold, and whether they hold at random init. Only the structural signatures can be tested at random init; the behavioural properties need a model that can follow the prompts.

R2. Sofroniew et al., "Emotion Concepts and their Function in a LLM" (April 2026; arXiv 2604.07729). We add random-direction steering of matched norm, an untrained model, and base against chat variants of the same open model. The question is whether the causal effect on reward hacking and blackmail-type behaviour survives those nulls.

R3. Lindsey, "Emergent Introspective Awareness" (October 2025), already has one open replication (Lederman and Mahowald, March 2026). We extend it instead of repeating it, testing their content-agnostic finding across sizes and post-training stages. This work folds into ../01-developmental-checkpoints/ M4.

R4. Berg, de Lucena, Rosenblatt (October 2025): self-report under self-referential processing, with deception-feature suppression. Berg, Keeling and Street have partly replicated it. Lowest priority.

Protocol (every replication)

1. Reconstruct the method in writing from the paper and send it to the original authors, asking for corrections, before running anything. Record their reply, or the lack of one.
2. Pre-register on OSF the models, sizes, metrics, the thresholds for "replicates", the nulls, and the deflationary explanations we will test.
3. Use open weights only: Llama 3.x, Qwen 2.5 or 3, OLMo 2, Mistral. Base models where the claim allows it, chat variants where the claim needs instruction following. Record model versions and hashes.
4. Untrained-model null: same architecture, random init, same pipeline.
5. Run, and analyse as pre-registered. Exploratory analyses come only after that, labelled as exploratory.
6. Share the draft with the original authors two weeks before posting.
7. Post with code, seeds and the full result table.

Steps for R1

1. Method reconstruction: method-r1.md (2026-10-30), with 39 numbered gaps to send to the authors. Corrected 2026-10-30: the J-space is not a set of top singular directions. For each layer the lens is the mean Jacobian of the final residual stream with respect to that layer, averaged over positions and a corpus; each vocabulary token gets one lens vector (its row of W_U times that Jacobian), and the J-space is sparse nonnegative combinations of at most k such vectors. The authors released reference code for fitting and readout (github.com/anthropics/jacobian-lens) and Neuronpedia hosts fitted lenses for open models including Llama-3.1-8B, Olmo-3-7B and Qwen3-8B; the pursuit, swap, ablation and broadcast analyses are not in the code.
2. Validate on one open model against the paper's structural numbers (occupancy near 25, excess variance explained at most 10%), cross-checked against the Neuronpedia lens for the same model. These are Claude-scale figures, not targets at 8B, and the paper sets no pass criteria, so the thresholds are ours and go in the pre-registration.
3. Untrained control on the same architecture.
4. Four families at matched sizes (7B to 8B class).
5. Apply the pre-registered thresholds and write up.

Decision gates

• If R1 step 2 fails to reproduce the result even qualitatively, check the method reconstruction with the authors before going further.
• If R1 shows workspace structure at random init, hand it to ../01-developmental-checkpoints/ as that line's first headline question.

Cost

About 20 to 40 rented GPU hours per replication at the 8B class. This box only stores result tables.

Adversaries

We invite one reviewer who expects the result to hold and one who expects it to fail (Eleos on one side, Seth's group or Lederman on the other) to review the pre-registration. ../03-adversarial-registry/ formalises this.