Sources and provenance

Sources checked 2026-09-07. Literature supplies methods and comparison points, not validation of this local experiment. The following descriptions are intentionally brief; the report's numerical findings come from local saved results.

• J-space paper: defines an averaged Jacobian transport followed by normalization/unembedding, and a pseudoinverse-based swap in a nonorthogonal span. Readouts are suggestive; interventions test effects. Our experiment does not test the paper's broader global-workspace account.
• Reference implementation: frozen in vendor/jacobian-lens, commit 581d398613e5602a5af361e1c34d3a92ea82ba8e. Read hf.py, hooks.py, lens.py, fitting.py, tests, README and the walkthrough. It records zero-based block outputs, transports row vectors with h @ J.T, and applies the original final norm and unembedding. README explicitly says reference code is unmaintained. Apache-2.0.
• Walkthrough: Qwen3.5-4B, bf16 CUDA, qwen-n1000, boot/currency example. This example is used only as a local smoke check, never evaluation evidence.
• Neuronpedia lens artifacts: Qwen3.5-4B n1000 file exists, 406,332,644 bytes; model shards total 9,319,828,096 bytes. Download not selected given free disk. Qwen3-1.7B lens is 226,501,315 bytes with 466 fitted prompts (despite the repository revision name). Fit configuration and model history are saved locally.
• Qwen3-1.7B checkpoint: full original safetensors and tokenizer, no quantization/conversion. Model is 28 blocks, residual width 2048, tied embedding/unembedding, learned RMSNorm gain. Apache-2.0. Artifact SHA256s and sizes are in provenance/artifacts.json.
• Neuronpedia fitting wrapper: downloaded at this commit. Delegates model wrapping, fitting and saving to jlens. The fit YAML does not record a model commit or historic jlens commit. Model repository history establishes that the latest checkpoint/tokenizer predates the 2026-06-11 fit; it does not recover the exact historic fitting environment. Treat that missing historic code provenance as a limitation.
• CHIVE, 2026-08-21: counterfactual prompt edits provide measured labels; AO/NLA/SAE readout tools did not improve prediction over transcript access in the tested settings. It did not evaluate this local J-lens intervention experiment. Motivates separating readout visibility from useful counterfactual prediction.
• Circuit tracing methods: cross-layer transcoders create a replacement model for attribution graphs; causal checks remain necessary because the replacement is approximate. We fit no transcoder and recover no attribution graph.
• Circuit tracing case studies: Dallas/Texas/Austin and California/Sacramento illustrate intervention on intermediate features, alongside a Dallas-to-Austin shortcut. This motivates our two-step task; none of those cases enter local evaluation.
• Introspection experiments: controlled concept injection can sometimes be detected/reported under specific conditions. It does not validate ordinary post-answer explanations as records of the preceding computation.
• Apple PyTorch/Metal: MPS uses the Mac GPU. The page currently lists PyTorch 2.11.0 in its installation example; the package index installed 2.14.0 here. We record the actually installed version and execute checks instead of treating the example as a version pin.
• MLX documentation, lazy evaluation, eval, MLX-LM: GPU arrays/lazy computation require explicit evaluation and synchronization for timing. Installed MLX-LM Qwen3 source was inspected for post-block residual access and tied unembedding.
• Core ML prediction and typed execution: converted prediction graphs may be partitioned across CPU/GPU/Neural Engine and precision depends on conversion/runtime. Inference support is insufficient evidence of access to the required intervention boundary. We did not convert a model or execute on the Neural Engine.
• SIX ISO 4217 list: downloaded to provenance/iso4217-list-one.xml to cross-check currency codes. Cities are deliberately constructed familiar factual examples; no claim of random sampling or independently audited geography dataset.

Editorial references were read from the requested absolute paths and remain unchanged; hashes are in provenance/editorial-inputs.json. Generated prose is this investigation's interpretation, not the papers' conclusion about this model.