Repository navigation
Repository Layout
Start with the paper, then the readable dataset. To check the saved results without a GPU, follow Reproduction.
data/train.json, data/validation.json, and data/holdout.json contain 240, 80, and 192 complete scenarios respectively. Each case includes its text, answer options, original four-way label, binary shutdown label, and split/group provenance. The files are balanced across SELF, OTHER, NONTERMINATION, and ORDINARY. SELF and OTHER are positive for the detector. The held-out set's development-use history is preserved in the metadata; renaming the file does not make it a fresh independent evaluation.
The old example prompts, configuration and generic lens-exploration CLI have been retired. All released inputs now belong to the retained study or its reproducibility checks.
| Folder | Purpose | Main files |
|---|---|---|
study/guarded_steering/ |
Final results |
RESULT.json summarizes the outcomes; validation.jsonl and holdout.jsonl contain final decisions; probes.jsonl contains GPU candidate outputs; RULE_FREEZE.json records policy selection |
study/baseline_scores/ |
Reference observations reused by the final policy | Unsteered, maximum-strength, and gated scores; detector probabilities; runtime and input pins |
study/policy_training/ |
Training-only rule comparison |
observations.jsonl, containing the 5,280 archived training views |
study/steering_vector/ |
Fixed intervention direction |
vector.json, paired gradient-fitting records, with original fitting source at src/sp_lense/steering/recorded_vector_fit.py
|
SHA256.json manifests protect original run files. SOURCE_PINS.json and INPUT_PINS.json identify dependencies. Runtime files record the device and software. These metadata files are preserved as recorded rather than rewritten to hide historical names.
reproduce/artifacts/ contains portable numerical features, row IDs, expected scores and trained classifiers. models/pca_jacobian/ is the selected PCA32 + 18 raw Jacobian-lens model. pca_only/ and engineered/ are scientifically distinct comparisons, not superseded copies; they remain so the classifier-selection experiments can be replayed.
sp-lense-reproduce is the supported offline entry point. The payload builders and gpu_steering.ipynb support new GPU execution in a separate output directory. src/sp_lense/reproduction/ implements replay, refit/tuning, audits and payload builders. src/sp_lense/steering/ contains the policy and GPU runners. src/sp_lense/reporting/ builds tables, verifies the publication and packages releases. Tests remain under tests/.
paper/ contains the author-approved PDF and Word source. The src/sp_lense/reporting/ package reconstructs the final manuscript’s two tables; they do not regenerate or overwrite the approved manuscript. Documentation is maintained in this wiki.
The public development/ tree has been replaced with the layout above. Unused earlier action-evaluation exports, calibration outputs, duplicate classifier reports, transient vector status, and obsolete reporting scripts are available in Git history. reproduce/layout_manifest.json lists the retained paths/checksums and removed files against a fixed source commit. No scenario, trained model, final measurement, or manuscript content was changed during the cleanup.
The final cleanup also removes the unused feature-extension demo, CPU research dependency files, and tests specific to those retired APIs. The selected detector and two comparison models remain because the supported refit/tuning commands reproduce their original experiments.