docs/VALIDATION.md

FeatureLens final validation

Run the complete local software gate before publishing:

python3 -m pytest -q
python3 -m compileall -q app.py featurelens experiments scripts
python3 -m ruff check app.py featurelens experiments tests scripts
python3 scripts/ui_smoke.py
python3 scripts/release_check.py
python3 -m scripts.validate_artifacts

Public study checks

The committed artifacts/ bundle must:

1. contain both causal_results_final_token.csv and causal_results_max_active.csv;
2. use max_feature_activation as the primary causal policy in study_summary.json;
3. document causal-task-level paired inference;
4. report 224 discovery prompts and 28 causal tasks;
5. include the six report figures required by scripts.validate_artifacts;
6. contain no activation matrices, model weights, SAE checkpoints, or completion markers.

HF Space acceptance

No new model inference needs to be rerun for the final publication if the software checks pass. Verify visually that:

• the Study tab loads measured results rather than the empty-state message;
• the measured headline and causal-position comparison are visible;
• the Workbench and other previously validated live paths still render;
• the interface remains version-neutral and follows DESIGN.md.

Reproducibility

For a fresh study, use notebooks/FeatureLens_Offline_Study_Colab.ipynb or:

python -m experiments.run_all --resume

The causal-addendum notebook is retained only as a migration/reproduction utility for an already-completed final-token baseline.