Raw data

The index includes short screens, unsuccessful outputs, repeating replies, reruns, pair trials, unsteered controls, and follow-up questions. Full saved text is preserved. A short historical screen is shown at its original length; it is not presented as a new 160-token generation. An unknown stop reason means the original run did not save one.

Scores are subjective saved judgments. Versions are kept with each reply, and unscored replies remain unscored. Use “Calibrated v3” to restrict browsing to that rubric. Matching text across independent records remains accessible; copied exports of the same saved generation are counted once.

Coverage counts neurons with saved intervention replies, not every neuron inspected in a weight scan. Activation-only scans, vector steering, and fine-tuning studies are outside this neuron reply index. Download coverage manifest ↓