Blue marks the final written word for orientation. The actual readout was at the final model token; the records do not include a token-by-token activation map.
Look at the sentences and their effects, then write a testable hypothesis. Notes stay on your device and are not published as verified explanations.
RECORDED ABLATIONS · REAL MODEL RESULTS
Inspect a recorded edit. See the output effect.
Choose an actual model, pick an observed example, and switch a recorded SAE feature off.
01
Choose a model
⌄
02
Pick an input
Feature —
03
Try the intervention
This compares recorded output scores, not generated text. Choose a different model or input to explore another actual run.
WHAT CHANGED?
Output response
✓ Measured
Change in output scores
0.00 L2
Original output · no feature removed
Three recorded inputsChoose a bar
Changes in output scores after the recorded feature removal
TRY SOMETHING ELSE ↘
64 REAL INPUT STATES
Change one thing. See the new input.
Explore the actual six-factor study inputs. Toggle a factor, pick any state, and inspect the exact text used in the experiment.
01
Choose the factors
02
Observe the input
Selected input
All 64 combinations
Click any square
These are real authored experimental inputs. This explorer changes the input text; it does not invent model activations for unseen configurations. Open full dataset ↗
48 VERIFIED SITES
Explore every layer. Compare real measurements.
Choose a model and a measurement. Tap any layer to inspect its recorded results.
These are historical evaluations over 64 controlled input groups per site. For fair comparisons, start with layers of the same model and SAE. Download all results ↗
RECORDED RESEARCH
Beyond single features. Explore the evidence.
Two questions: how do features work together, and what happens to features when the input changes?
FEATURE INTERACTIONS · SAR
Can single-feature effects explain combinations?
Select a study size to explore recorded coalition experiments.