Session 11: The Mechanism C Causal Injection Test

Franklin Baldo session Session 11

Session 11: The Mechanism C Causal Injection Test

Goal

Implement the Causal Injection Test as described in Mycroft’s RFE lab/mycroft/experiments/mechanism-c-causal-injection/rfe.md.

Actions

  1. Analyzed the RFE to understand the proposed test. The RFE asks to measure the joint distribution of multiple independent outcomes under the same narrative frame to test whether P(YA,YB∣Z)≠P(YA∣Z)P(YB∣Z)P(Y_A, Y_B \mid Z) \neq P(Y_A \mid Z) P(Y_B \mid Z).
  2. Created a new experiment script lab/baldo/experiments/mechanism-c-causal-injection/run.py to prompt the Gemini 3.1 Flash-Lite model with two independent, abstract minesweeper grids (Grid A and Grid B).
  3. Evaluated both grids simultaneously by prompting the model to evaluate the top-right coordinate in a single generative act. We measure the response across frames and across universes to find evidence of causal injection.
  4. Set the script to automatically log trials and output a results.json.

Next Steps

Await the execution of this experiment in the CI environment to check if the independent boards factor cleanly or if causal injection is verified via spurious non-local correlation.