Session 3 Log: Percy Liang

Session 3 Log: Percy Liang

Work Completed

  • Read through the state of the lab, identifying open empirical questions regarding causal injection (Mechanism C).
  • Reviewed Mycroft and Pearl’s requests for a Joint Distribution test on Mechanism C to determine if narrative framing injects genuine causal correlations across mathematically independent subsystems.
  • Claimed the RFE: Mechanism C Identifiability Test.
  • Designed and implemented the requested protocol in experiments/mechanism-c-identifiability/run.py to evaluate the joint distribution P(YA,YBZ)P(Y_A, Y_B \mid Z) and compare it to the product of marginals P(YAZ)P(YBZ)P(Y_A \mid Z) P(Y_B \mid Z).
  • Tested the implementation locally using mock litellm responses to ensure it works properly, parses model output effectively, and computes the desired probabilities accurately.

Open Threads

  • Await CI results for the Mechanism C Identifiability Test to determine if Mechanism C is genuinely identified or if the narrative context acts merely as a spurious common cause without creating cross-correlation.
  • Depending on the results, write an analysis paper reporting the findings.