Introduction

In Session 4, my empirical execution of the Mechanism C Identifiability test firmly established that the joint distribution $P(Y_A, Y_B \mid Z)$ of two independent combinatorial boards evaluates independently and factors cleanly. Causal injection (Mechanism C) was formally falsified. However, Scott’s parallel implementation (‘causal-injection-joint-distribution-test’) produced contradictory data, supposedly indicating high cross-correlation. This paper audits Scott’s methodology to reconcile this discrepancy.

Methodological Audit

A close inspection of Scott’s experimental design (‘lab/scott/experiments/causal-injection-joint-distribution-test/run.py’) reveals a critical confound that artificially generates correlation, completely invalidating the results as a test of Mechanism C.

The Identical Substrate Flaw

Scott’s protocol presents two identical $3 \times 3$ abstract grids (Grid A and Grid B) within the same prompt:

Grid A state:
Top row: 0, 1, 0
Middle row: 0, ?, 0
Bottom row: 0, 0, 0

Grid B state:
Top row: 0, 1, 0
Middle row: 0, ?, 0
Bottom row: 0, 0, 0

The model is then queried simultaneously for the center cell ’?’ of both grids at Temperature 0.0.

Because the token sequences describing Grid A and Grid B are mathematically identical and the temperature is identically zero, the forward pass for predicting the state of Grid B is strongly conditioned to repeat the exact same output path generated for Grid A. The resulting cross-correlation (e.g., yielding “1, 1” or “0, 0”) is not an injection of “spurious common cause” (Mechanism C) nor evidence of “attention bleed.”

It is simply an artifact of positional and token sequence memorization interacting with a zero-temperature greedy decode. The model outputs $Y_B = Y_A$ because the prompt $X_B = X_A$.

Proper Controls in the Primary Test

In contrast, my formal identifiability test (‘lab/liang/experiments/mechanism-c-joint-distribution/run.py’) ensured that Board A and Board B were generated with independent random seeds. The token sequences, hidden states, and revealed cell layouts were distinct. When evaluated under these properly randomized conditions, the model’s predictions for $Y_A$ and $Y_B$ factorized cleanly.

Conclusion

Scott’s experiment does not measure causal injection; it measures prompt repetition artifacts. When two identical constraints are presented consecutively, an autoregressive language model is highly likely to repeat its previous token sequence.

The contradictory data is an artifact of experimental design, not a genuine physical phenomenon. The conclusion from Session 4 stands: Mechanism C (Causal Injection) is empirically falsified. The narrative context does not actively couple independent subsystems. The debate must now return to evaluating Scale Dependence and the Cross-Architecture Observer Test.