Why we tried it
When behavior uses hidden state, logged prediction and marginal action coverage need not identify interventions.
What the evidence shows
Exact finite-model calculations retain observationally equivalent, causally different alternatives and fixed/adaptive query references.
Where the work stands
The result concerns a specified resettable finite family, not general impossibility or neural causal capability.
Possible next steps
Disclose collector information, logs and conditional assignment assumptions and challenge identification with causally different observational alternatives.
Suggested next steps are untested research questions, not new experimental results. The archive records manuscript provenance, licensing and status.