AffectLoop: From Experience to Conditional Self-Reaction Prediction with Counterfactual Twins

Persistent agents can experience the same aggregate outcomes yet develop different expectations because those outcomes occurred in different contexts. We introduce AffectLoop, a counterfactual-twin assay connecting experience-dependent individualization to conditional self-reaction prediction. Matched histories isolate contextual associations; state exchange tests whose experience supports a prediction; local intervention and experience tracing explain the dependence. In a waveform world, correct-state numerical error is .0580, compared with .3050 after twin exchange and .1659 after global pooling. A precommitted 12-pair ordinal confirmation yields 95.83% correct-state accuracy, 12 positive pair contrasts, and 10 complete protocol passes (median contrast .4375; exact two-sided p = .000488). A structurally different object-field world yields 100% correct-state accuracy and 12/12 composite passes; local intervention changes 98.96% of targeted labels while preserving all non-target labels. Beyond the designed Reaction kernels, we apply the correspondence test to the public genagents implementation's own categorical responses. In a separate precommitted 12-pair experiment, prediction agrees with those responses in 100% of cases using the corresponding memory, 0% after exchange, and 50% without memory. A memory-method comparison traces a numerical difference to selected experiences, and a collaborator reproduces that computation. Reader substitution identifies the explicit final operation in the original ordinal interface. The contribution is a reproducible method for establishing how experience forms individual state, whose prediction that state supports, and where prediction differences arise, demonstrated in two designed worlds and one external agent implementation. Release note: This record provides the manuscript PDF for version 11.1. Code, experimental data, and supplementary materials are not included in this release. This is a preprint and has not been peer reviewed.