[Endorsement Request] cs.LG — When does a predictive system learn it's the one predicting? (40 experiments, 12 falsified hypotheses)

Hi everyone,

I’m an independent researcher looking for an arXiv endorsement for cs.LG.

My paper asks: under what conditions does a minimal neural system (192-dim GRU, <100K params) learn to distinguish self-caused from world-caused changes? We propose agency gain — the predictive gap between a self-aware and a self-blind predictor sharing the same hidden state — as a directly trainable, density-free alternative to empowerment.

Key results from 40 controlled experiments:

• A 93.7% prediction gap survives removal of all auxiliary components

• Awareness must consolidate before intention (simultaneous learning fails in all configurations)

• 12 falsified hypotheses mapping the boundary between “systems that predict” and “systems that know they are the ones predicting”

• Forward-sampled action selection succeeds; two gradient-based alternatives degenerate

Preprint: https://doi.org/10.5281/zenodo.20523162

If you’re qualified to endorse for cs.LG and find this relevant, my endorsement code is 7XJ7XN.

https://arxiv.org/auth/endorse?x=7XJ7XN

Endorsement confirms the work is legitimate scholarly material, not an endorsement of conclusions.

Happy to answer any questions about the paper. Thanks!