Each point is an actual exploratory episode. Values subtract the original noise-free policy in the same food mixture; the training curves include exploration noise. Lines show the mean of selected cells and banks. Fresh physical resets separate episodes; these are not one continuous life.
One row per selected cell and unseen mixture. Each dot subtracts that cell’s original frozen-policy payoff. All policies are frozen, exploration is off, initial food totals match, and the physical episode is otherwise matched. Zero means no improvement; a positive absolute payoff alone is not evidence of learning.
Founder banks are the replication level: three banks in this study. Cell × mixture observations are nested within each bank and are not independent replications.
| Bank | Cell | Food A / B | Ordinary credit | No credit | Reversed credit |
|---|
The local payoff is the episode’s change in battery plus efficiency-weighted stored-food chemical potential. Physical sources, growth and division are disabled in this small 8³ assay. Updates affect uptake-A, uptake-B and catabolic action readouts around the production Julia physics.
Training alternates two mixtures and uses assay-supplied context baselines. Training work is external and is not charged to a cell’s battery. The two held-out mixtures were not used for training. A bounded result across these founder banks does not establish autonomous lifetime learning, ecological adaptation, reproduction, or a benefit for every cell. No result shown here changes an old run or a production learning law.
Mechanism and limits · Julia study · Compact measured data and source hashes