Restricting what a module can read may improve what a system learns to compute. We tested this hypothesis in a preregistered confirmation experiment using sixty four‑cell systems that share a frozen language‑model backbone and communicate through learned continuous packets.
Five conditions were varied: evidence masking, ownership markers, and replacement of foreign evidence with neutral filler. Each condition was evaluated across six initialization clusters, two data orders, on a single fresh task world.
When markers were available in both regimes, masking boosted accuracy on held‑out two‑ and three‑operation compositions, with median paired differences of 0.846 and 0.859 respectively; all twelve pairs cleared the required margins and the full preregistered behavioral criterion was met. The unmarked replication also passed.
No globally visible system passed the marker‑following check, leaving the effect of usable role information unresolved. The filler condition yielded seven full generalizers, but its decomposition criteria were inconclusive. Packet interventions in all eighteen audited masked systems showed the predicted intermediate‑value changes on eligible cases; these finite, success‑conditioned audits do not establish mediation.
The results confirm a large advantage of the tested masking regime while leaving finer attribution and generality open. Protocols, results, and checkpoints are public.
Review