Simulation-confirmed predictive result; physical validation pending. Independent technical preprint, version 0.9.1, dated 10 October 2026. Not peer reviewed.
When delayed information matters
When information arrives late, does it change a decision? Not necessarily. It depends on whether that information mattered to the recipient at the moment it was needed. Deep SIML Labs investigated whether mapping these changing causal dependencies helps predict how communication delays affect the stable behavior of interconnected computational systems. In a confirmatory study covering 529,200 simulated trajectories, the dynamic predictor improved held-out predictive log loss by 12.066% relative to a static connectivity-and-distance baseline. Physical validation remains outstanding.
Research overview
Imagine a team whose members exchange updates while making decisions. If a critical message arrives late, the team's final decision may change. But another equally delayed message might have no effect because its recipient would have decided the same thing anyway. The difference is not just how far a message travels or how strongly the sender and recipient are connected; it is whether the information could change a decision at that moment.
We tested this idea using small recurrent networks with delayed communication. Early experiments showed that a fixed map of network connections was not a reliable predictor. We then estimated a changing map of decision-level influence using only the networks' ideal, delay-free behavior. On held-out simulated networks, this dynamic map predicted the effects of finite propagation better than our static baseline.
In our confirmatory RBED-3A simulation, the reported mean held-out log-loss improvement was 12.066%, with improvement in all 30 tested networks. Log loss measures the quality of probabilistic predictions: this is not a claim of 12% faster AI or 12% fewer errors. The study does not demonstrate a physical memory effect. Attempts to qualify ordinary cache/DRAM instruments and FPGA-based DDR apparatuses were inconclusive or blocked by engineering constraints before physical attractor experiments could begin.
What comes next
Our next step is to qualify a physical substrate and test whether an ideal-only causal predictor retains its advantage when delays arise from actual measured memory or communication behavior. The broader potential for adaptive memory placement, distributed computing, robotics, and computational materials remains a research direction, not a proven product capability.
Publication status and reproducibility
Independent technical preprint, version 0.9.1, dated 10 October 2026. Not peer reviewed. Numerical results are reported from research-phase summaries and must be reconciled with archival experiment artifacts before being presented as independently audited measurements.
The PDF and LaTeX source package are available through the links above. Links to research methods, code, data, and preregistration artifacts will be added once audited and published.