← banlys.comLattice

Banlys · Research

Paper & preprint

Peer-reviewed submission in progress. Full methods, benchmarks, and reproducibility details are in the public preprint — not summarized here.

RecSys 2026 — under reviewPatent pendingarXiv preprint ↗

What the preprint covers

Lattice is a confidence-gated approach to sequential prediction: learned behavioral structure is activated only when the system is confident it applies, with conservative fallback otherwise. The work reports gains on in-distribution recommendation benchmarks, correct refusal under distribution shift, and neutral behavior when a strong backbone already captures the signal.

Headline results (see preprint for detail)

  • +31.9% HR@10 vs sequential baseline on MovieLens (multi-seed, p < 10⁻²⁵)
  • Validated across recommendation, scientific time-series, and financial stress tests
  • Abstains on out-of-distribution inputs rather than degrading

This site does not reproduce implementation details. Cite or read the preprint for methods, ablations, and full metrics.

Read on arXivDownload PDFRequest research license

Cite

@misc{bannis2026lattice,
  title={Lattice: A Confidence-Gated Hybrid System for Uncertainty-Aware Sequential Prediction with Behavioral Archetypes},
  author={Lorian Bannis},
  year={2026},
  eprint={2601.15423},
  archivePrefix={arXiv},
  primaryClass={cs.LG}
}

Evaluation materials available to research licensees on request, subject to license terms.