Banlys · Research

Confidence when it matters.

Patent-pending architecture for sequential prediction. Lattice activates learned behavioral structure only when confidence is high — and stays conservative when it is not.

RecSys 2026 — under reviewPatent pendingarXiv ↗

Interactive demos are illustrative. Hosted inference and commercial licensing are open for research partners and pilots.

When Lattice speaks vs waits

The design goal is restraint — not always-on personalization.

GATE ON

Speaks

  • Input aligns with learned behavioral structure
  • Advanced scoring path enabled
  • Strong gains on in-distribution benchmarks
  • Validated on recommendation data (MovieLens)
GATE OFF

Waits

  • Short history or unfamiliar input
  • Advanced structure fully disabled — not down-weighted
  • Conservative fallback predictions only
  • Correct refusal under distribution shift (e.g. LIGO, financial stress tests)

Earned personalization

Lattice does not assume identity from the first click. It earns personalization as evidence accumulates — and defers to conservative baselines when the match is weak. Full methodology, benchmarks, and reproducibility are in the arXiv preprint.

Explore gating in the illustrative demo →

Validation snapshot

+31.9%HR@10 vs sequential baseline (MovieLens)
30 seedsMulti-seed evaluation
OOD-safeAbstains when patterns do not apply

Reported on MovieLens with rigorous multi-seed testing (p < 10⁻²⁵). Under distribution shift, Lattice abstains rather than forcing a match — a successful outcome in the preprint. See the paper for full metrics across domains.

Research & commercial pilots

We are accepting research license requests and scoped design partnerships while peer review is in progress. Hosted API access follows pilot validation — not self-serve yet.