03 · Confidence · Composition layer

Lattice

Confident recommendations

A patent-pending gating layer on your sequential backbone — compose learned structure when support is sufficient; defer to baseline when it is not.

Move your cursor — confidence crystallizes near the probe.

RecSys 2026 — under reviewPatent pendingarXiv preprint ↗
+31.7%Gated HR@10 vs LSTM backbone (MovieLens, 30 paired seeds)
OFF ≡ backboneGate off: predictions match your model exactly (parity verified)
ON ≡ fusionGate on: same behavior as always-on fusion on supported rows

Not a new recommender — a composition policy

Lattice sits on top of the backbone you already ship. Binary ON/OFF controls whether an auxiliary scoring path is blended in — not whether items are filtered from a list.

1

Your backbone

LSTM, transformer, SASRec — any sequential ranker you already train and serve. Lattice never replaces it.

2

Confidence gate

When in-support confidence is below threshold, the gate stays OFF. Metrics and rankings match baseline exactly — auditable, not down-weighted.

3

Backbone-native fusion

When ON, archetypes fit in the backbone's embedding space and fusion activates. Ungated mode reaches higher pooled HR; gating trades peak benchmark for control.

Validated results

MovieLens primary eval: 30 paired seeds, full-catalog ranking, identical test rows per seed. Lifts are vs each backbone alone — not literature SOTA comparisons. Benchmark access for independent verification: LorianBannis@banlys.com.

BackboneDatasetGated lift (HR@10)Ungated liftSeedsNote
LSTMMovieLens+31.7%+58.7%30Primary estimand
TransformerMovieLens+13.3%+20.9%30Backbone-native archetypes
SASRecMovieLens+17.0%+26.9%30Contextual baseline + Lattice
LSTMAmazon Electronics+124.0%+211.6%15Supporting — high variance

Primary headline: LSTM +31.7% gated (+58.7% ungated). Transformer +13.3% and SASRec +17.0% use backbone-native archetypes. Amazon +124.0% is supporting evidence only (small test set, high variance). We do not claim gating maximizes in-distribution accuracy — we claim auditable ON/OFF with zero harm when OFF.

When Lattice composes vs defers

GATE ON

Compose

  • Session in support for learned behavioral structure
  • Auxiliary fusion path enabled on top of your backbone
  • On gate-ON rows, matches always-on fusion in our eval
  • Primary lift: +31.7% HR@10 vs LSTM on MovieLens (30 seeds)
GATE OFF

Defer

  • Low support, short history, or unfamiliar input
  • Auxiliary path fully disabled — not down-weighted
  • Predictions match your backbone exactly (parity verified)
  • Under distribution shift, gate tends to stay off — metrics match baseline

The trade-off (stated plainly)

Always-on fusion scores higher on pooled benchmarks in our ablation (+58.7% vs +31.7% gated on MovieLens). We do not sell peak benchmark mode. We sell an auditable ON/OFF switch: when off, predictions match your backbone; when on, you capture full fusion gain on supported rows.

Baseline vs gated vs always-on

Lattice gates whether an auxiliary scoring path is composed with your backbone — not whether individual items are filtered from a list. When the gate is OFF, ranked output matches the backbone exactly; when ON, it matches always-on fusion on the same session.

Session support (illustrative)55%
Gate threshold θ40%

GATE ON — auxiliary fusion active

Gated ranking matches always-on fusion on this session (same behavior as our MovieLens eval: OFF ≡ LSTM, ON ≡ ungated fusion).

Backbone only (baseline)

  1. 1.Viral TikTok Hit(Comedy)
  2. 2.Trending Sequel(Action)
  3. 3.Hype Train Movie(Drama)
  4. 4.Inception(Sci-Fi)
  5. 5.The Dark Knight(Action)

Lattice (gated)

  1. 1.Inception(Sci-Fi)
  2. 2.The Dark Knight(Action)
  3. 3.Interstellar(Sci-Fi)
  4. 4.The Matrix(Sci-Fi)
  5. 5.Pulp Fiction(Crime)

Always-on fusion (ungated)

  1. 1.Inception(Sci-Fi)
  2. 2.The Dark Knight(Action)
  3. 3.Interstellar(Sci-Fi)
  4. 4.The Matrix(Sci-Fi)
  5. 5.Pulp Fiction(Crime)

Mock rankings for education only — composition policy (OFF ≡ backbone, ON ≡ fusion), not item filtering. Does not expose production confidence formulas or training pipelines.

Research & commercial pilots

Eval protocol and pilot KPIs focus on coverage, lift when ON, and zero harm when OFF. Hosted API access follows pilot validation — not self-serve yet.

A Banly's product · Patent pending · RecSys 2026 under review · LorianBannis@banlys.com