How the Landvex Consensus Engine Works: From Observations to Findings

Published July 20, 2026 · Data Quality

When multiple contributors observe the same asset, the naive approach is to take a vote. Three people say the lamp is working, one says it is not — majority wins. The problem is that votes treat every input as equally reliable, which is never true.

The Landvex Consensus Engine takes a different approach. It does not vote. It weighs.

Five signals, one finding

Every observation that enters the engine carries metadata beyond the raw answer. The Consensus Engine combines five weighted signals into a single scored finding:

1. AI model confidence. The underlying detection model's certainty about what it sees — a crack, a working lamp, an obscured sign.

2. Human validation. Where available, expert review of a sample of observations, used to calibrate model performance per category and condition.

3. Historical consistency. Whether the current observation aligns with the known history of the same asset. A lamp reported as broken for the third time in two months carries different weight than a first-time report.

4. Cross-contributor agreement. How closely independent observations of the same asset agree, weighted by each contributor's established reliability profile.

5. Official record alignment. Where official GIS or maintenance registers exist, the engine factors in whether the observation confirms or contradicts the registered state.

Weighted aggregation, not averaging

The engine does not simply average the five signals. Each signal's weight is adjusted dynamically based on the category, the asset type, and the available data quality. For a lighting asset, model confidence and cross-contributor agreement may dominate. For a vegetation assessment, historical consistency and official record alignment may matter more.

The output is not a binary yes/no. It is a confidence-scored finding: "Asset condition: degraded, confidence 87%, based on 4 observations, 2 validators, 1 contradiction with register."

What happens when signals conflict

Conflict is information, not noise. When the AI predicts "intact" and a trusted contributor reports "damaged," the engine flags the asset for priority re-observation rather than forcing a resolution. The contradiction itself becomes a finding: "Condition uncertain — conflicting evidence, requires verification."

This design avoids the most dangerous failure mode of consensus systems: false certainty. An unresolved conflict is more useful than a wrongly resolved one.

Traceability by design

Every finding produced by the Consensus Engine retains a full provenance chain: which observations contributed, what each signal scored, how the weights were applied, and what the confidence calculation yielded. A customer who questions a finding can trace it back to source.