How the Landvex Consensus Engine Works: From Observations to Findings
Three people say the streetlight works. One says it's dead. Majority rules, right?
Wrong. That's how you end up with a dark street and a surprised city council.
I've watched this play out in the field. A veteran contributor — 800 verified observations, zero disputes — reports a failed lamp. Two newcomers, still learning the app, mark it working because they photographed it at noon and the bulb looked intact. The lamp hasn't lit up in six weeks. Majority voting would call it fine. Reality would call it broken.
The Landvex Consensus Engine doesn't vote. It weighs. Every input is judged on its merits, not its popularity.
Five signals, one finding
Each observation carries more than a thumbs-up or thumbs-down. The engine blends five signals into a single scored finding:
1. AI model confidence. How sure is the machine? A crack detector screaming "90% certain" carries weight. One mumbling "maybe 40%" gets sidelined.
2. Human validation. Expert review of samples, calibrating model performance per category. We know which models overreact and which ones sleepwalk.
3. Historical consistency. A lamp reported broken three times in two months isn't having a bad day. It's dying. That pattern changes everything.
4. Cross-contributor agreement. Independent observers agreeing is gold. But we weight each contributor by their track record. The veteran beats the rookie, every time.
5. Official record alignment. Does the observation confirm the register or blow it up? Both are useful. One is more interesting.
Weighted aggregation, not lazy averaging
The engine doesn't average the five signals like a high school GPA. Weights shift dynamically based on asset type and data quality. Lighting assessment? Model confidence and contributor agreement dominate. Vegetation overgrowth? Historical consistency and register alignment matter more.
The output isn't yes/no. It's a confidence-scored finding: "Asset condition: degraded, confidence 87%, based on 4 observations, 2 validators, 1 contradiction with register."
That 87% tells you something. It tells you there's evidence, but not certainty. It tells you to act, but not to panic. It respects your intelligence.
When signals fight, we don't pick a winner
Here's where most systems break: conflict. AI says intact. Trusted contributor says damaged. What now?
We don't flip a coin. We flag the asset for priority re-observation. The contradiction itself becomes the finding: "Condition uncertain — conflicting evidence, requires verification."
False certainty is the most dangerous output any system can produce. An unresolved conflict is honest. A wrongly resolved conflict is a lie wearing confidence intervals.
Traceability by design — show me the receipts
Every finding carries full provenance: which observations, what each signal scored, how weights were applied, what the math yielded. Question a finding? Trace it back to the exact photo, the exact contributor, the exact model version.
No black boxes. No "the algorithm said so." Just evidence, documented and inspectable.
Request a pilot →
The street doesn't care about your majority vote. It cares about what's actually there.