MOSAIC: multi-observer analysis with intelligent consensus

The problem with trusting a single model

A language model can misread a figure, mix up two dates or put a document in the wrong category — and do so with complete apparent confidence. When the result feeds a decision — a cover, an amount, a clause — that error has a cost.

How MOSAIC works

• Multiple observers: the same content is analyzed by several specialized models.
• Cross-checking: their outputs are compared against each other.
• Intelligent consensus: MOSAIC reconciles those readings into a single result, reducing errors and inconsistencies.

Where it adds the most

• Extracting critical data from contracts and legal documents.
• Classifying ambiguous documents that could belong to several sectors.
• Processes where human review is expensive and best focused on uncertain cases.

Example: a date two models read differently

Illustrative example of how it works, not a specific customer.

In a scanned contract, the expiry date appears as “01/02/2027”. One model reads it as 1 February and another as 2 January. MOSAIC detects the disagreement, reconciles it using the rest of the document — language, format of other dates, signature date — and flags it so review focuses on that value, which VERA lets you locate in the original.

Limits and good practices

• Cost and latency: running several models uses more resources than one; it pays off where a mistake is expensive.
• Consensus is not infallible: several models can agree on an error, which is why every value stays verifiable at its source.
• Focused review: use disagreements to direct human review to the uncertain cases.

Together with VERA

MOSAIC decides which value is correct; VERA shows where it comes from by anchoring it to its region in the document. Together they produce results you can audit. This approach is one of the differences from using ChatGPT or Claude directly.