From model answers to a traceable verdict.
Five components turn parallel model outputs into claims, evidence, verdicts and a traceable synthesis.
Follow one question through the architecture.
Click a component to see its input, role and output.
See how model answers become comparable claims.
Click an answer or a claim. The Claim Engine normalizes different wording before detecting agreements and contradictions.
The DPA covers processing and appears to restrict region.
A DPA exists, but exclusive EEA processing is not established.
SCCs may cover some transfers.
Subprocessor location still needs verification.
Current configuration may contradict regional exclusivity.
Comparable atomic statement · #1
Comparable atomic statement · #2
Comparable atomic statement · #3
Comparable atomic statement · #4
The graph preserves which models support, qualify or contradict each claim.
One question. Multiple AIs in parallel. Then verification.
The same question is sent to multiple systems. Recherche.fr compares independent answers before verifying the claims that matter.
Each model answers separately.
Answers are decomposed into comparable statements.
Model positions are compared at claim level.
Documents, registries, standards and authoritative sources.
Our specialized model adjudicates against the evidence.
The conclusion remains linked to models, claims and evidence.
Five components, five distinct responsibilities.
The architecture deliberately separates model selection, claim extraction, verification, adjudication and synthesis. Some foundations are implemented today; the evidence-aware components remain active R&D and are being benchmarkinged and calibrated.
Recherche Router
Selects or recommends models and execution strategies according to task, domain, diversity, cost, latency and observed performance.
Claim & Contradiction Engine
Breaks long answers into comparable claims and separates material contradictions from differences in wording or style.
Verification Router
Decides which claims need independent checking, what type of evidence is appropriate and how much verification is justified by risk, cost and latency.
Recherche Judge
Adjudicates at claim level using evidence quality, source authority, freshness, contradictions, provenance and uncertainty — not model popularity.
Recherche Synthesizer
Builds a final answer without hiding unresolved points and preserves the path back to the models, claims, evidence and verdicts that produced it.
What already exists
The current platform is not a concept deck. Work already covers authentication, workspaces, provider configuration, encrypted BYOK, research sessions, response persistence, multi-provider adapters, comparison, contradictions, synthesis, queued execution, staging and CI/E2E.
More models are not the research contribution.
The difficult part is deciding when agreement is misleading, which claims justify the cost of verification, what counts as authoritative evidence, how Judge decisions should be calibrated and how reliability changes as models, providers and sources evolve.
Explore the research →Trace what each layer actually changes.
Select a stage and scenario. The module shows its input, transformation, output, and what fails if it is removed.
The hard part is not calling eight models.
Fan-out and orchestration are standard engineering. The difficult work sits in the transformations that must remain inspectable and testable between model output and final answer.
Resolve paraphrases, negation, scope and coreference without erasing meaningful disagreement.
Track authority, freshness, provenance and conflicting evidence so multiple copies of the same assertion do not become independent proof.
Evaluation design must test blinded provider identity, ordering effects, persuasive wording and uncertainty handling.
The Synthesizer should remain constrained by adjudicated claims, unresolved states and provenance.
Verified Claim–Evidence Graph.
The architecture itself is reproducible. The potentially harder-to-replicate asset is the accumulated history of claims, contradictions, evidence, adjudications, corrections and later outcomes — if it is built with rigorous provenance and governance.
Track paraphrases and revisions without collapsing materially different statements into one node.
Evidence is attached with time and source lineage so historical decisions can be re-evaluated.
Each decision records the evaluation context rather than storing only a final label.
Corrections can become evaluation material for Judge, Router and verification-policy research.