Europe should not need one AI provider to verify another.
Verification should remain portable across models, clouds and sovereign infrastructure. Recherche.fr separates the logic used to verify a claim from the provider being evaluated.
Change the provider without changing the verification logic.
Choose a provider: the verification layer underneath stays the same.
The independent layer remains portable.
The provider can change. The verification logic stays.
Click a provider above: only the source changes; the verification chain remains stable.
Why verification independence matters in Europe.
European organisations are unlikely to run on a single model, cloud or data environment. A verification layer therefore has to remain independent from the system it evaluates, portable across deployments and compatible with different infrastructure choices.
Provider independence
Compare and verify heterogeneous systems while keeping the verification logic separate from any single model provider.
Interoperability
Design interfaces that can operate across different model APIs, deployment environments and organisational stacks.
Sovereign compatibility
Support architectures where models, evidence services or workloads may run on European or sovereign infrastructure.
Cross-border R&D
Benchmark and validate the verification layer with research and industrial partners in representative environments.
Industrial validation
Measure quality, cost, latency, portability and failure modes under real operational constraints.
Open interfaces
Keep model selection, evidence acquisition and adjudication modular enough to integrate with existing systems.
Verification should survive a change of model, cloud or deployment environment.
Claims
The same claim representation should remain comparable even when the underlying provider changes.
Evidence
Evidence acquisition and qualification should not depend on the provider that generated the answer.
Verdicts
Adjudication criteria should remain stable enough to compare systems and versions over time.
Provenance
The path from model output to claim, evidence and verdict should remain inspectable across deployments.
Three concrete collaboration paths.
Prospective collaboration, not a partnership claim.
Recherche.fr is seeking research and industrial environments in which the verification architecture can be tested, measured and improved. Collaboration scope is described as prospective until formally agreed.
Portability test design: what should survive a provider change.
Illustrative test design — not a measured portability result. It shows which variables a real provider-removal test would need to track.
These concepts remain stable when provider or infrastructure changes.
Compatibility and portability are architecture goals, not certifications.
Portability must be measured, not asserted.
The interactive module on this page is an illustrative test design. It does not prove sovereignty or portability. A real portability claim requires provider-removal tests with recorded changes in evidence coverage, unresolved states, Judge output, cost and latency.
Keep claim schema, evidence requirements, provenance rules and adjudication criteria constant.
Run the same case with one provider removed, then with a different provider mix, including an EU/open-weight configuration where feasible.
Compare claim coverage, contradictions, evidence coverage, unresolved states, Judge outcome, cost and latency.
Every changed dependency must remain visible in the decision trace rather than being hidden behind a “portable” label.
What would count as evidence of portability
- The same verification contract remains usable after provider substitution.
- Rerouting and quality degradation are explicitly measured.
- Provenance and unresolved states remain intact.
What would not count
- Pre-programmed UI toggles.
- Provider logos or “sovereign-compatible” labels alone.
- A claim of invariance without a recorded comparison.