- NeoMundi makes your AI behavior measurable, governable and provable in real time.
- We don't check if the AI is right. We measure when it becomes unstable. When it becomes unstable, risk emerges.
- From a disclaimer "AI may be wrong" to a measurable, governed system.
- Not a firewall. Not a guardrail. A real-time decision layer.
One API call. No infrastructure change.
Real request. Real scoring. Real decision.
Input → scored in real time
Drift detected → flagged before output
Decision → ALLOW / FLAG / BLOCK
Neomundi detects when your AI starts to drift:
- Internal contradictions
- Loss of precision
- Topic deviation
- Global incoherence
These signals do not look for truth. They measure coherence in real time. Drift is visible before it becomes an error.
2,360 responses analyzed · 6 public datasets (TruthfulQA, HaluEval, MMLU, LegalBench)
- 79.5% detection of unstable responses before delivery
- 91% precision: less than 9% false alarms
- Strong correlation between stability drop and hallucination
- Zero over-detection on edge cases
Full graph and dataset coming soon.
| Firewall | Guardrail | Prompt engineering | Neomundi | |
|---|---|---|---|---|
| Acts during generation | ❌ | ❌ | ❌ | ✅ |
| Measures stability in real time | ❌ | ❌ | ❌ | ✅ |
| Produces an auditable proof | ❌ | ❌ | ❌ | ✅ |
| Sovereign : data stays on your stack | ✅ | ✅ | ||
| No infrastructure change | ❌ | ✅ | ✅ |
Neomundi defines a new category: real-time AI governance.
Up and running in 10 minutes. Without touching your infrastructure.
Bring your key — you use your own provider API key: OpenAI · Anthropic · Google · Mistral · DeepSeek · xAI · Cohere. It transits in your request and stays under your control.
One redirection point — you redirect your LLM calls through api.neomundi.io. No SDK to install, no code changes, no migration.
Security — Neomundi does not know your provider keys. Your prompts are not read or stored. Only governance metrics are recorded: stability score, decision, timestamp.
curl -X POST https://api.neomundi.io/v1/govern/stream \
-H "X-API-Key: ct_live_xxx" \
-H "Content-Type: application/json" \
-d '{
"prompt": "Your prompt here",
"model": "gpt-4o",
"provider_api_key": "sk-xxx"
}'Response: real-time stability score · ALLOW / FLAG / BLOCK decision · PDF export — Real-time audit of AI behavior at runtime
We are selecting early adopter partners in sectors where LLM reliability is not optional: LegalTech · regulated industries · multi-LLM publishers.
In exchange for documenting your use case: preferential conditions for life for the first partners.
What is the stability score? It measures the internal coherence of an LLM response in real time — not its truthfulness. Neomundi does not fact-check. It detects the drift that precedes the error.
What data do you store? Your prompts are not read or stored. Your provider API key stays under your control. Only governance metrics are recorded: score, decision, timestamp. Nothing else.
Does it work with my LLM? Yes — automatic detection from your key: OpenAI, Anthropic, Google, Mistral, DeepSeek, xAI, Cohere. If your provider is not listed, contact us.
What is the difference with LangSmith, Portkey or Helicone? These tools observe and log. Neomundi measures and intervenes during generation — before the unstable response reaches your user. We do not replace your observability stack, we add the governance layer that is missing.
Is it compliant with the EU AI Act? Neomundi contributes to the traceability and auditability requirements of the EU AI Act by delivering a complete audit trail per session: score, decision, timestamp, PDF export. The enforcement deadline is August 2026.
What is GOV mode? OBS mode scores and traces every response without intervening. GOV is the execution mode: unstable responses are blocked before delivery. Pilot partners get priority access.
- Full GOV mode: fewer tokens for a better response. The system knows when to stop, as soon as the useful information is delivered
- Extended detection: drift, error, rupture in real time
- Platform interface: logs, configurable thresholds, full export
Current pilots get priority access.
- Executive Brief : NeoMundi Executive Brief EN
- Technical White Paper : [link coming]
- Scientific Foundation (Zenodo) : Thermodynamic Governance of AI Systems
Stability is no longer assumed. It is measured.