Frequently Asked Questions

You can. Here's what you lose. One model asked to "give four perspectives" is one brain roleplaying. Our four independent models have genuinely different training data, different biases, and different blind spots. Push back on any single LLM and it caves — ours don't see each other's output until cross-examination. When all four flag the same risk independently, that signal means something. When one model lists five risks, you have no idea which ones matter.

Seven to twelve minutes for a full adversarial stress test. Four independent analyses, then five rounds of structured cross-examination. You submit at 6pm, walk into your IC meeting prepared. We optimised for depth, not speed — if you wanted fast surface-level answers, you'd use ChatGPT.

Four top-ranked AI models from different providers, selected daily based on independent analytical benchmarks. We don't disclose specific model names or versions — our edge is the adversarial process, not any single model. When a provider releases a stronger model, our benchmark system detects it and rotates it in automatically.

Enterprise-grade. AES-256 encryption at rest, TLS in transit, EU-hosted infrastructure (Germany). Each analysis is cryptographically isolated — no cross-client data access, no query aggregation, no model fine-tuning on your inputs. Ephemeral mode available: your question is processed and immediately purged, with only a cryptographic attestation retained. Full details on our Security page.

By default, encrypted analysis records are retained for 90 days so you can access your reports. After 90 days, they're automatically purged. You can enable ephemeral mode for any analysis — zero retention, cryptographic proof only. You can also request immediate deletion of any record under GDPR Article 17.

Yes. The adversarial debate architecture works across any domain — strategy, legal, technical, medical, policy. The system auto-detects your question's domain and adapts scenario labels, framing, and disclaimers accordingly. A strategy question gets Challenging/Realistic/Optimistic scenarios. A finance question gets Bear/Base/Bull with market data. Same rigour, different vocabulary.

Three things you can't replicate in ChatGPT tabs. First: structured cross-examination — each model must defend its position against specific objections from the other three, not just respond to a generic prompt. Second: steelman testing — the system forces models to strengthen opposing arguments before attacking them. Third: quantified agreement — you get actual scores showing where conviction is real (4/4 flagged the same risk) versus where it's fragile (2-2 split on a key assumption). The process is what creates the signal, not the models.