Questions we get.
Honest answers.

Different shapes of question, different tools. Phronesis is for when you have options and need to rank them — "Which of these three CRMs should we pick?" It runs three blind analyst models and returns a weighted ranking with confidence and per-dimension scores. Synthesis is for when you need a research brief — "What are the strategic tradeoffs of entering this market?" It runs five models in parallel and merges their output into a single structured brief.

Synthesis runs GPT-4o-mini, Claude Haiku, Gemini 2.5 Flash, DeepSeek Chat, and Grok-3 in parallel. Expert mode swaps in the flagship tier (GPT-4o, Claude Sonnet 4.6). The merge/strategist layer is Claude Sonnet. Phronesis uses three different models (from the same pool) as its blind analyst jury, selected based on the depth you choose (quick/standard/deep).

In our internal benchmark across 20 expert-labeled decisions, Phronesis scored 7.3/10 overall with winner accuracy at 7.7/10. More usefully: when the three analysts agreed, human experts agreed 84% of the time; when analysts split, experts split 71% of the time. The agreement signal is genuinely predictive of decision quality. Full methodology here.

Diversity of priors. A single frontier model has one training-shaped view of a question. Five models from different vendors have five different priors — and the union of their coverage beats any single model's output on genuinely open questions. Our head-to-head benchmark found a 65% win rate for orchestration over single-Claude.

Phronesis: 15–90 seconds depending on depth (quick / standard / deep). Synthesis: 30–90 seconds in standard mode, up to several minutes in expert mode. We stream progress events live so you always know which phase is running.

Not via the hosted product — we use our own keys so pricing stays simple and predictable. If you want BYOK for compliance or cost reasons, the codebase is designed to support it and we can talk about a self-hosted tier. Get in touch.

Your run history lives in your browser's localStorage, not on our servers. We store completed runs temporarily so you can share a result URL, but we don't associate them with a user account (there are no user accounts on the free tier). Paid tiers can opt in to server-side history. See the full privacy policy.

No. We pass your inputs through to the underlying model providers (OpenAI, Anthropic, Google, etc.) under their zero-retention enterprise terms where available. We do not train any of our own models. We do not retain inputs beyond the run lifecycle.

Yes. Every decision and research brief can be exported as PDF, JSON, or Markdown. Your localStorage history can be cleared from the History page or from your browser's settings. For server-side data (paid tier), email security@intellcluster.com and we'll honor deletion within 7 days.

We follow GDPR principles: local-first data where possible, right-to-delete on request, fonts served from EU-friendly providers (fonts.bunny.net), and no third-party trackers loaded unless you're on a paid account that opts in.

Yes. The free plan includes five Phronesis runs and five Synthesis briefs per month. No credit card. Upgrade when you outgrow it.

Buy a top-up credit pack (5, 15, or 30 credits) or wait until your next monthly reset. Credits from top-ups never expire.

Yes. Cancel from the Stripe billing portal. Paid features stay active until the end of your billing period. No cancellation fees, no long-term contracts.

The Team plan covers 5 seats with shared history, SSO, and priority support. For larger deployments, on-prem, BYOK, or custom SLAs — talk to us.

Yes. The /phronesis/api/decide and /synthesis/api/run endpoints accept the same inputs as the web UI and return structured JSON (or streaming SSE). See /docs for the full reference including rate limits, error model, and SSE frame format.

Not supported publicly yet, but the codebase is designed to run anywhere a FastAPI monolith can. If self-hosting is a hard requirement for compliance, get in touch.

30 API requests per minute per IP by default. We return standard X-RateLimit-* headers so clients can back off cleanly. Paid tiers get higher limits.