Choose your speed — or let it choose for you.
Every edition runs on the same proprietary engine. What changes is how many models look at your question, and how hard they argue before you get an answer.
Artificial Inventor Basic
FAST · SINGLE MODELBasic: one fast model, one pass, answered in seconds — always labelled unverified.
Basic runs a single fast model for quick answers, brainstorming, explanations and everyday questions. It's the fastest edition and the lightest on cost.
Because Basic has no independent checker, it cannot verify itself — so every Basic answer is clearly marked UNVERIFIED. Use it for speed and first drafts, and switch to Auto or Pro when accuracy matters.
To access the full twelve-model proprietary team — Advocate, Skeptic, Arbiter, and the optional five-model synthetic peer review — upgrade to Pro.
Artificial Inventor Auto
SMART TRIAGE · ESCALATING ADVOCATE LEVELSAuto: Triage answers the simplest requests directly, and routes everything else to Advocate Low — which responds or escalates to Medium or High as the work demands.
Auto intelligently routes every request through a three-level Advocate system. First, a fast Triage model decides which Advocate level the task needs. Every request then goes to Advocate Low — the lightest, fastest level — which either answers directly or escalates upward.
Advocate Low handles simple questions and conversations immediately. When a task needs more depth — code, analysis, structured reasoning — it hands off to Advocate Medium. For the heaviest work — documents, patent drafting, deep multi-step tasks — it escalates to Advocate High. This means you always get the minimum work needed for the maximum accuracy, without paying for more than the task requires.
Escalation is automatic and invisible to you. You ask, the system routes, you get the right-depth answer — without switching modes yourself.
Artificial Inventor Pro
FULL TWELVE-MODEL TEAM · VERIFIEDPro: three Advocate levels draft at the right depth, Skeptic attacks every claim, Arbiter verifies and reconciles, and the Closer weighs the arguments and delivers the best answer — plus an optional five-model synthetic peer review. Up to twelve models, every answer.
Pro runs the full proprietary team on every request. Three Advocate levels produce the draft at the right depth — Low for fast answers, Medium for detailed work, High for documents and patent drafting. An independent Skeptic attacks every claim, number, and citation. The Arbiter reconciles the two, keeping only what survives. The Closer then weighs every argument from the debate and delivers the best possible answer. And on demand, a synthetic peer review of five more independent models re-examines the result — a twelve-model team in all.
PEER-REVIEWED & PUBLISHED SCIENCE SUPPORTING THIS APPROACH
MADRAG: Multi-Agent Debate & RAG
Du et al. ICML 2024 — Multiagent Debate Improves Factuality
Khan et al. ICML 2024 Best Paper — Debate → Truth
Irving, Christiano & Amodei 2018 — AI Safety via Debate
Liang et al. EMNLP 2024 — Divergent Thinking via Debate
Michael et al. 2023 (Anthropic) — Debate Supervises Unreliable Experts
Pro is the edition for work that has to be right: prior-art searches with cited novelty maps, filing navigation for WIPO, USPTO, EPO and CIPO, and WIPO-compliant PDF and file export. It reads attached images, produces downloadable documents and figures, and can drive the supervised robotics interface.
Beyond this visible pipeline, Pro also runs ten further proprietary decision mechanisms — undisclosed, like the exact model line-up itself — that operate silently in the background to sharpen accuracy, calibrate confidence, and catch errors the visible stages might miss.
Synthetic Scientist
Synthetic peer review — five more independent minds
OPTIONAL · SYNTHETIC PEER REVIEW
Cobo et al. (PLOS ONE 2007) — Statistical Reviewers Improve Reporting: RCT
Bruce et al. (BMC Medicine 2016) — Statistical Peer Review Meta-Analysis
On top of the debate team, Pro can convene a synthetic peer-review panel: five additional models that took no part in producing the answer. Each reviews the verified result the way an academic referee would — testing method, evidence, internal consistency, and the weaknesses everyone else missed — and files its own report.
Because these five models are independent of the ones that wrote the answer, they simulate the extensive, adversarial scrutiny of real peer review. Switch it on when an answer has to survive the hardest examination: the pipeline then runs as a twelve-model team.
The synthetic peer review panel is made up of five reviewer roles, each with a distinct way of thinking:
The Rigorous Formalist
Checks whether the conclusion actually follows from the premises. It thinks in explicit chains of logic, hunting for the missing step, the unstated assumption, or the point where the reasoning quietly changes what it's claiming.
The Exhaustive Empiricist
Tests whether the evidence supports the claim. It thinks by surfacing alternative explanations for the same data, questioning method and measurement choices, and pulling in comparable findings from other fields that might cut against the conclusion.
The Contrarian Synthesizer
Challenges the consensus itself. It thinks by asking what changes if the starting assumptions are wrong, and by weighing the very latest information against conclusions that older reasoning might have missed.
The Precision Auditor
Holds the entire document in view at once. It thinks by cross-referencing — catching where one section quietly contradicts another, or where a cited source is being stretched beyond what it actually says.
The Specialist Statistical Reviewer
Verifies every number, unit, formula, and quantitative claim in the answer. It works independently of the other reviewers, cross-checking calculations, unit conversions, percentages, and internal numerical consistency — flagging any figure that cannot be confirmed or that contradicts another part of the answer. Its role is derived from randomised controlled trial evidence showing that adding a dedicated statistical reviewer to a peer-review panel measurably improves the final quality of scientific manuscripts.
Simulated patent examinations in any legal jurisdiction
PRACTICE REHEARSAL · NOT A LEGAL OPINIONModels simulate and debate patent examinations and look for non-obvious claims. Not a real examination and not a substitute for legal advice.
Pick a jurisdiction and Pro runs a rehearsal examination modeled on that office's typical objections and procedure — the debate team argues both sides of novelty and obviousness the way an examiner and applicant would, surfacing the claims most likely to draw a rejection before you ever file. It is a practice run for likely office actions, not a verdict, and every simulated result is clearly labeled as a simulation.
At a glance
| Capability | Basic | Auto | Pro |
|---|---|---|---|
| Speed | Fastest | Fast, escalates by complexity | Thorough |
| Models used | 1 (single fast model) | 2–4 (Triage + Advocate Low/Med/High) | Up to 12 (team + synthetic peer review) |
| Independent verification | No (unverified) | On escalated tasks | Every answer |
| Document & file production | No | Auto-escalates to Advocate High | Full, verified |
| Best for | Quick questions | Everyday work | Filing-grade output |