Skip to content

Built from Virtuous Machines:
Towards Artificial General Science.

One engine underneath every feature. It assembles the relevant context, orchestrates a mixture of frontier models, runs them inside agents that think like scientists, reasons across hours of compute, and checks its own conclusions. Jump to autonomous pipeline →

Why this works

Three featured elements, of many.

Science that stands up to scrutiny is hard, and needs more than a good model. The platform utilises the same architectures behind systems that have run original research autonomously for hundreds of hours.

01Mixture of models

A mixture of models,
better than any alone.

Many frontier models are orchestrated as one system alongside our Explorer One model. Every sub-task routes to the model that handles it best; every conclusion is cross-checked across model families. One lab's blind spots are caught by another's strengths, so single-model bias is gone by construction.

You never have to choose a model. You get the strongest answer at every step.

Why this matters

Every frontier lab has its own inductive bias. Use one and you inherit it. Orchestrate all of them, and you've built something only an independent lab would.

Live mixture · live routing
ClaudeAnthropic
GPTOpenAI
GeminiGoogle
MistralMistral AI
GrokxAI
Explorer OneExplore Science
Explore Science orchestratorRoutes · cross-checks · synthesises
02In-house infrastructure

An in-house science stack scaling discovery.

The full scientific method, running autonomously. Verification is front-loaded where errors are costliest and backed by tools that can prove a claim wrong. Code that runs, sources that check out, novelty tested against the field.

The system finds literature and data outside of its training set. It's not an LLM talking to itself in a closed loop.”

Dynamic retrieval
Live
Literature pulled and indexed at run time. Never a frozen training set.
DOI verification
Per-citation
Every reference resolved to a live source before it enters any output.
Novelty assessment
Corpus-grounded
Newness checked against the published corpus, not the model's prior.
Memory systems
Hours of context
Coherence is maintained across hundreds of agents and hours of compute.
03Scientific agents

Scientific agents that
think like the field.

Dozens of agents orchestrate behind the scenes, built on how scientists actually reason: abstracting patterns, breaking problems down, checking their own work, collaborating on problems, knowing when they're done. A single task - a review, a novelty check, a full study - draws on the whole ensemble in parallel, and adapts at run-time to the task demands.

Our agents were built to do science - the published foundation is Virtuous Machines: Towards Artificial General Science.

A single task · agents spawned

Live decomposition

Orchestrator agentCoordinates
Field-ingest agentReads literature
Research-design agentCritiques design
Methods & statistics agentAudits analysis
Power-analysis sub-agentSpawned · n=12
Interpretive-rigor agentChecks claims
Citation-health agentDOI verify · novelty
Adversarial agentWrites case against
Synthesis agentFinal output
What it looks like end-to-end

Five phases, no human in the loop.

A single agentic system runs all five cells. The pipeline forks at Phase 03: either the system builds an instrument and collects fresh data, or it interrogates a dataset you already hold.

Input

A research domain

A field, a question space, a corpusNo pre-specified hypothesis

Phase 01

Question

Literature readHypothesis generationNovelty & feasibility

Phase 02

Design

OperationalisationPower analysisProtocol draft

Phase 03

Collect or interrogate

Build instrument & deployorRead your existing data

Phase 04

Analyse

Pipeline developmentStatistical modellingRobustness checks

Phase 05

Write

Figures & tablesManuscript proseInternal review

Output

A submittable paper

Manuscript · figures · codeDOI-verified citations
For more on autonomous findings, see research
The score

Calibre. Zero to one hundred.

Every manuscript review ends with a Calibre score, carried as a tier chip, but the number is only the entry point. Beneath it, the review lays out the reasoning criterion-by-criterion: what holds up, what doesn't, and why.

The issues are ranked by criticality, so the biggest problems surface first. Each comes with a specific, actionable fix, turning the score into a roadmap for the next revision. Address the top issues, and the quality lifts measurably.

Read the scoring methodology →

Platinum85-100Rigor at the level demanding venues expect.
Gold70-84Strong work; the review's findings are refinements.
Silver60-69Sound core, with named gaps to close before submission.
Bronze50-59Substantive revision needed; the review says where.
Slate30-49Several major issues; treat the review as a redraft map.
Flaggedbelow 30Fundamental problems, documented candidly.

That's how. Here's what.

no card required

How it works · Explore Science