Building the AI Scientist

Architecture

The scientific harness, opened up

The model supplies intelligence. The harness plans the investigation, selects methods, acquires context, preserves state, evaluates quality, involves people and produces a reviewable work product. Effectiveness depends on how coherently these systems exchange context, confidence and control.

Layers: 9
Capabilities: 45
Reusable across loops: 26
Cross-cutting layers: 3

fig. 01 — the same nine components, four ways— / 09

Above — composition, exchange, coordination. Nine components, assembled as a pipeline, recomposed as a cycle so evaluation can send work back, given a gate that fires only when its inputs agree, then given the two spines that touch every stage. The last arrangement is the diagram below.

fig. 02 — the architecture, layer by layer

Open a layer and ask what happens without it

The nine layers are drawn as a cycle rather than a pipeline, because planning, execution, evidence and evaluation feed back into planning; memory and human judgment are vertical spines because they are not stages. Every capability states what it does, why it is necessary, and what a system does wrong when it is missing — and which class it belongs to, since that is what explains how a general AI scientist emerges from several validated vertical loops rather than one generic agent.

Flow

Feedback

Cross-cutting

Stopping gate

Level 1

fig. 02 — the scientific harnessclick a layer to open it

Memory and human judgment are drawn as vertical spines because they are not stages — they touch the investigation throughout. Evaluation returns work to execution, evidence and planning as often as it advances it. The moving traffic is the direction work flows; brand-colored traffic is evaluation sending it back.

Level 1 · The whole system

Nine layers, one calibrated architecture.

The model supplies intelligence. Everything in the diagram is the harness that wields it — planning the investigation, selecting methods, acquiring context, preserving state, evaluating quality, involving people, and producing a reviewable work product.

Work loops, it does not flow

Planning, execution, evidence and evaluation form a cycle. Evaluation sends work back as often as it sends it forward.

Two spines cut across everything

Memory and human judgment are not stages. They touch the investigation from the objective to the artifact.

One gate to the output

Every work product must pass a scientific stopping decision.

Select any layer to go deeper. You can understand the architecture without opening every capability.

Four kinds of capability, deliberately distinguished

This distinction is the strategic core of the architecture. It explains how a general AI scientist can emerge from several validated vertical work loops rather than from one generic autonomous agent.

Generic AI scientist capability: Reusable infrastructure. Built once, used by every work loop — runtime, memory, provenance, orchestration.
Scientific skill: A validated scientific method, reusable across related scientific problems. Versioned and testable, like a library function for science.
Loop-specific configuration: Unique to one work loop. The recipe, thresholds and rubric that make omics-to-target different from safety-signal investigation.
Organization-specific context: Your therapeutic strategy, internal datasets, platform capabilities, prior failures and scientific preferences. Cannot be bought off the shelf.

in depth · scientific evaluation

Method selection is where the architecture is tested

Two layers make a claim that can be measured: planning selects the method, and evaluation scores how applicable that method is to the data in front of it. On one high-volume task the consensus method holds 92.5% of the literature while the context-appropriate choice rests on five verified papers — and a top-50 retrieval window contains none of them. The measurement, the instrument that flips the choice on a real dataset, and the capabilities that carry it are on their own page.

head share: 92.5%
verified advocacy: 5
in the top 50: 0

Open method selection

See the architecture doing real work

An architecture diagram is a claim. The work-loop explorer shows these layers applied to one scientific problem, stage by stage.

Open the work loop