Anatomy of a RISE pipeline

The landing-page diagram is a coarse model. This page expands each block into its operational components and points to the projects in the catalog that realize them concretely.

Inputs

A RISE system is specified in part by what it accepts as a starting point. Catalog entries declare an inputs: field; common values:

• Human idea — an underspecified prompt from a researcher ("something about momentum in crypto markets").
• Agentic idea — an idea produced by an upstream ideation loop; characteristic of systems that chain ideation and execution (e.g., the Sakana AI Scientist lineage).
• Research question — a fully specified, identification-aware RQ with a stated population, treatment, outcome, and design.
• Replication target — a published paper whose results the pipeline must reproduce (social-science-replicability).
• Submitted paper — for review-focused systems (ape, reviewer, marg).

The breadth of supported inputs is one of the eleven evaluation dimensions; an input-narrow system can still be excellent at what it does.

Knowledge Production (the core)

The agentic pipeline itself. Stages defined in projects/VOCABULARY.md:

Stage	What happens	Reference systems
rq-formulation	Sharpen an underspecified prompt into a research question.	e2er, agent-laboratory
hypothesis-generation	Generate testable hypotheses.	sakana-ai-scientist, robin, agent-laboratory
literature-discovery	Find relevant papers.	storm, paper-qa, open-scholar, gpt-researcher
literature-synthesis	Summarize, integrate, identify gaps.	storm, paper-qa, open-scholar
research-design	Choose method, identification, data plan.	e2er, agent-laboratory, robin
data-acquisition	Fetch, scrape, request data.	e2er, social-science-replicability
data-analysis	Clean, explore, estimate.	e2er, agent-laboratory, social-science-replicability
formal-modeling	Theoretical / mathematical models.	e2er
code-generation	Produce analysis or replication code.	agent-laboratory, sakana-ai-scientist
paper-drafting	First-draft paper sections.	zeropaper, agent-laboratory, sakana-ai-scientist
revision-editing	Refine prose, structure, argumentation.	refine-ink, coarse-ink
referee-simulation	Pre-submission peer review.	ape, reviewer, marg
replication	Reproduce a published paper's results.	social-science-replicability
dissemination	Format for journals/preprints.	(none yet in catalog)

Implementations differ on three orthogonal axes:

• Autonomy — copilot, supervised agent, autonomous, or society-of-agents. The evaluation rubric scores this 0–3.
• Topology — directed acyclic graph, iterative loop, debate / consensus, or hybrid. Tagged in architectural_features.
• Stage coverage — single-stage tool ↔ end-to-end pipeline. Tagged via pipeline_stages and headline focus.

Data layer (left input)

The empirical material the pipeline operates on. Sources are heterogeneous and per-domain: market data feeds, administrative records, biological-sequence databases, user-supplied PDFs.

See Data layer for the typology and access-pattern discussion.

Knowledge layer (right input)

Prior scholarship: literature, citations, theory, reusable methods. The catalog's literature-focused projects (paper-qa, open-scholar, storm) live primarily in this layer.

See Knowledge layer.

Outputs

A RISE system is characterized by what it produces durably. Catalog entries declare an outputs: field; common values:

• Artifacts — code, figures, tables, prompts, intermediate notes.
• Papers / preprints — the most public output type.
• Datasets — curated, often a by-product of the pipeline.
• Replication reports — structured comparison of pipeline output vs. a published target.
• Referee reports — structured peer reviews.

Output reproducibility is scored on the evaluation rubric: a system that drops outputs to disk in a versioned, re-runnable form scores higher than one whose terminal artifact is ephemeral.

What is not in this diagram (yet)

Three blocks are conspicuously absent and likely need explicit treatment in a future revision:

• Submission and venue routing. Where do papers go, and how is that decision made?
• Ethics and IRB-equivalent gates. For studies involving human subjects, animal subjects, or sensitive data.
• Post-publication revision and retraction. How an output is maintained over time.

These belong in the diagram but no project in the current catalog implements them. Pull requests welcome.