MVLab — Vibe Researching for Everyone

Turn any question into a Personal Research Lab. Help people, schools and teams plan, investigate, verify and package reproducible research.

Comment

Inspiration

AI is making it possible to explore unfamiliar subjects, read large bodies of literature, write analysis code and test ideas through natural language.

This emerging workflow is sometimes called vibe researching: the human provides the question, intuition and critical judgment, while AI handles much of the operational work.

That accessibility creates a major opportunity, but also a major risk. A fluent answer, polished report or generated experiment can look like research without having a clear hypothesis, sufficient evidence, an accountable owner or a reproducible process.

We were also inspired by the idea of a Minimum Viable Laboratory: instead of building every possible capability, start with the first experiment and identify the smallest set of people, tools, data and infrastructure required to run it correctly.

We combined these ideas into Minimum Lab:

A platform that lets anyone create a small, structured and accountable research laboratory around a serious question.

A Personal Research Lab may belong to an individual, a classroom, a school, an association, a startup, a company or an existing research team.

What it does

Minimum Lab turns a natural-language question into an operational research environment.

A user can start with something as simple as:

“Does shade materially reduce surface temperatures in our school playground?”

The platform creates a Personal Research Lab containing:

  • a precise research question;
  • a Lab Charter defining the mission and scope;
  • testable hypotheses;
  • a first minimal experiment;
  • required controls and quality gates;
  • available and missing capabilities;
  • safety and supervision requirements;
  • claims connected to supporting or contradicting evidence;
  • a transparent record of AI and human contributions;
  • a portable research package.

The product is organized around three core components.

Lab Steward

The Lab Steward is the AI guide for the laboratory.

It helps the user clarify the question, identify assumptions, design the first experiment, search for evidence, expose limitations and decide when specialist or institutional review is required.

The Steward executes work, but the human remains responsible for direction and approval.

Capability Ledger

The Capability Ledger shows what the lab can and cannot currently do.

Capabilities can include:

  • human expertise;
  • datasets;
  • software;
  • AI models;
  • compute;
  • instruments;
  • time;
  • budget;
  • permissions;
  • external reviewers.

The Ledger distinguishes between available capabilities, blockers and optional future additions.

This prevents the AI from pretending that a complete laboratory exists when essential resources or expertise are still missing.

Lab Crate

Every lab can be exported as a structured Lab Crate based on RO-Crate.

The crate contains the research question, hypotheses, sources, experiment plan, artifacts, contributors, software, AI model information, provenance and reuse conditions.

This makes the work portable, inspectable and easier to reproduce outside Minimum Lab.

The Minimum Lab principle

Minimum Lab works backward from the first meaningful experiment.

The goal is not to create the largest laboratory. It is to create the smallest credible laboratory capable of producing the next useful piece of evidence.

How we built it

We built Minimum Lab as a multi-tenant web application using Next.js, TypeScript and Supabase.

Supabase provides:

  • authentication;
  • PostgreSQL data storage;
  • Row Level Security;
  • artifact storage;
  • realtime laboratory updates;
  • structured audit records.

We used the OpenAI Responses API with GPT-5.6 to compile an initial question into a structured laboratory blueprint.

The model produces strict structured outputs for:

  • the Lab Charter;
  • hypotheses;
  • the first experiment;
  • the Capability Ledger;
  • safety classification;
  • quality gates;
  • next actions.

A separate critical-review pass challenges the proposed experiment, searches for unsupported assumptions and identifies missing expertise.

Web search is used to ground factual claims in external sources. Claims without sufficient evidence remain explicitly marked as hypotheses.

We used Codex as our primary coding agent to implement and review the application, database migrations, AI schemas and RO-Crate export workflow.

The Lab Crate exporter generates a machine-readable ro-crate-metadata.json file together with human-readable Markdown documents and project artifacts.

Challenges we faced

Keeping the experience simple without pretending research is simple

The main design challenge was preserving the low-friction experience of vibe researching while introducing enough structure to make the result useful.

We avoided turning the product into a long compliance form. The user begins with a conversation, while the platform progressively generates the required scientific structure.

Turning probabilistic AI output into durable objects

A conversational answer is easy to generate but difficult to audit.

We therefore converted GPT-5.6 outputs into versioned database entities: hypotheses, claims, evidence, capabilities, experiments and approvals.

Distinguishing an investigation from a validated result

Minimum Lab never treats generated content as an automatic scientific conclusion.

Every claim has an explicit status, such as:

  • hypothesis;
  • planned;
  • observed;
  • calculated;
  • contradicted;
  • reproduced;
  • externally reviewed.

Supporting many kinds of laboratory owners

The same product must work for an individual, a classroom and an organization.

We modeled ownership through workspaces, roles and laboratory memberships rather than assuming that every lab belongs to a single consumer account.

Exporting interoperable research artifacts

RO-Crate is flexible, but mapping AI conversations, model versions, evidence and human approvals into a useful research object required deliberate decisions about provenance and structure.

What we learned

We learned that the primary bottleneck in AI-assisted research is not generating more ideas.

The difficult part is turning an idea into:

  • a bounded question;
  • an executable first experiment;
  • an explicit capability requirement;
  • a traceable chain of evidence;
  • a process with a responsible human owner.

We also learned that accessibility and rigor do not have to be opposites.

AI can reduce the amount of specialized knowledge required to begin an investigation, while the product can make uncertainty, missing expertise and validation requirements more visible.

The future of research may involve millions of small, temporary laboratories created by individuals and organizations.

The infrastructure surrounding those laboratories will determine whether they produce useful knowledge or only polished content.

What is next

The next version of Minimum Lab will allow independent laboratories to exchange:

  • evidence;
  • datasets;
  • expertise;
  • compute;
  • experiment templates;
  • replication requests.

Each lab will remain sovereign over its private data while sharing selected artifacts through portable Lab Crates.

Longer term, Minimum Lab could connect Personal Research Labs to universities, public datasets, scientific infrastructure, external laboratories and European research services.

The long-term vision is simple:

Everyone should be able to begin researching. Evidence, replication and accountable review determine what becomes science.

Built With

Share this project:

Updates