TECHNICAL DOCUMENTATION

Inspectable methods.
Reproducible work.

The workflow, interfaces, and deployment boundaries behind LabVector’s scientific workspace.

Explore the sample workspace Save sample to your workspace

analysis.config.json

// An explicit, recorded configuration
{
  "action": "train",
  "features": [
    "temperature", "catalyst", "time"
  ],
  "target": "yield",
  "seed": 42,
  "split": "random"
}

LABVECTOR / DOCS

Start with the demonstration study

1. Open the sample project

Explore the read-only synthetic polymer formulation study: 48 observations with temperature, catalyst, time, and yield. Save a copy to your workspace to rerun analyses and export reports. This is an educational demonstration, not laboratory evidence.

2. Inspect and analyze

Open Dataset Library and Data Quality. Select the numeric input variables and response in Statistical Studio. Inspect summaries and correlations before modeling.

3. Train and evaluate

In Model Workbench, choose the dataset, response, features, split method, and seed. Run candidates, inspect the comparison, and evaluate held-out errors and residuals.

4. Propose, version, and report

Use supported continuous factors in Optimization Lab. Review proposed measurements, save an experiment version, and download the report and reproducibility JSON from Reports.

Python computation downloads a pinned WebAssembly scientific runtime on first use. Network access to the configured Pyodide package CDN is required on the first run; initialization and calculation progress are displayed.

DEPLOYMENT ARCHITECTURE

What runs where.

COMPONENT	CURRENT LIVE WORKSPACE	SUPPLIED SERVICE
Interface	React + TypeScript	Same interface and payload contract
Identity	Verified email sign-in + organization roles	Trusted API token; identity forwarded by server
Scientific compute	Python in an isolated Pyodide worker	FastAPI + durable, isolated worker jobs
Scientific libraries	NumPy, SciPy, scikit-learn	Same Python kernel and recorded versions
Persistence	SQLite, organization scoped	PostgreSQL queue, results, and fitted model artifacts
Status	Deployed with this Site	Source supplied; separate deployment required

A browser-computed result is recorded with its origin. Download the scientific service source below to inspect its schema, deployment configuration, and tests. Durable queued execution, fitted model artifacts, cancellation, and centralized validation require the separately configured compute service. Its source is supplied; a hosted service has not been connected to this Site. Research records are stored in SQLite on the hosting account, outside the public website directory.

Download the Python service source

APPLICATION API

A small, explicit contract.

GET /api/workspace

Authenticated organization, projects, experiments, datasets, runs, members, and history.

POST /api/workspace

Action-based mutations create and edit research records, preserve experiment versions, archive and restore work, manage member access, and save independent reviews, discussions, and notifications. Authorization is enforced server-side.

GET /api/report?runId=…

Download an authorized scientific HTML report containing recorded inputs, methods, results, and reproducibility metadata. It can be printed to PDF.

GET /api/compute · POST /api/compute

Authenticated gateway for an optional separately deployed FastAPI service. The live workspace uses browser Python until its compute URL and service token are configured.

GET /api/billing · POST /api/billing

Read the plan catalogue, recorded usage, subscription state, and plan requests. Administrators can save or withdraw requests. Hosted payment and portal actions are enabled only after the payment service is configured.

DELETE /api/compute?jobId=…

Cancel a server job as its author or an administrator. The compute service also provides organization-scoped job history, fitted-model artifacts, and reproduction bundles.

runScience(action, payload)

Compute actions: analyze, train, predict, optimize, and design. Payloads carry the rows, features, response, and explicit configuration.

Example training configuration

{
  "action": "train",
  "payload": {
    "features": [
      "temperature",
      "catalyst",
      "time"
    ],
    "target": "yield",
    "seed": 42,
    "split": "random",
    "rows": "<dataset observations>"
  }
}

SCIENTIFIC INTEGRITY

Evidence has a scope.

Observed ≠ predicted

Dataset measurements and calculated model estimates have separate views and labels. The sample dataset is explicitly synthetic.

Exploratory ≠ confirmatory

Exploratory plots and candidate designs support investigation. Confirmatory claims need a prespecified design, assumptions, and suitable inferential method.

Association ≠ causation

A correlation or fitted coefficient does not by itself establish an intervention effect. Randomization, controls, and domain knowledge matter.

Reproducible ≠ universally valid

A run preserves its inputs, seed, split, parameters, and versions. Its result still depends on measurement quality and the represented domain.

START WITH A QUESTION

Your next experiment.
A better starting point.

Explore the sample study, then bring your own research.

Explore the sample workspace Open your workspace

Inspectable methods.Reproducible work.