podbit-labs

Status: alpha (0.6.1-alpha.1, tracking Podbit releases). APIs, config schemas, database schemas, and prompts may change without notice. Not yet recommended for production use.

Lab subsystems for the Podbit project.

This repository contains:

• lab-core shared infrastructure (@lab/core) used by every lab
• math-lab computational verification of mathematical claims
• nn-lab neural network training and experiment execution
• critique-lab LLM based critique of knowledge graph nodes

Installation

lab-core must be built first because every lab depends on it via file:../lab-core.

cd lab-core
npm install
npm run build

cd ../math-lab
npm install

cd ../nn-lab
npm install

cd ../critique-lab
npm install

Each lab can then be run with npm run dev (tsx) or npm run build && npm start.

Python

Python is required by every lab. lab-core ships executor.py, the sandbox process that runs generated code on behalf of the labs. Each lab's config.json points at the interpreter via sandbox.pythonPath (default python).

Per lab Python requirements:

• math-lab needs a substantial scientific Python stack: numpy, scipy (with optimize, integrate, special, stats, interpolate, linalg, signal, fft, sparse), sympy, mpmath, networkx, and numba (JIT compilation for tight loops, Matsubara sums, Monte Carlo). All spec types in math-lab/config.json (math, parameter sweep, convergence analysis, curve shape, quantum model, wave system, coupled dynamics, many body model) assume this stack is present
• nn-lab needs torch (with CUDA if you want GPU training), torchvision (lazy imported by nn-lab/runner/datasets.py for MNIST, FashionMNIST, CIFAR10 loaders), numpy, and matplotlib (lazy imported by nn-lab/runner/charts.py with the Agg backend for headless chart generation). The full runner lives in nn-lab/runner/
• critique-lab is LLM only and does not execute generated code, but still resolves pythonPath at startup so a working python on PATH is expected

See the Podbit repo for the recommended Python environment and version pinning.

Setup

These labs are not standalone. They are designed to run alongside a Podbit instance, which owns model assignment, semaphore based concurrency, rate limit cooldowns, and budget tracking. Each lab routes its LLM calls through Podbit via the podbit.url and podbit.subsystem fields in its config.json.

For installation, configuration, model registry setup, and how to wire each lab into Podbit, see the main Podbit repository. Start there before attempting to run anything in this repo.

Configuration

Each lab ships a config.example.json with documentation comments and a sanitized config.json with placeholder model blocks. Real API keys belong in .env files (gitignored) or in Podbit's model registry, never in config.json.