stochadex

---

title: "Home"

So what is the 'stochadex' project?

It's a simulation engine written in Go which can be used to sample from, and learn computational models for, a whole 'Pokédex' of possible real-world systems entirely in YAML.

The framework abstracts away the machinery that sampling algorithms have in common behind a single configurable interface; so a whole simulation, the analysis and inference layered on top of it can be stated as pure configuration.

This simulation engine is designed based on the simulation software fundamentals described in this collection of blog posts.

When to use it

The stochadex fits best when you want stochastic simulation, online inference and simulation-based decision-making over one composable primitive and you want a whole run to be a single config file.

The model, its wiring, the run mode, the data in/out and any inference or optimisation on top are all data: one YAML file, run by one prebuilt binary, with no Go toolchain anywhere in the loop. Maths outside the built-in catalogue can be written as expressions in the same file, so Go is only for genuinely new primitives.

Other declarative formats are narrower (SBML, Modelica) or are languages with their own compilers (Stan, GAML). However, you should probably reach for something else when:

• Large fixed-shape, GPU, or autodiff-heavy Bayesian modelling → Stan, PyMC, or Julia's SciML.
• Pure discrete-event simulation (entities through queues and servers) → godes or SimPy.
• A standards-based interchange format (systems biology, physical plant) → SBML with COPASI, or Modelica.
• Plain numerics or classical ML in Go → gonum, which the stochadex is built on.
• Training neural networks or deep RL → train in Python, then import a frozen ONNX model to run inference behind a OnnxInferenceIteration.

Install

Four ways in, depending on whether you're writing YAML, running it as a pipeline step, letting an agent write it for you, or writing Go.

As a CLI → describe a whole run in one YAML file and execute it with a prebuilt binary. A config that names no Go anywhere runs in-process, so no Go toolchain is needed. See running from a config file.

curl -L "https://github.com/umbralcalc/stochadex/releases/latest/download/stochadex-$(uname -s | tr '[:upper:]' '[:lower:]')-$(uname -m | sed 's/x86_64/amd64/;s/aarch64/arm64/')" -o stochadex
chmod +x stochadex

As a container → the same YAML surface, with every integration already built in (Arrow, Postgres, S3, DuckDB, accelerated BLAS) and nothing to install. See running from a config file.

docker pull ghcr.io/umbralcalc/stochadex:latest

With an agent → using stochadex from an AI agent is robert's job; see its documentation. It drives the same CLI, with --check, inspect --io, graph and catalogue as its machine-readable contract, and the engine itself stays deterministic and agent-free.

As a Go library → implement the Iteration interface to add a primitive the catalogue doesn't have, or embed the engine in your own service. Start with the quickstart.

go get github.com/umbralcalc/stochadex

Integrations

Integration	What it does	Where
	Load state history into a simulation and write output back over database/sql. Point it at any Postgres-wire database.	read · write
	Build simulation output directly as Apache Arrow for columnar interchange (Polars / pandas / Parquet). Opt-in module.	read · write
	Land the Arrow output in DuckDB for SQL analytics, zero-copy. Opt-in module.	write
	Read and write runs to Amazon S3 or any S3-compatible store (MinIO, Cloudflare R2, Ceph). Opt-in module.	read · write
	Run a frozen ONNX model trained upstream in Python (with sklearn, XGBoost or a small neural net). Opt-in module.	run