# MIST Stack

> Eval-driven infrastructure for AI systems. MatchSpec, InferMux, SchemaFlux, TokenTrace.

## About

The MIST stack is eval-driven infrastructure for AI systems. Four Go tools, zero external dependencies, one universal message protocol.

MIST stands for MatchSpec, InferMux, SchemaFlux, TokenTrace.

## Tools

- MatchSpec: Define correctness as code. Run eval suites with statistical rigor. Gate deployments on passing thresholds.
- InferMux: Route inference requests across providers. Load balance, failover, and cost-optimize model calls.
- SchemaFlux: Compile structured data from any source. Schema validation, entity extraction, batch processing.
- TokenTrace: Trace every token. Track cost, latency, and quality across your entire AI stack. Alert on regressions.


## Use Cases

- RL Environments: Reward hacking, reproducibility crises, and the debugging abyss. Why RL training fails and what's missing from the toolchain.
- Model Harnesses: Bad data, silent failures, and catastrophic forgetting. The real problems behind fine-tuning that tutorials skip.
- AI Agents: Cascading errors, context rot, and agents that ignore instructions. What the benchmarks don't show about production agents.


## Links

- Site: https://miststack.dev
- Repo: https://github.com/greynewell/mist-go
- Methodology: https://evaldriven.org
- Author: https://greynewell.com

## Pages

- MIST Stack: https://miststack.dev/
- AI Agents: https://miststack.dev/pillars/ai-agents/
- Model Harnesses: https://miststack.dev/pillars/model-harnesses/
- RL Environments: https://miststack.dev/pillars/rl-environments/