Microcosm
This page

Getting started

Overview

Microcosm is a runnable public inspection surface for an AI-native workflow system. You run bounded components locally and read what they produced; every component declares the evidence behind it and a plain boundary showing where its scope stops.

What it is

Microcosm is the part of a larger working system that can stand on its own in public: components that are runnable, readable, and checkable against their own source. The rest stays private: parts of it are live and handle private data, so the whole thing is not online.

It is not a hosted service and makes no external model calls. When you run it locally, it writes its state into a folder beside your project so you can follow what it did and trace each step back to the source behind it.

What you can do here

  • Run one local witness by cloning the repository and running the first component against a folder.
  • Read the result record in the local .microcosm/ state, then open the component card it points to.
  • Check the evidence line: evidence kind, independence rank, real-tool marker, and the scope limit.
  • Open the whole-system map after the first loop is concrete.
  • Browse the seven areas, the system's components grouped by what they do.
  • Open the source, and follow any page back to the files behind it.

The seven areas

There are 78 components in total. That's a lot to meet at once, so they're grouped into seven areas; the one that matches why you came is the way in:

  • Entry & orientation. How a newcomer first meets Microcosm and follows a short guided path through it.
  • Architecture & navigation. The core primitives, pattern rules, and routing that give the system its shape.
  • Formal math & proof. Inspectable pieces of a proof pipeline: premise retrieval over a copied Lean Std index, tactic routing, verifier-trace repair, and claim-separation result records. Three components run the real Lean/Lake prover locally on bounded examples; the rest publish the pipeline's checking layers as contracts you can open.
  • Agent reliability & safety. Source-open replays of agent failure modes as inspectable specimens.
  • Research & science. Replays that stand in for scientific and forecasting workflows, run over synthetic fixtures.
  • Import & drift control. The boundary that brings non-secret material into the public tree and flags drift.
  • Work & continuity. How reversible work is recorded, landing decisions are made, and runs resume.

Next steps

Quickstart →Clone and run the first local witness. What it is, in full →A guided read after the evidence loop is concrete. What the pass proves →Evidence kinds, ranks, real-tool markers, and boundaries. Whole-system map →The rendered map and scene packet behind the architecture. Components →Browse the generated component index.
Open on GitHub →