Getting started

What Microcosm is

A plain walk through what Microcosm publishes today: a source-linked public map of a working local system, the components you can inspect from public records, the evidence each one names, and the line where every claim stops.

What it is

Microcosm is the public map of a larger working system. The larger system is AI-native: software where the ideas come from a person and AI agents execute the building and the upkeep, structured so that an agent's work is stored as evidence a separate check can read. Most of that system stays private. What is open here today is the public source slice: component cards, evidence classes, source paths, synthetic-fixture boundaries, scope limits, source files, and pages that say plainly how far each claim goes. The live public repository carries the runnable standalone slice, not the private working root behind it.

Start with the Quickstart or the repository README.md. A coding agent can clone the repository, read AGENTS.md, run the bounded quickstart check, and inspect the .microcosm/ result record. The public slice stays local: it writes beside your files, does not call a model provider, and keeps each claim tied to the source, evidence record, and scope limit that support it.

What a component is, and how to read one →

The public map contains seventy-eight components grouped into seven areas. Each one takes a single job and turns it into a public contract: what the component is for, what evidence class backs it, which source path owns it, and where the claim stops. Some contracts run over bounded public files; others are fixture-bound, card-only, or evidence records. Every component carries a card that says what it does, which class of evidence backs it, the command or route associated with it, and the line that matters most, the thing it refuses to claim. The evidence rank is how independently the check can fail, not a maturity badge: a high rank means the result is hard to fake, and even the strongest check stops at the scope its card declares. One card read this way carries the idea; the full map comes easier after it.

Doctrine →

Underneath all of it is a short set of rules the whole slice obeys, the floor every claim has to meet before it is allowed to mean anything. Recompute a verdict from the evidence underneath rather than trust a label, since an assertion on its own carries no weight. Keep the trust in a small checker you can rerun, the way a proof is only ever as strong as the kernel willing to accept it. Authority comes from proof, policy, and the state of the world, never from a standing account secret or a confident voice. When the evidence is missing, fail closed and stay blocked until something genuinely earns the pass, and never read silence as success. When a computation cannot be done honestly, refuse it with a reason instead of returning a number that looks fine and means nothing. Carry where every value came from, hold a live count to a way of refreshing it, and keep a generated page below the source it was built from. The rule I lean on hardest is the last one: this page meets the same floor as everything else, which is why it can tell you what it cannot prove. The axioms and principles behind these are laid out on the Doctrine page.

The seven areas →

What makes them one system is a shared path most of them plug into rather than calling each other, project -> catalog -> pattern -> standard -> route -> work -> event -> evidence -> explanation -> assimilation.

• Entry and orientation is the front door and the short guided path, honest that it is a walk-through and not the whole system.
• Architecture and navigation is the shape itself, the kernel primitives and routing that every other component binds to.
• Import and drift control is the membrane between this open slice and the private system, bringing in material by manifest and digest, watching for drift when a projection wanders from its source, and never letting a copy outrank the original.
• Work and continuity is how a run actually lands, with landing records, leases, resumes after a pause, and concurrency control so two writers cannot quietly clobber each other and a cheerful done is never mistaken for finished work.
• Agent reliability and safety keeps the ways agents go wrong as specimens you can open, from sandbox escape and prompt injection through memory poisoning, sabotage, tool authority, and benchmark gaming, all as study specimens and not live defenses.
• Formal math and proof is a proof pipeline under glass, where most steps check premises, tactics, and traces, and only a few like the Lean witness compile a tiny public example through Lean.
• Research and science runs the same honesty over synthetic fixtures, showing what replication, forecasting, interpretability, and a finance forecast would each have to carry before anyone should believe them, with the advice boundary kept shut.

Where to start →

If you have ten minutes today, open the overview, run or read the quickstart, then inspect one component card, its evidence line, its source path, and its scope limit. That one loop is the idea in small: a public claim, the evidence record it names, and a clear edge where the claim ends. The site and AI packet are the reading surface; the repository is the source of record.