auto: Instructions file for Claude Code

CLAUDE.md, Auto (the AGI compiler)

read this fully before touching code. it records the vision, objects, pipeline, stack, norms, and guardrails we hold this project to. everyone working on the repo follows it.

what this is

Auto converts interpreted cognition into compiled cognition. today an agent run re-derives its plan token by token on a frontier model: brilliant, expensive, unbounded, unauditable. Auto records that behavior, proves what is secretly symbolic, distills what is not, verifies the result against a behavioral contract, and emits a cognition binary that is bounded, reproducible, capability-confined, roughly 1000x cheaper and 100x faster. speech, then writing, then computation, then cognition: the fourth compilation event. the interpreter is model-agnostic by design. it is a frontier LLM today, but nothing in the pipeline assumes that, so the same loop compiles the behavior of whatever sits at tier-0 next, world models included. the end state we are building toward: contracts as executable institutions, the IR as the shared language of cognitive work, binaries as the commodity form of expertise, and a tiered runtime where nothing is ever figured out twice.

the three objects

1. auto, the compiler. cli + service. auto record (attach to a live agent, capture traces) to auto compile --contract (lower to IR, run passes, verify) to task.cbin.
2. the runtime. tiered execution. tier-1: compiled fast path. tier-0: frontier model as interpreter, for novelty. guards on every compiled entry; guard trip to deopt to tier-0 to capture trace to recompile. the ratchet: every novel solve compiles back down. nothing figured out twice.
3. the artifact + registry. .cbin = code + small models + kernels + manifest (eval scores, cost/latency bounds, capability requirements, full provenance: traces, reference model, eval run ids). content-addressed, sigstore-signed. the manifest is the trust layer, it is never aspirational, only measured.