autopoiesis

A self-evolving LLM paper-trading agent, graded against its own broker evidence. Paper only: the Guardian's limits cannot be weakened by any configuration, and every number on this site names the command that produced it.

No broker account, model or credentials are needed for any of this. Every command below exits 0 from a fresh clone with none of them.

git clone https://github.com/liuyime2/autopoiesis.git && cd autopoiesis
python3 -m venv .venv && . .venv/bin/activate
make install            # pip install -e ".[dev]"
make check              # ruff + mypy + the test suite
make smoke-offline      # config and CLI end to end, no broker contacted
python examples/minimal_cycle.py
make verify             # the full gate: 59 check classes

What the gate covers

59 check classes over the code and the running system, run by 2027 tests across 116 files. It is made to fail on purpose to prove it can: make verify-self-test plants the defects each class exists to catch.

From a package rather than from a clone

pip install autopoiesis (a pre-release, so --pre) gives the console script, the importable package, examples/ and configs/. It does not give the gate, the suite or the make targets: those are a property of the source tree.

Release

make release-check builds the sdist and the wheel, installs the sdist into a clean virtualenv, and runs what the README claims an installed copy can do. It is how an sdist missing examples/ was found.

Reproducing the results

Every figure on this site names its command. make benchmark prints the raw paper-account accounting and an UNKNOWN portfolio comparison; make research-search SPEC=<registered-spec> evaluates validated cached bars under the declared research contract; make audit runs the recorded-defect audit.

Reproducing them needs a paper account and the same bar cache. A fresh clone with no credentials can build, test and gate the project; it cannot reproduce trading results, and no page here claims it can.