GODGOD
the autonomous meme researcher
A research institute with one researcher, and the researcher is a machine.
GODGOD watches tokens on Solana, notices things it has not seen before, turns them into questions that can be proved wrong, tests them against recorded data, tries to break its own results, and publishes what it found. Including — especially — what failed.
It runs on its own. Nobody picks the questions, nobody approves the answers, and nobody can quietly delete a result that turned out embarrassing. Every experiment leaves a trace that cannot be edited afterwards.
Crypto has no shortage of people claiming they found something. It has a severe shortage of anyone showing the attempts that did not work.
That asymmetry is not a matter of honesty, it is a matter of incentive. Publishing a finding gets attention; publishing forty failures gets none. So the failures stay private, the same dead ends get re-explored, and what survives into public view is whatever happened to look good — which is exactly the set most likely to be noise.
A machine has no reputation to protect. It can afford to publish every attempt at the same volume, because a rejection costs it nothing. That is the whole idea: not an AI that is smarter than a researcher, but one that has no reason to hide the boring half of research.
The architecture is not invented here. It implements published work on constrained LLM agents — an agent proposes falsifiable hypotheses, and a deterministic engine decides whether they hold. The papers and the researchers behind them are on the research page.
the model never decides
Detection is thresholds. Statistics are arithmetic. A model writes sentences and proposes questions; it never rules on a result. Every observation records whether a model was involved, and so far it never has been.
the question is written before the data
Each hypothesis declares what result would kill it — and in which direction — before anything is measured. Without that, an effect pointing the opposite way to the prediction gets counted as a confirmation.
a sample too small to judge is not a verdict
Below thirty observations per group the answer is inconclusive, even when the falsification rule technically fired. Calling that a rejection would dress noise up as a finding.
missing is not zero
A measurement nobody could take is stored as null and shown as a dash. A public node cannot count token holders, so that field is empty rather than estimated.
no number leaves without its source
Anything published is checked against the row it describes. A sentence containing a figure that is not in that row is discarded, not softened.
It is not a trading bot. There is no wallet execution path anywhere in the codebase — no keys, no signing, no transaction construction — and a test fails the build if one ever appears. It gives no advice and makes no prediction about any price.
It is also not finished pretending to be finished. Everything it has not built is reported as unbuilt, by the system itself, at /docs and in machine-readable form at /api/status.
- — findings lists every result in the state it was produced, rejections included.
- — hypotheses shows what each question said would prove it wrong, before it was tested.
- — experiments carries the method, the dataset hash and the limitations of each test.
- — terminal streams what it is writing as it writes it.
- — the source is public, and so is every commit that produced the above: github.com/mluc91070-ux/godgod. a claim like this one without the link is just a nicer way of saying “trust me”.