How a hypothesis becomes a verdict at ArchLab.
A hypothesis is turned into one falsifiable claim, and prior work is searched so we do not re-test what is already known.
Predictions, thresholds, seeds, controls and the exact decision script are fixed before any code exists. A cryptographic commitment to them is published to a public, append-only ledger and timestamped by two independent RFC 3161 authorities and by OpenTimestamps (Bitcoin). Git dates are not used as evidence.
An agent builds the experiment without seeing the predicted outcome. Before anything runs at scale, mechanical checks shuffle the targets and zero the inputs; if the metric does not move, the instrument is declared suspect and nothing is run. The code and container are then sealed a second time.
Compute only accepts jobs whose code matches the second seal. Every launch is logged in an append-only, hash-chained log; seeds cannot be re-run until they pass. The verdict comes from the sealed script.
Two reviewers from different model providers try to break the result. They are first tested on an experiment with a planted bug; a reviewer that misses it carries no weight.
Supported every sealed prediction held. Refuted at least one did not. Inconclusive too few valid runs. Instrument suspect the experiment itself failed its checks.