API reference

These pages are generated from the installed package on every change, so a signature here is the signature the interpreter sees. Each name shows the signature with its defaults, the file that defines it, and its docstring. When to use which call, in the order a run happens, is on the five calls; every knob with its default is on parameters; how a page is laid out is on design.

Page	Names	What it holds
whileai	29	The platform client: sign in, push traces, read the account.
data	6	SimulationData: the object simulate() returns, and what it can do next.
score	99	Grading, judge checks, pass@k, selection for SFT and RL, hack scans.
generate	18	Situation writers, coverage axes, agent adapters, model backends.
world	2	The mock world that answers tool calls and fails on schedule.
verify	3	Verifiable rewards: programmatic checks a judge cannot game.
constants	7	Module-level constants and their defaults.
environment	3	A verifiers environment for on-policy trainers.
export	7	JSONL, preference pairs, and RL environments out of graded rows.
ingest	39	Platform calls: datasets, traces, cuts, training runs, hosted models.
monitor	2	Watch a live training run for reward hacking.
run	1	The search loop that spends the budget.
schema	10	The typed row schema and its conversions.
simulation	1	simulate(): the entry point, and every argument it takes.
training	14	Training runs, trainer callbacks, TRL export.