Quick start: solvi.build

The question, labelled examples, the promise and — if you have one — a slow path in; System 1 fitted, its guarantee and who answers what it hands over calibrated on examples it did not see, the store wired, and a plain account of every choice out:

import solvi
from solvi.models import llm

s = solvi.build(question, examples, catalog=cat, max_risk=0.02,    # examples: [(state, correct answer)]
                slow=llm(URL, "openai/gpt-oss-120b"), price=(0.037, 0.17), total=solvi.Budget(usd=5),   # optional
                storage="decisions.jsonl")
res = s.ask(state)            # res.answer, res.by ("s1", "s2" or "human"), res.reasons, res.cost
print(s.explain())            # what System 1 is, its signal and promise, who answers each slice, what is not covered
print(s.report())             # what it did, read from the store

System 1 is the catalog's rule for the question if it has one, your own fitted part (learner=), or a head fitted on the facts the catalog computes (System.fit). The signal its guarantee reads is chosen (the act probability, the confidence, or — for a rule — the computed number that best separates right from wrong); a choice among more than two options gets an open-set gate for answers no example shows. A slow path gets a share of the examples to be calibrated on only when System 1 actually hands it something; inputs the open-set gate holds back go to a person. Every piece is the one described further on (System.fit, System.guarantee, solvi.core.guarantees.openset, solvi.core.dispatch), and everything stays replayable. On four tasks of the task stand (Banking77, Abt-Buy, CUAD, RAGTruth) it matched or beat the hand-written setups with every promise kept on eval, in a quarter to half of the code; its eval numbers sat closer to the promised level than theirs (risk 9.3% of 10% on RAGTruth), its drift flag came later, and the slow path was rarely given anything (examples/24_one_entry_point.py runs without a model).