Liquid’s D1.2 extracts structured decisions in a single pass. Value extraction returns source positions, parsed values and a probability of finding the value. Below 0.5, the result is absent or null. Parsing does not convert units: “4 percent” becomes 4 plus a percent unit, not 0.04; an unparseable value can still retain its source span.
Multiple choice assigns independent yes probabilities to options, rather than a distribution over mutually exclusive answers. Span and box limits can truncate counts. The documented TypeSafe SDKs drop typed answers and fail on multiple-choice questions, so these answers currently require direct HTTP.
Liquid’s reported calibration errors—0.8 points for Ultra and 1.2 for D1.2—use 22,505 questions answered by all five compared models across thirteen benchmarks. They are not production-wide accuracy rates. Input-only billing is $0.02 or $0.042 per million tokens; extra questions and larger span or box limits add billable input.