Fairwind

Fairwind routes every incoming task through a cost-ordered pipeline, escalating only when a cheaper tier can't confidently answer: Deterministic solvers (0 tokens) — bare arithmetic, and a constraint solver for row/adjacency-style logic puzzles (brute-forces permutations, commits only when exactly one satisfies every clue). Local models (0 tokens, confidence-gated) — Qwen3-4B-Instruct-2507 handles factual/sentiment/NER/summarization; a dedicated Qwen2.5-Coder-1.5B-Instruct handles code_gen/code_debug. Every local answer must clear a calibration curve fitted from 191 labelled examples pulled from public eval datasets (TriviaQA, CoNLL-2003, tweet_eval, HumanEval, MBPP, CNN/DailyMail) plus hand-written cases — self-consistency sampling (1 deterministic + 2 stochastic generations) produces a category-appropriate agreement signal, looked up against the fitted curve, and only trusted above an 85% reliability bar. Categories that can never clear that bar (factual, sentiment) skip the local attempt entirely rather than waste time on a guaranteed rejection. Fireworks AI (batched, parallelized) — everything the above can't confidently answer, merged into as few calls as possible per model, with automatic fallback to a different allowed model if one fails, <think>-tag stripping for reasoning models, and a truncation-retry safety net.