Open General Decision Model
Amos serves as 'System 1' for fast-thinking decisions (built on a bidirectional base with ultra-fast forward passes that output calibrated confidence), working with 'System 2' deep-reasoning models to drive AI toward an efficient, reliable productivity loop.
Generative LLMs answer a question by sampling across countless candidate text branches: the more answers they produce, the harder truth is to tell, and the less predictable cost and latency become. Production systems never needed more candidates — they need the single, executable, accountable answer.
Unfork believes the endpoint of intelligence is not more text but the one answer. Amos runs constrained systemone-protocol output on a generative base, collapsing millions of possibilities into one decision (≈1s public e2e, zero errors across 2,000 real calls) with probabilities and confidence — the engineering expression of 'The World Needs One Answer'.
'The World Needs One Answer' is delivered by a dual-system design: System 1 (Amos) outputs the single answer in a single inference pass, resolving high-throughput routine decisions; upon low confidence, high risk, or unknown territory, it escalates to System 2 (reasoning models or humans) for deeper verification — yet every path converges into the same executable conclusion.
"Exhaust every possibility. Deliver one answer."
— Unfork Research Manifesto · The World Needs One Answer01Convergence First
Amos defines every task as a structured decision problem, outputting argmax and a probability distribution in a single forward pass — converging countless possibilities into one answer.
02Honest Uncertainty
Every answer ships with a temperature-scaled calibrated confidence; below threshold, the context escalates to System 2 for verification, so every conclusion is grounded.
03Boundaries Are Promises
The model card lists both intended and out-of-scope uses as they are: decision-only, 896-token context, three-layer OOD disclosure — capability boundaries are clear and checkable.
04Data Before Narrative
2,000 real calls across four independent-seed blind rounds, 17-category details, and the full known-weakness registry are all public, presenting the model's true level through reproducible data.
Qwen3-8B generative base, QLoRA adapter, choice / score / noul primitives, and scope bounds.
Learn more → 02 / RESEARCH97.45% four-round blind-test mean, 96.6% #1 on the three-model shared board, 17-category details, and honest weaknesses.
Learn more → 03 / LIVE DEMOExperience real-time convergence across ticket routing, moderation, and trading commands.
Learn more → 04 / DOCSSix standalone sections: quickstart, model card digest, public benchmarks, API reference, private deployment, and about us.
Learn more →