Janakiram MSV writes that a new category of "decision models" has standardized around TypeSafe AI's System One API schema within three weeks of its launch. These lightweight, non-generative models return structured, probability-weighted answers to typed questions rather than prose, functioning as an "if statement" to route requests, gate tool calls, or provide fallback logic for LLM agents. While AWS, Upstage, Ollama, and Perplexity have aligned with the System One contract to allow for portable, inexpensive code-level interoperability, OpenAI has opted out with its own proprietary Decisions API, mirroring the historical standardization pattern of LLM APIs. However, researchers have demonstrated that an adversarial optimizer can easily redirect these models toward wrong answers by adding short, natural-looking context, meaning reliability depends on rigorous calibration and safeguards rather than just API convergence.
- Decision models use only three question types: "choice," "score," and "noul" (yes/no).
- Open-weight implementations often bypass autoregressive decoding by using a small readout head to score options in a single pass.
- AWS's Strands Decider achieves a median response time of 115 milliseconds per question on an RTX 3090.
- A USC study found that adding context of a median length of 31 words can redirect 61.4% of initially correct Jev decisions.
- Decision models use only three question types: "choice," "score," and "noul" (yes/no).
- Open-weight implementations often bypass autoregressive decoding by using a small readout head to score options in a single pass.
- AWS's Strands Decider achieves a median response time of 115 milliseconds per question on an RTX 3090.
- A USC study found that adding context of a median length of 31 words can redirect 61.4% of initially correct Jev decisions.
