Strands Decider 2B is a 2‑billion‑parameter open‑source decision model built by stripping the LM head from a Qwen3.5‑2B backbone and replacing it with a ~1 M‑parameter pointer head that scores each candidate answer against the hidden state at an ` ` token; the torso is fine‑tuned with a rank‑16 LoRA adapter. The release includes full training scripts, data, and model weights on GitHub and Hugging Face, targeting local CPU/GPU experimentation. Evaluation on the public JevBench set shows the model ranks 3rd of 33 in the 2 B class and 1st of 30 when excluding just‑over‑2 B models, achieving 100 % correctness on easy tasks and competitive Brier‑score calibration that improves across iterative versions (v19). Latency is roughly 115 ms on an RTX 3090 and 153 ms on an M3 MacBook for small‑size prompts, scaling linearly with token count. The design emphasizes fast, low‑latency decisions with reliable confidence scores, enabling use cases such as model routing, tool selection, guardrails, policy classification, and hybrid agents that combine LLMs for complex reasoning with Decider for routine binary choices. An example CLI and an InterventionHandler integration demonstrate how the model can intercept tool calls to verify grounding and timing, returning Proceed/Deny/Confirm/Guide actions. The project aims to lower barriers for research and deployment of decision‑model‑based agentic workflows.