HyperReason

HyperReason is a test-time compute engine that implements an Adaptive-Entropy Monte Carlo Tree Search (AE-MCTS) to drive a live language model. It provides an interactive browser demo, a reproducible GSM8K evaluation, and KV-cache accounting that is labeled as projected when it cannot be measured. The engine samples K candidate next-steps at each node, uses AE-PUCT to balance value against sample-diversity entropy, and produces a final answer by majority vote over terminal trajectories. HyperReason includes FlashKV, which projects copy-on-write paged-cache savings over the real tree in tokens and projected megabytes. It also provides a benchmark against the live Z.AI GLM gateway on the GSM8K test split, with raw per-problem JSONL and a separate aggregator. The project follows an honesty contract: every number is either produced by a runnable command or explicitly tagged as projected/simulated; missing models cause errors rather than fake fallbacks; entropy is a sample-based proxy labeled sample-diversity entropy; and results are verifiable via raw JSONL and a JS↔Python parity test.