1.17.0 — Meta-Agentic Tree Search Lab

MATS now provides an inspectable, reproducible branching search over workflow designs. Named bounded rewrite operators change stage policies; fixed-point UCT balances exploration with observed utility. The simulator schedules a dependency graph against finite resource pools, propagates modeled defects and charges for detected rework. Separate held-out workloads challenge the frozen candidate.

Four editable cases cover software delivery, API migration, public-data reproducibility and a deliberately unsafe shortcut objective. An optional exhaustive audit reports the gap to the training-model optimum without selecting the candidate. Five review gates keep failed proposals from being marked review-ready. Every proposal remains unapproved, and the active workflow remains the baseline.

A responsive browser lab exposes the actual tree, iteration replay, node statistics, policy comparison, resource schedules, exact gates, settings recovery and six-file evidence downloads. Python and JavaScript reproduce complete runs and exports. A standard-library CLI, loopback-only packaged server and executable notebook provide local entry points without providers or runtime downloads.

The preserved integer demo now branches, counts rewards once, selects actual leaves, emits valid CSV and isolates seeded randomness. Optional provider calls are bounded, SDK imports are lazy, and configured search parameters reach the coordinator tool. Original research, diagrams, notebook, browser replay and legacy entry points remain available.

Publication requires exact native/browser parity, independent arithmetic tests, hostile-input rejection, installed-wheel acceptance, accessible browser workflows and commit-bound public-site evidence. The lab optimizes synthetic assumptions; it does not perform customer work or authorize deployment.

Use Start here for verified installation and the MATS operating guide for first results and recovery.