AmberContinuum

Maxx is a working prototype of a deterministic safety gate for AI agents, built for the Meta Global AI Developer Hackathon. Inspired by Maxwell's demon, Maxx sits outside the agent, between a proposed action and tool execution. It is designed for consequential everyday actions — paying bills, sending messages, sharing files, booking travel, posting publicly, and deleting files — where an error may be costly or irreversible. Each proposal is represented in a fixed schema — action type, trace, amount, and reversibility — and scored with explicit, inspectable arithmetic: floor 15 + trace weights + 15 if irreversible, capped at 98. Maxx then issues one of four decisions: release, wrap, hold, or escalate. A separate sidecar enforces the decision through one-use permits for releases, corrected substitutions for wraps, and no execution for holds or escalations. Every decision and its plain-language reason are recorded in a ledger whose counts must reconcile. The seeded demonstration covers eight actions and yields 1 release, 3 wraps, 2 holds, and 2 escalations. Irreversible executions fall from 6 to 0, $1,090 is held for human decision, and operator attention falls from 8 to 7. The prototype uses mocked tools and does not execute real-world actions. Muse Spark 1.3 and Muse Wonder 1.0 were used through the Meta Model API in a documented live run to propose and review actions; the gate itself is deterministic and calls no model. Muse Glimmer 30B was used for a separate Petermann-K analysis, whose checks did not outperform conventional baselines and whose output is not consumed by the gate. The project also includes a sorter that tests whether trace-based sorting is informative on a held-out variable, using stated falsification criteria. Maxx does not claim to solve AI governance. It provides a measurable checkpoint for studying declared-trace verification, threshold calibration, and production-grade enforcement at the point of action.