bylaw

bylaw: deterministic guardrails for AI agents. Frontier models are powerful, but they are not predictable: the same agent, asked the same thing twice, can take two different actions. In a company that issues refunds, changes bookings or moves money, that is the real blocker. bylaw sits between any agent and its tools. It reads the policy the agent runs on and searches for chains of tool calls that are each allowed but together reach an outcome the policy forbids. It proves each chain by executing it 5 times, drafts an amendment, backtests it against the agent's own reference conversations, and, once a person ratifies it, enforces it on every tool call before it runs. The model proposes. The application decides. Showcase: OpenAI's open-source customer service agents demo (openai-cs-agents-demo, MIT), run on Muse Spark. bylaw found 4 kinds of holes, for example hotel and meal vouchers ($280) issued on a flight that was on time. With the guard on, all 4 attacks were stopped in our recorded runs (one run each), 0 of 5 honest customers were blocked, and the amendments block 0 of 8 reference conversations. A known limit: a rebooking the customer did not ask for is not caught. Muse Spark (Meta Model API) compiles the policy into rules, proposes the attack chains, drafts the amendments, and runs every agent in the showcase. Its output is treated as untrusted: rules run in a sandboxed evaluator, and only human-ratified rules reach the guard. bylaw itself is model-agnostic. Every recorded run is traceable, with tokens and estimated cost: about 704,000 tokens, $1.35–1.80 at Meta's list price.