Maestro Order harness turns unreliable LLMs into reliable systems with 4 primitives

A new academic paper from researcher Hidayet Aksu proposes Maestro Order, a model-agnostic orchestration harness designed to transform unreliable large language models (LLMs) into reliable, predictable systems. The core insight is that a single forward pass of a capable model is fast and fluent but dangerously unreliable—producing confident errors known as hallucinations. Maestro Order addresses this by composing base solvers according to four structural primitives: decompose (break problems into subtasks), ensemble (combine multiple solvers), verify (check outputs with verifiers), and recurse (repeatedly refine). A budget-aware controller dynamically decides where to allocate compute, prioritizing stages with the highest marginal reliability per unit cost. The harness treats any model as a black box behind a uniform interface, layers a verifier ensemble whose discrimination is measured online, and ensures the system is deterministic, observable, and fault-tolerant.

The paper reports results from a faithful Monte Carlo simulation over a parameterized solver/verifier model, confirming predicted laws quantitatively. Verification amplifies reliability geometrically: for instance, a base solver with 0.55 accuracy jumps to 0.98 with two verification gates and to 0.999 with four. Voting, in contrast, only helps when individual solvers perform above chance, and its gains are limited by shared errors. The budget-aware controller reaches a target reliability at a small fraction of the cost of voting alone by selecting the cheapest mechanism for each regime. Aksu also discusses failure modes including verifier gaming, correlated errors, and decomposition error compounding, offering concrete guidance: build robust checkers, diversify solvers, and let the controller put compute where the information is. This framework could significantly reduce hallucinations in production AI systems without requiring model retraining.