> ## Documentation Index
> Fetch the complete documentation index at: https://docs.mindlab.io/llms.txt
> Use this file to discover all available pages before exploring further.

# Advanced Aggregation

> Ensuring fairness and quality with speculative decoding, procedural verifiers, and Best-of-N sampling.

To further amplify the efficiency and reliability of the MindLab platform, the Orchestrator employs a range of advanced, market-validated techniques for aggregating and verifying the work of its agentic workforce.

### Speculative Decoding

In this workflow, the Orchestrator first routes a task to a fast, lightweight "draft" agent (e.g., a highly quantized 7B parameter model). This agent generates a candidate sequence of actions or text at extremely low latency. This draft is then passed to a more powerful, but slower, "verifier" agent (e.g., a GPT-4o class model).

The verifier's task is not to generate from scratch, but simply to check and, if necessary, correct the draft. Because verification is a computationally simpler task than generation, this can dramatically reduce the end-to-end latency and cost of the entire workflow—often doubling the effective tokens per second—while maintaining the quality standard of the larger model.

### Procedural Verifiers

For tasks in domains with deterministic outputs, such as code generation or financial calculations, the Orchestrator uses **procedural verifiers**. A code-writing agent's output can be passed to a compiler and a suite of unit tests. A financial agent's calculations can be checked by a traditional calculator tool. This "generate then verify" loop provides a level of reliability that no single generative model can currently achieve on its own.

### Best-of-N Sampling

For critical or creative tasks, the Orchestrator manages the **"Best-of-N" sampling** process. It can route the same prompt to multiple agents and then use a "judge" agent to select the best response. This is made economically viable by dynamically pruning unpromising reasoning paths early, achieving the quality benefits of a large sample size without the high latency and computational cost.