Skip to content

Methodology

Reusable research methodologies developed and validated through BLIS experiments. These approaches are domain-agnostic — while they were refined on LLM inference serving, the techniques apply to any system with configurable policies and a simulation or benchmark harness.

Methodology Pages

Page Description
Strategy Evolution Iterative policy discovery through simulation: hypothesis-bundle-driven methodology with multi-judge review, convergence-gated verification, Bayesian parameter optimization, and cumulative principle extraction
Hypothesis Bundles in Practice Detailed examples of hypothesis bundles from PR #452 (scheduling) and PR #447 (routing), prediction error analysis, bundle sizing, and writing guidelines

When to Use Strategy Evolution

Strategy Evolution is the right approach when:

  • Your system has multiple interacting policy layers (routing, scheduling, memory, admission) where interactions produce non-obvious emergent behaviors
  • The optimal configuration cannot be derived analytically because layer interactions are too complex
  • You have a deterministic simulator or benchmark that accepts parameterized configuration and produces machine-parseable metrics
  • You need defensible results — not just "it works" but "here's why it works, here's the evidence, and here are the principles"

Relationship to Other Processes

Strategy Evolution integrates existing BLIS processes:

  • Hypothesis Experiments — Each strategy iteration is formulated as a hypothesis bundle. The hypothesis experiment framework provides the per-arm workflow (experiment design standards, review gates, convergence protocol).
  • Convergence Protocol — Three convergence-gated review stages per iteration: Design Review (5 perspectives), Code Review (5 perspectives), FINDINGS Review (10 perspectives).
  • PR Workflow — Implementation of winning strategies follows the standard PR workflow.