experiment-driven-development

Evidence-driven development is a disciplined process that coordinates changes in LLM systems by enforcing real observed behavior rather than assumptions. It requires documenting hypotheses, baseline evidence, focused validation, and traceable git commits aligned with an iteration journal.

To validate prompt tuning changes, begin with verifiable system evidence, state a clear hypothesis, run a focused validation batch, and commit with a documented plan. This ensures every prompt adjustment is traceable and aligned with observed outcomes in the iteration journal.

Yes, this structured change workflow is applicable to any implementation task in LLM systems, including node logic updates, routing decisions, prompt tuning, and architectural changes during development, ensuring all modifications rely on evidence-driven validation.

The best way to stop assumption-based coding is to enforce an evidence-driven workflow that mandates collecting real-system evidence, formulating a testable hypothesis, executing focused validation, and producing a traceable git commit before finalizing changes.

Yes, you must document a hypothesis before tuning LLM routing logic. The process requires formulating a testable hypothesis alongside baseline evidence to ensure routing modifications are validated against real observed behavior rather than unverified assumptions.

LLM development cycles often fail when architectural changes rely on assumptions instead of real observed behavior. Enforcing an evidence-driven process with documented hypotheses, baseline evidence, and focused validation prevents unverified modifications from destabilizing the system.