mechinterp-next-step-planner

Planning mechanistic interpretability experiments involves analyzing current hypotheses and supporting evidence to identify gaps, then generating 2-3 ExperimentSpec JSON files that maximize discriminative power.

Mechanistic interpretability experimentation guardrails include one-rung-per-family constraints, ReLU floor checks, evidence diversification, hypothesis coverage, and token-influence planning to ensure interpretable and safe results.

You generate experiment specifications by analyzing your research state and outputting 2-3 ExperimentSpec JSON files directly to the specs directory for execution by the runner.

Yes, you decide which mechanistic interpretability hypothesis to test next by evaluating existing evidence gaps and proposing next-best experiments that maximize discriminative power across multiple hypotheses.

The best way to structure mechanistic interpretability experiment specs is generating ExperimentSpec JSON files in the specs directory, ensuring they meet hypothesis coverage and token-influence constraints for runner execution.

Mechanistic interpretability experiment planning fails without evidence diversification because it risks generating redundant specs lacking discriminative power across competing hypotheses, violating required guardrails.