← Ideas & Research

Classification Without Sisyphean Taxonomies: Bisimulation as ‘Distinguishable by Experiment’

“The Pledge, Part IV.” — Lucius Meredith (Feb 2026)

In brief: Classification by naming tends to expand without converging on stable, testable boundaries. Bisimulation provides a formal model for grouping behaviors into equivalence classes defined by what observations can separate them. This framing emphasizes reproducibility: classification improves when experiments are explicit and repeatable.

Key takeaways

  • Classification by naming tends to expand without converging on stable, testable boundaries.
  • Bisimulation provides a formal model for grouping behaviors into equivalence classes defined by what observations can separate them.
  • This framing emphasizes reproducibility: classification improves when experiments are explicit and repeatable.

Key questions

  • How do you choose which experiments define the classification?
  • What happens when the classification granularity doesn’t match the use case?
  • Where does bisimulation-based classification offer advantages over statistical clustering?

Why it matters for developers

A behavior-first approach pushes you to define explicit probes and keep evidence of outcomes, so you can audit why something was classified the way it was.

Related: Bisimulation for AI: Why “Behavior, Not Structure” Changes the Story of Intelligence

Read the full essay on Substack →

classification bisimulation equivalence classes reproducibility experimental method