Formal Methods

Overview

This area tracks specification, synthesis, abstraction, verification, and proof-oriented tools for intelligent systems.

Active Questions

  • Which formal guarantees survive contact with neural function approximation and partial observability?
  • What abstractions give the best tradeoff between tractability and fidelity?
  • When should verification optimize expected behavior, full distributions, or explicit tail-risk measures?
  • How should verification results be surfaced back into the wiki as reusable conceptual knowledge?
  • When does a hierarchical or compositional specification preserve behavior while improving tractability?
  • How can model-building and causal-temporal formalisms expose where expressivity, decidability, and strategic agency trade off?
  • Which product-game abstractions are strong enough to shield learned agents without requiring a full transition model?
  • How far can compositional assume-guarantee reasoning push shield synthesis before local obligations become too conservative or manually brittle?
  • How can formal responsibility criteria support strategy selection before a trace is realized?
  • How can data-derived safe sets be certified conservatively when the model is unknown but structured?
  • When does a proof or specification require constructive evidence rather than a classical truth-value argument?
  • When does a specification depend on full second-order semantics, and when is a Henkin or automata-restricted reading enough?
  • Which proof-assistant workflows should be treated as reusable formal-methods infrastructure rather than just implementation tooling?
  • How can safety shields remain interpretable when their synthesized policy is too large to inspect directly?
  • Which numerical model-checking bounds are strong enough to serve as runtime safety certificates rather than only approximate diagnostics?
  • Which learned safety value functions are conservative enough to serve as certificates or runtime filters?
  • Which constrained-learning guarantees are optimization guarantees, and which are formal runtime-enforcement guarantees?
  • When do restricted temporal fragments genuinely lower synthesis complexity, and when do they only make automata construction cleaner?
  • Which game objectives are expressive enough for controller synthesis while still supporting symbolic solution methods?
  • Which real-time duration requirements can be compiled into executable monitors or shields without falling into undecidable fragments?
  • Which specification logics are best handled by direct automata translations, and when does translation blowup force a specialized fragment or symbolic method?
  • When is a formal specification naturally a linear trace property, and when is it a branching-time property over possible executions?
  • When can subsystem-level RL success guarantees be composed into a verified system-level probability guarantee?
  • When does a safety claim need an auditable world model, safety specification, and verifier rather than empirical evaluation alone?

Key Concepts

Key Sources

Adjacent Foundations