• Sources: post, HN discussion
  • Summary: Wayne states that TLA+ properties are implicitly quantified over every individual behaviour, so it cannot natively express reachability, meaning that there exists a behaviour where a property holds, nor hyperproperties defined over sets of behaviours, nor properties spanning two or more steps, nor floating point or real time as opposed to logical time. He notes that hyperproperties cover many security properties and all statistical properties, giving a 95th percentile response time as the example. The named workarounds are described as hacks with costs, because auxiliary variables break refinement, self-composition exponentiates the state space, and TLC's REACHABLE keyword and TLCGet cover only the most basic cases.
  • Why it matters: The argument is about what can be stated rather than about how hard the checking is, so a team pointing an agent at TLA+ gets no coverage of the security and latency properties it is most likely to care about.