Update report.md

Author: dsyme

docs/report.md

@@ -747,4 +747,4 @@ Further:
 
 1. We speculate that with further prompting properties such as units (e.g. dollars) and multiplicative constants (e.g. millions) can be really carefully tracked and assessed. Unit mistakes are pernicious and the automated extraction of unit derivations can help here. Code could potentially be generated in languages such as F#, which support [statically-checked units of measure](https://learn.microsoft.com/dotnet/fsharp/language-reference/units-of-measure), or systems where units are attached to numeric types and checked dynamically.
 2. It is now clear that there are many other classes of "equips" that can be dealt with this way. For example, we believe GPT-4 can be "equipped" with Datalog, or an SMT-solver, or Wolfram, or further programmatic data retrieval, or SQL queries. It remains open whether these are best handled through generating specific DSLs or arbitrary code in languages such as Python.
-3. Given the relative simplicity of this technique, we speculate that any use of LLM AI outputs in high-trust applications may eventually require the generation of explanatory documentation for each written number. Unlike words, numbers must essentially always have a coherent calculational explanation, and so it seems logical that any AI system writing numbers will soon be required to also provide such explanations.
+3. Given the relative simplicity of this technique, we speculate that any use of LLM AI outputs in high-trust applications or regulated domains may eventually require the generation of explanatory documentation for each written number. Numbers are always "rational" (no pun intended) - that is, unlike words, numbers must always have a coherent calculational explanation - you're never justified writing a number in an output without having an explicable reason to write it. So it seems logical that any AI system writing numbers may be required to provide such explanations.