Isabelle is an interactive theorem prover. It comes with an IDE where users can type mathematical specifications and proofs which will be checked continuously. Specifications consist of a sequence of type and value definitions. All high-level tools inside Isabelle must justify their proofs and constructions against a small inference kernel. This greatly diminishes the risk of unsound proofs.

Additionally, Isabelle features powerful proof automation, for example:

  • classical and equational reasoning
  • decision procedures (e.g. for linear arithmetic)
  • interfaces to automated provers (e.g. Z3 and CVC4)

However, most non-trivial proofs will require the user to give proof hints. In general, interactive provers like Isabelle trade better guarantees for weaker automation.

Normally, users write Isabelle specifications manually. Leon comes with a bridge to Isabelle which allows it to be used as a solver for program verification.


You don’t have to obtain a copy of Isabelle. Leon is able to download and install Isabelle itself. The installation happens in the appropriate folder for your operating system, e.g. %APPDATA% under Windows or $HOME/.local under Linux.

Basic usage

For most purposes, Isabelle behaves like any other solver. When running Verification, you can pass isabelle as a solver (see Command Line Options for details). Isabelle is a bit slower to start up than the other solvers, so please be patient.

Advanced usage

Isabelle has some peculiarities. To accomodate for those, there are a set of specific annotations in the object leon.annotation.isabelle.

Mapping Leon entities to Isabelle entities

Isabelle – or it’s standard logic HOL, to be precise – ships with a rather large existing library. By default, the Isabelle backend would just translate all Leon definitions into equivalent Isabelle definitions, which means that existing proofs about them are not applicable. For example, Isabelle/HOL already has a concatenation function for lists and an associativity proof for it. The annotations @typ, @constructor and @function allow the user to declare a mapping to existing Isabelle entities. For lists and their append function, this looks like this:

@isabelle.typ(name = "List.list")
sealed abstract class List[T] {

  @isabelle.function(term = "List.append")
  def ++(that: List[T]): List[T] = (this match {
    case Nil() => that
    case Cons(x, xs) => Cons(x, xs ++ that)


@isabelle.constructor(name = "List.list.Cons")
case class Cons[T](h: T, t: List[T]) extends List[T]

@isabelle.constructor(name = "List.list.Nil")
case class Nil[T]() extends List[T]

The above is a simplified excerpt from Leon’s standard library. However, it is no problem to map functions with postconditions to their Isabelle equivalents.

By default, mapping for types is loosely checked, whereas mapping for functions is checked:

  • For types, it is checked that there are exactly as many constructors in Leon as in Isabelle and that all constructors are annotated. It is also checked that the given constructor names actually exist and that they have the same number of arguments, respectively. There is no check whether the field types are equivalent. However, in most cases, such a situation would be caught later because of type errors.
  • For functions, a full equivalence proof is being performed. First, the function will be translated to Isabelle as usual. Second, the system will try to prove that both definitions are equivalent. If that proof fails, the system will emit a warning and ignore the mapping.

The checking behaviour can be influenced with the strict option (see Command Line Options).

Scripting source files

The script annotation allows to embed arbitrary Isabelle text into Leon source files. In the following example, this is used together with mapping:

  name = "Safe_Head",
  source = """
    fun safe_head where
    "safe_head [] = None" |
    "safe_head (x # _) = Some x"

    lemma safe_head_eq_head[simp]:
      assumes "~ List.null xs"
      shows "safe_head xs = Some (hd xs)"
    using assms
    by (cases xs) auto
@isabelle.function(term = "Safe_Head.safe_head")
def safeHead[A](xs: List[A]): Option[A] = xs match {
  case Nil() => None()
  case Cons(x, _) => Some(x)

script annotations are processed only for functions which are directly or indirectly referenced from the source file which is under verification by Leon. The effect of a script is equivalent to defining an Isabelle theory with the name and content as specified in the annotation, importing the Leon theory. Theories created via script annotations must be independent of each other, but are processed before everything else. As a consequence, any entities defined in scripts are available for all declarations.


Invalid proofs will raise an error, but skipped proofs (with sorry) are not reported.

Proof hints

The system uses a combination of tactics to attempt to prove postconditions of functions. Should these fail, a custom proof method can be specified with the proof annotation.

@isabelle.proof(method = """(induct "<var xs>", auto)""")
def flatMapAssoc[A, B, C](xs: List[A], f: A => List[B], g: B => List[C]) =
  (xs.flatMap(f).flatMap(g) == xs.flatMap(x => f(x).flatMap(g))).holds

The method string is interpreted as in Isabelle:

lemma flatMapAssoc: ...
by (induct "<var xs>", auto)


In annotations, the function parameters are not in scope. That means that referring to the actual Scala variable xs is impossible. Additionally, in Isabelle, xs will not be called xs, but rather xs'76 (with the number being globally unique). To be able to refer to xs, the system provides the special input syntax <var _>, which turns an identifier of a variable into its corresponding variable in Isabelle. Think of it as a quotation for Scala in Isabelle. There is also a counterpart for constants: <const _>.

The proof annotations admits a second argument, kind, which specifies a comma-seperated list of verification conditions it should apply to. The empty string (default) means all verification conditions.

Influencing the translation of functions

By default, the system will only translate the body of a function, that is, without pre- and postconditions, to Isabelle. Sometimes, the precondition is required for termination of the function. Since Isabelle doesn’t accept function definitions for which it can’t prove termination, the presence of the precondition is sometimes necessary. This can be achieved by annotating the function with @isabelle.fullBody. If, for other reasons, termination can’t be proven, the annotation @isabelle.noBody leaves the function unspecified: It can still be called from other functions, but no proofs depending on the outcome of the functions will succeed.

Advanced example

The following example illustrates the definition of a tail-recursive function. The challenge when proving correctness for these kinds of functions is that “simple” induction on the recursive argument is often not sufficient, because the other arguments change in the recursive calls. Hence, it is prudent to specify a proof hint. In this example, an induction over the definition of the lenTailrec function proves the goal:

def lenTailrec[T](xs: List[T], n: BigInt): BigInt = xs match {
  case Nil() => n
  case Cons(_, xs) => lenTailrec(xs, 1 + n)

@isabelle.proof(method = """(induct "<var xs>" "<var n>" rule: [[leon_induct lenTailrec]], auto)""")
def lemma[A](xs: List[A], n: BigInt) =
  (lenTailrec(xs, n) >= n).holds

The attribute [[leon_induct _]] summons the induction rule for the specified function.


  • Mutually-recursive datatypes must be “homogeneous”, that is, they all must have exactly the same type parameters; otherwise, they cannot be translated.
  • Recursive functions must have at least one declared parameter.
  • Polymorphic recursion is unsupported.
  • The const and leon_induct syntax take a mangled identifier name, according to the name mangling rules of Scala (and some additional ones). The mangling doesn’t change the name if it only contains alphanumeric characters.
  • The const and leon_induct syntax does not work for a given function f if there is another function f defined anywhere else in the program.