diff --git a/doc/expected.md b/doc/expected.md
index 3e22e9e..35bdaa2 100644
--- a/doc/expected.md
+++ b/doc/expected.md
@@ -1,5 +1,173 @@
 # expected<T, E...>
 
+## Description
+
+The class `expected<T, E...>` presented here is an extended version of `expected<T, E>` as
+proposed in [P0323R1](http://www.open-std.org/jtc1/sc22/wg21/docs/papers/2016/p0323r1.pdf)
+and the subsequent [D0323R2](https://github.com/viboes/std-make/blob/master/doc/proposal/expected/d0323r2.md).
+
+The main difference is that this class takes more than one error type, which makes it more
+flexible. One example of a type of the `expected` family, [outcome<T>](https://ned14.github.io/boost.outcome/),
+can store either an error of type `std::error_code`, or an exception in the form of `std::exception_ptr`.
+This can be represented naturally in this implementation via `expected<T, std::error_code, std::exception_ptr>`.
+
+In addition, libraries would generally differ in their choice of error types. It would be a
+common need in practice of having to combine the results of calling two different libraries,
+each with its own error type, such as the two examples below:
+
+    // Library 1
+
+    namespace lib1
+    {
+
+    enum class error
+    {
+        division_by_zero,
+        other_error
+    };
+
+    expected<double, error> div( double x, double y );
+
+    } // namespace lib1
+
+    // Library 2
+
+    namespace lib2
+    {
+
+    enum class error
+    {
+        division_by_zero,
+        negative_logarithm
+    };
+
+    expected<double, error> log( double x );
+
+    } // namespace lib2
+
+In this proposal, combining the results of `lib1::div` and `lib2::log` can be achieved via
+simple composition:
+
+    expected<double, lib1::error, lib2::error> log_div_mul( double x, double y, double m )
+    {
+        auto r1 = lib1::div( x, y );
+        if( !r1 ) return r1.unexpected();
+
+        auto r2 = lib2::log( r1.value() );
+        if( !r2 ) return r2.unexpected();
+
+        return m * r2.value();
+    }
+
+An alternative approach that requires more effort is also supported:
+
+    enum class common_error
+    {
+        division_by_zero,
+        negative_logarithm,
+        other_error,
+        unknown_error
+    };
+
+    common_error make_common_error( lib1::error e );
+    common_error make_common_error( lib2::error e );
+
+    expected<double, common_error> log_div_mul2( double x, double y, double m )
+    {
+        static const auto rm = []( auto x ) { return make_common_error(x); };
+
+        auto r1 = lib1::div( x, y ).remap_errors( rm );
+        if( !r1 ) return r1.unexpected();
+
+        auto r2 = lib2::log( r1.value() ).remap_errors( rm );
+        if( !r2 ) return r2.unexpected();
+
+        return m * r2.value();
+    }
+
+When an attempt to access the value via `r.value()` is made and an error is present,
+an exception is thrown. By default, this exception is of type `bad_expected_access<E>`,
+as in D0323R2, but there are two differences. First, `bad_expected_access<E>` objects
+derive from a common base `bad_expected_access<void>` so that they can be caught at
+points where the set of possible `E` is unknown.
+
+Second, the thrown exception can be customized. The implementation calls
+`throw_on_unexpected(e)` unqualified, where `e` is the error object, and the user can
+define such a function in the namespace of the type of `e`. Two specialized overloads
+of `throw_on_unexpected` are provided, one for `std::error_code`, which throws the
+corresponding `std::system_error`, and one for `std::exception_ptr`, which rethrows
+the exception stored in it.
+
+For example, `lib1` from above may customize the exceptions associated with `lib1::error`
+via the following:
+
+    namespace lib1
+    {
+
+    enum class error
+    {
+        division_by_zero,
+        other_error
+    };
+
+    class exception: public std::exception
+    {
+    private:
+
+        error e_;
+
+    public:
+
+        explicit exception( error e ): e_( e ) {}
+        virtual const char * what() const noexcept;
+    };
+
+    void throw_on_unexpected( error e )
+    {
+        throw exception( e );
+    }
+
+    } // namespace lib1
+
+In this implementation, `unexpected_type<E...>` has been called `unexpected_<T...>` and is
+an alias for `variant<T...>`. It is unfortunately not possible to use the name `unexpected<T...>`,
+because a function `std::unexpected` already exists.
+
+The `make_...` helper functions have been omitted as unnecessary; class template argument deduction
+as in `expected{ 1.0 }` or `unexpected_{ lib1::division_by_zero }` suffices.
+
+Other functions have also been dropped as unnecessary, not providing sufficient value, dangerous, or
+a combination of the three, although the decision of what to include isn't final at this point. The aim
+is to produce a minimal interface that still covers the use cases.
+
+`expected<T, E1...>` can be converted to `expected<T, E2...>` if all error types in `E1...` are
+also in `E2...`. This allows composition as in the example above. Whether value convertibility ought
+to also be supported is an open question.
+
+A single monadic operation ("bind") is supported in the form of `operator>>`, allowing
+
+    auto log_div_mul3( double x, double y, double m )
+    {
+        return lib1::div( x, y ) >> [&]( auto && r1 ) {
+
+            return lib2::log( r1 ) >> [&]( auto && r2 ) -> expected<double, lib1::error, lib2::error> {
+
+                return m * r2;
+
+            };
+        };
+    }
+
+as well as the more concise in this example, although limited in utility for real world scenarios,
+
+    auto log_div_mul3( double x, double y, double m )
+    {
+        return lib1::div( x, y ) >> std::bind<expected<double, lib1::error, lib2::error>>( lib2::log, _1 ) >> m * _1;
+    }
+
+The more traditional name `then` was also a candidate for this operation, but `operator>>` has two advantages;
+it avoids the inevitable naming debates and does not require parentheses around the continuation lambda.
+
 ## Synopsis
 
     // unexpected_