So yeah maybe it is implicit in the comic, but the math is backed by a specific set of assumption on how the words were selected. Or I could be wrong, I haven't verified it personally.
First character is 4 bits: c = 4
The next 7 chars are 2 bits/each: "orrect " = 14
Characters 9-20 are 1.5 bit/each: "horse batter" = 18
Charactes 21-n are 1 bit/each: "y staple" = 8
No bonus for including both upper and non-alpha chars.
No bonus for passwords of length less than 20 chars, not containing dictionary words, because the password is longer than 20 characters.
Total entropy: 4 + 14 + 18 + 8 = 44 bits of entropy.
The NIST entropy estimation is based off of characters that aren't chosen at random (I think?) and it is a heuristic.
For the approach I think Randall intended to describe they aren't really words, just glyphs chosen randomly from a set of glyphs. That these glyphs are easy to memorize and drop right into existing password interfaces is orthogonal.
The expected entropy is derived from "correct horse battery staple", including the spaces, using NIST SP 800-63.