diff --git a/doc/bootstrap.txt b/doc/bootstrap.txt
index 270e54e..dcae434 100644
--- a/doc/bootstrap.txt
+++ b/doc/bootstrap.txt
@@ -16,7 +16,7 @@ This binary can be separated in 5 distinct layers:
 2. Arch-specific boot words (B305 for Z80)
 3. Arch-independant core words (low) (B350)
 4. Drivers, might contain arch-specific code
-5. Arch-independant core words (high) (B380)
+5. Arch-independant core words (high) (B390)
 
 # Boot code
 
@@ -28,9 +28,7 @@ It also of course does core initialization: set RSP/PSP, HERE
 CURRENT, then call BOOT.
 
 It also contains what we call the "stable ABI" in its first
-0x100 bytes. The beginning of the dict is intertwined in this
-layer because EXIT, (br), (?br) and (loop) are part of the
-stable ABI.
+few (<0x20) bytes.
 
 # Boot words
 
@@ -53,25 +51,26 @@ See B260 for details.
 
 # Drivers
 
-Up until now, we haven't implemented EMIT or KEY yet: those
-words are defined in the "high" part of core words because we
-generally need machine-specific drivers to implement (emit) and
-(key).
+Core words don't include (key) and (emit) implementations be-
+cause that's hardware-dependant. This is where we need to load
+code that implement it, as well as any other code we want to
+include in the binary.
 
-Well, now is their time to shine. We split core in two
-precisely to fit drivers in there. This way, they have access
-to a pretty good vocabulary and they're also give the oppor-
-tunity to provide (emit) and (key).
+We do it now because if we wait until the high layer of core
+words is loaded, we'll have messed up immediates and ":" will
+be broken. If we load our code before, we won't have access to
+a wide vocabulary.
 
 # Core words (high)
 
-Then come EMIT, KEY and everything that depend on it, until
-we have a full Forth interpreter. At the very end, we define
-tricky IMMEDIATEs that, if defined earlier, would break cross
-compilation.
+The final layer of core words contains the BOOT word as well
+as tricky immediates which, if they're defined sooner, mess
+cross compilation up. Once this layer is loaded, we become
+severly limited in the words we can use without messing up.
 
-We end that with a hook words which is also where CURRENT will
-be on boot.
+After having loaded that last layer, we end that with a hook
+word which is also where CURRENT will be on boot, which is then
+followed by the initialization string (see below).
 
 So that's the anatomy of a Collapse OS binary. How do you build
 one? If your machine is already covered by a recipe, you're in
@@ -110,4 +109,9 @@ part. The boot sequence is designed to interpret whatever comes
 after LATEST as Forth source, and this, until it reads ASCII
 EOT character (4). This is generally used for driver init.
 
+To produce a Collapse OS binary, you run that xcomp unit and
+then observe the values of "ORG @" and "H@". That will give you
+the start and stop offset of your binary, which you can then
+copy to your target media.
+
 Good luck!