diff --git a/Makefile b/Makefile
index 6ad8604..188447a 100644
--- a/Makefile
+++ b/Makefile
@@ -25,8 +25,8 @@ endif
 
 default: kernel.rom
 
-kernel.rom: kernel/*.asm lib/*.inc
-	nasm -s -o $@ -I lib -I kernel kernel/main.asm && scripts/fix-rom.sh $@
+kernel.com: kernel/*.asm lib/*.inc
+	nasm -s -o $@ -I lib -I kernel kernel/main.asm
 
 debug.rom: debug/*.asm lib/*.inc
 	nasm -s -o $@ -I lib -I debug debug/main.asm && scripts/fix-rom.sh $@
@@ -34,11 +34,11 @@ debug.rom: debug/*.asm lib/*.inc
 boot/%.bin: boot/%.asm lib/*.inc
 	nasm -s -o $@ -I boot -I lib $<
 
-fdimage.img: boot/floppy.bin kernel.rom
+fdimage.img: boot/floppy.bin kernel.com
 	dd if=/dev/zero bs=512 count=$$(( $(FD_CYLINDERS) * $(FD_HEADS) * $(FD_SECTORS) )) of=$@
 	mformat -i $@ -t $(FD_CYLINDERS) -h $(FD_HEADS) -n $(FD_SECTORS) -B boot/floppy.bin ::
-	mcopy -i $@ kernel.rom ::kernel.rom
-	mattrib -i $@ +s ::kernel.rom
+	mcopy -i $@ kernel.com ::kernel.com
+	mattrib -i $@ +s ::kernel.com
 
 clean:
 	rm -f *.com *.bin *.rom *.img *.log boot/*.bin
@@ -47,7 +47,7 @@ qemu-rom: kernel.rom $(DEBUGROM)
 	$(QEMU) $(QEMU_ARGS) -option-rom kernel.rom
 
 qemu-floppy: fdimage.img $(DEBUGROM)
-	$(QEMU) $(QEMU_ARGS) -boot c -hda fdimage.img
+	$(QEMU) $(QEMU_ARGS) -boot c -fda fdimage.img
 
 qemu-serial: boot/serial.bin $(DEBUGROM)
 	$(QEMU) $(QEMU_ARGS) -hda boot/serial.bin
diff --git a/boot/chs.inc b/boot/chs.inc
deleted file mode 100644
index 1cceb56..0000000
--- a/boot/chs.inc
+++ /dev/null
@@ -1,47 +0,0 @@
-fix_chs:
-  mov ah, 0x08
-  int 0x13
-  jc chs_end ; skip if function does not exist
-  inc dh
-  mov [cs:fdc.nos], dh
-  push cx
-  and cx, 0x003F
-  mov [cs:fdc.spt], cx ; no adjustment because sectors are 1-indexed
-  pop ax
-  xchg al, ah
-  mov cl,6
-  shr ah,cl
-  inc ax ; convert from maximum number (0-based) to total number (1-based) of cylinders
-  mul word [cs:fdc.nos] ; number of tracks = number of cylinders * heads
-  mul word [cs:fdc.spt] ; number of sectors = number of tracks * sectors per track
-  mov [cs:fdc.ts], ax
-  jmp chs_end
-
-; Load a single sector into memory
-; Does not return on error
-; in    dx:ax   sector number (will be trashed)
-;       es:bx   buffer
-load_sector:
-  push cx
-  div word [cs:fdc.spt]    ; ax:temp     = (lba / spt)
-  inc dx                   ; dx:sector   = (lba % spt) + 1
-  mov cl, dl ; sector number
-  xor dx, dx
-  div word [cs:fdc.nos]    ; ax:cylinder = (tmp / heads)
-                           ; dx:head     = (tmp % heads)
-  mov ch, al ; cylinder number
-  mov dh, dl ; head number
-  mov dl, [cs:fdc.drv] ; driver number
-  mov ax, 0x0201 ; ah=0x02 al=0x01
-  int 0x13
-  jc disk_error
-  ; return
-  pop cx
-  ret
-
-disk_error:
-  xchg al, ah
-  call error
-  db "DISK ERROR", 0
-
-chs_end:
\ No newline at end of file
diff --git a/boot/floppy.asm b/boot/floppy.asm
index 7a8ce27..5028537 100644
--- a/boot/floppy.asm
+++ b/boot/floppy.asm
@@ -1,8 +1,17 @@
-org 0x0000
+org 0x7600
 
-jmp 0x07C0:main
+; Memory layout:
+; 060:0000 PSP for loaded program
+; 060:0100 Loaded program
+; 060:7800 Our code (= 0:7C00)
+; 060:8000 FAT table
+; 060:E000 Root directory
+; 060:FF00 Stack area
 
-times (0x0B - ($-$$)) db 0 ; space-padded
+init:
+  jmp 0x060:main
+
+times (0x0B - ($-$$)) db 0
 fdc:
 .ss:
   dw 0x200	; sector size
@@ -41,119 +50,207 @@ fdc:
   db "FAT12"
   times (62 - ($-$$)) db " "
 
-main:
-  cli
-  ; setup stack area growing down from directly below us
-  xor ax, ax
-  mov ss, ax
-  mov sp, 0x7C00
-  ; setup base segments
-  mov ax, cs
-  mov ds, ax
-  mov es, ax
-  ; save initial environment
-  push bx
+; Read sectors from disk
+; Does not return on error
+; ax and bx will be incremented, cx decremented
+; in    dx:ax   sector number
+;       es:bx   buffer
+;	cx	number of sectors to read
+; out	dx:ax	next sector to read
+;	es:bx	next free buffer
+;	cx	zero
+load_sectors:
+  ; fail instantly if reading sectors > 16 bit
+  test dx, dx
+  mov bp, 0x0030
+  jnz error
+
+.loop:
+  push ax
   push cx
   push dx
 
-  mov [fdc.drv], dl ; fix drive number in fdc
+  ; calculate CHS data
+  div word [cs:fdc.spt]    ; ax:temp     = (lba / spt)
+  inc dx                   ; dx:sector   = (lba % spt) + 1
+  mov cl, dl ; sector number
+  xor dx, dx
+  div word [cs:fdc.nos]    ; ax:cylinder = (tmp / heads)
+                           ; dx:head     = (tmp % heads)
+  mov ch, al ; cylinder number
+  mov dh, dl ; head number
+  mov dl, [cs:fdc.drv] ; driver number
+  mov ax, 0x0201 ; ah=0x02 al=0x01
+  int 0x13
+  mov bp, 0x0031
+  jc error
 
-%include "chs.inc"
+  pop dx
+  pop cx
+  pop ax
 
-  mov di, 0x1000 ; loading address for roms
-  call search_rootdir
+  ; count up for next sector
+  add bx, 0x0200
+  inc ax
 
-  mov ax, 0xFFFF
-  call error
-  db "THING", 0
+  loop .loop
+  ret
 
-search_rootdir:
+; Load the file in [filename]
+; or exit with carry set
+load_file:
+  mov si, 0xE000
+  mov cx, [fdc.rde]
+.loop:
+  call file_match
+  je read_clusters
+  add si, 0x20
+  loop .loop
+  ret
+
+; Compares the FAT dirent against [filename]
+; in	ds:si	filename
+; out	ZF	zero flag set if equal
+file_match:
+  push cx
+  push si
+  mov di, filename
+  mov cx, 11
+  repe cmpsb
+  pop si
+  pop cx
+  ret
+
+; Read the file given by cluster number
+; into the target program area
+; in	ax	cluster number
+read_clusters:
+  add si, 0x1A
+  lodsw
+  mov bx, 0x0100
+.loop:
+  ; read cluster into area for target file
+  push ax
+  sub ax, 2
+  mul BYTE [fdc.sc]
+  add ax, [cluster_offset]
+  xor dx, dx
+  mov cx, [fdc.sc]
+  xor ch, ch
+  call load_sectors
+  pop ax
+
+  ; calculate index in FAT
+  mov si, ax
+  shr si, 1
+  add si, ax
+  add si, 0x8000
+
+  ; load entry from FAT, truncate to 12 bit
+  mov dx, [si]
+  test ax, 1
+  jz .noshift
+  mov cl, 4
+  shr dx, 4
+.noshift:
+  mov ax, dx
+
+  and ax, 0x0FFF
+  cmp ax, 0x0FF8
+  jc .loop
+
+  ret
+
+main:
+  cli
+  ; Everything in one segment, stack starts on top
+  mov ax, cs
+  mov ds, ax
+  mov es, ax
+  mov ss, ax
+  mov sp, ax
+  mov [fdc.drv], dl ; save drive number in fdc
+  sti
+
+  ; load fat table into memory
+  mov ax, [fdc.rsc]
+  mov cx, [fdc.sf]
+  xor dx, dx
+  mov bx, 0x8000
+  call load_sectors
+
+  ; calculate length of rootdir
+  mov ax, [fdc.rde]
+  dec ax
+  mov cl, 4
+  shr ax, cl ; 32 bytes per entry
+  inc ax
+  mov cx, ax
+
+  ; load root dir
+  xor dx, dx
   mov ax, [fdc.sf]
   mul byte [fdc.fn]
   add ax, [fdc.rsc]
-  mul word [fdc.ss]
-  mov si, ax ; now pointing to start of rootdir
-  mov cx, [fdc.rde]
-  jmp .start
-.next:
-  dec cx
-  jz .end
-  add si, 0x20
-.start:
-  call lazy_load
-  cmp byte [si], 0 ; end of directory
-  je .end
-  cmp byte [si], 0xe5 ; deleted entry
-  je .next
-  cmp word [si+8], 0x4F52 ; first 2 chars of extension: "RO"
-  jne .next
-  mov ax, [si+10] ; 3rd char of extension, file flags
-  and ax, 0x14FF ; zero out flags except directory and system
-  cmp ax, 0x044D ; "M" , system flag set, directory flag unset
-  jne .next
-  mov ax, [si+26]
-  push si
-  push cx
-  call loadrom
-  pop cx
-  pop si
-  jmp .next
-.end:
-  ret
+  mov bx, 0xE000
+  call load_sectors
 
-; in	ax	cluster number
-loadrom:
-  call print16
-  ret
+  ; remember where we left off
+  ; clusters start after rootdir
+  mov [cluster_offset], ax
 
-; Makes sure the block starting with (or under SI) is
-; already fetched from disk
-; in	ds:si	data expected to be fetched from disk
-; trashes ax, bx and dx
-lazy_load:
-  cmp si, [max_si]
-  jc .end
-  ; calculate next sector number
-  mov ax, [max_si]
+  call load_file
+  mov bp, 0x0032
+  jc error
+
+  ; get length of arguments
+  mov cx, 0x007F
+  mov di, arguments
+  xor ax, ax
+  repe scasb
+  sub di, arguments
+  mov cx, di
+
+  ; setup arguments field
+  mov [0x080], cl
+  mov di, [0x081]
+  mov si, arguments
+  rep movsb
+
+  ; setup int 19h call at 0000, and push its address to stack
+  mov WORD [0000], 0x19CD
+  push sp
+
+  ; clear out registers and jump into target
+  xor ax, ax
+  xor cx, cx
   xor dx, dx
-  div word [fdc.ss]
-  ; write to es:bx
-  xor dx, dx
-  push cs
-  pop es
-  mov bx, [max_si]
-  call load_sector
-  ; update max_si counter
-  mov ax, [max_si]
-  add ax, [fdc.ss]
-  mov [max_si], ax
-  ; loop until requirement is satisfied
-  jmp lazy_load
-.end:
-  ret
+  xor bx, bx
+  mov dl, [fdc.drv]
+  jmp 0x0100
 
 error:
-  call print16
-  pop si
-  mov bx, 0x0000
-  mov al, 0x20
-.loop:
+  mov ax, bp
   mov ah, 0x0e
+  mov bx, 7
   int 0x10
-  lodsb
-  cmp al, 0x00
-  jne .loop
-.hlt:
-  hlt
-  jmp .hlt
+  mov al, 0x21
+  int 0x10
+  xor ax, ax
+  int 0x16
+  int 0x19
 
-%include "print.asm"
+cluster_offset:
+  dw 0
 
-times (0x1FA - ($-$$)) db 0
+times (0x175 - ($-$$)) db 0
 
-dw 0x2B2B
+filename:
+  db "KERNEL  COM"
 
-max_si:
-dw 0x200
+arguments:
+  db " ", 0
+
+times (0x1FE - ($-$$)) db 0
 
 dw 0xAA55
diff --git a/kernel/main.asm b/kernel/main.asm
index 32cb154..5791244 100644
--- a/kernel/main.asm
+++ b/kernel/main.asm
@@ -1,60 +1,5 @@
 cpu 8086
-org 0x0000
-rom:
-  db 0x55, 0xAA
-.sectors:
-  db 0x00
-.init:
-  push cs
-  pop es
-
-  mov di, start
-  mov ax, 0x18
-  call intr_register
-
-  retf
-.name:
-  db "ROM DOS", 0
-
-  times (0x18 - ($-$$)) db 0
-.pcir_ptr:
-  dw 0
-
-  times (0x1A - ($-$$)) db 0
-.pnp_ptr:
-  dw pnp
-
-pnp:
-  db "$PnP"
-.version:
-  db 1 ; version 1
-.length:
-  db 2 ; 2 * 16 length
-  dw 0 ; offset of next header
-  db 0
-.checksum:
-  db 0 ; checksum (filled by fix-rom)
-  dd 0 ; device identifier
-  dw 0 ; manufacturer string
-  dw rom.name ; product name string
-  db 0,0,0 ; device type string
-  db 0x20 ; device indicator, bit for "read cacheable" set
-  dw 0 ; boot connection vector
-  dw 0 ; boot disconnect vector
-  dw start ; bootstrap entry point
-  dw 0 ; reserved
-  dw 0
-
-start:
-  xor ax, ax
-  mov ds, ax
-  mov es, ax
-  mov ss, ax
-  mov sp, 0x800
-  mov di, sp
-
-  call drvtab_create
-
+org 0x0100
   push cs
   pop es
   mov di, isr_dos_main
@@ -64,7 +9,6 @@ start:
 .loop:
   mov ah, 0x01
   int 0x21
-  int3
   jmp .loop
 
   cli
@@ -91,4 +35,4 @@ isr_return:
 
 %include "chario.asm"
 
-align 512
+times 1000 db 0xEA