diff --git a/flag.txt b/flag.txt
new file mode 100644
index 0000000..421376d
--- /dev/null
+++ b/flag.txt
@@ -0,0 +1 @@
+dummy
diff --git a/solve/solve b/solve/solve
index 5bf9bde..0e5906c 100644
Binary files a/solve/solve and b/solve/solve differ
diff --git a/solve/solve.s b/solve/solve.s
index c0ede8a..35c0538 100644
--- a/solve/solve.s
+++ b/solve/solve.s
@@ -1,10 +1,11 @@
 ; vim: ft=nasm
+s:
+
 %macro i_nop 0
     db 0x00
 %endmacro
 %define np i_nop
 
-
 %macro i_halt 1
     db 0x01
     db %1
@@ -42,7 +43,8 @@
     db 0x50
     db %1
     db %2
-    db %3
+    db (%3 >> 8) & 0xff
+    db %3 & 0xff
 %endmacro
 %define jl i_less
 
@@ -53,10 +55,19 @@
     db 0x51
     db %1
     db %2
-    db %3
+    db (%3 >> 8) & 0xff
+    db %3 & 0xff
 %endmacro
 %define le i_leq
 
+%macro break 0
+    sd disp, 0x00
+    sd disp, 0x01
+    sd disp, 0x02
+    sd disp, 0x03
+    np
+%endmacro
+
 %define inp 0
 %define disp 1
 %define arith 2
@@ -69,72 +80,324 @@ gt 0x02, inp
 gt 0x03, inp
 ; [a, b, c, d] -> mem
 sd tape, 0x00
-sd tape, 0x01
 sd tape, 0x02
+sd tape, 0x01
 sd tape, 0x03
 ; {a, b, c, d}
+
+
+sd arith, 0x00
+sd arith, 0x03
+
+sd tape, 0x00
+ld 0x00, 0x01
+sd arith, 0x00
+gt 0x00, tape
+
+gt 0x03, arith
 ; [i[0], i[1], i[2], i[3] + i[0]]
 ; - add i[3] and i[0]
 
+sd arith, 0x01
+sd arith, 0x03
+
+sd tape, 0x00
+ld 0x00, 0x03
+sd arith, 0x00
+gt 0x00, tape
+
+gt 0x03, arith
 ; [i[0], i[1], i[2], (i[3] + i[0]) ^ i[1]]
 ; - xor i[1] with i[3] + i[0], pop from arith
 
+sd arith, 0x00
+sd arith, 0x01
+
+sd tape, 0x00
+ld 0x00, 0x01
+sd arith, 0x00
+gt 0x00, tape
+
+gt 0x00, arith
 ; [i[0] + i[1], i[1], i[2], (i[3] + i[0]) ^ i[1]]
 ; - add i[0] and i[1], pop from arith
 
+sd arith, 0x00
+sd arith, 0x02
+
+sd tape, 0x00
+ld 0x00, 0x03
+sd arith, 0x00
+gt 0x00, tape
+
+gt 0x02, arith
 ; [i[0] + i[1], i[1], (i[0] + i[1]) ^ i[2], (i[3] + i[0]) ^ i[1]]
 ; - xor i[0] + i[1] with i[2], pop from arith
 
+
+sd arith, 0x00
+sd arith, 0x02
+
+sd tape, 0x00
+ld 0x00, 0x03
+sd arith, 0x00
+gt 0x00, tape
+
+gt 0x00, arith
 ; [i[2], i[1], (i[0] + i[1]) ^ i[2], (i[3] + i[0]) ^ i[1]]
 ; - xor (i[0] + i[1]) ^ i[2] with i[0] + i[1], pop from arith
 
+sd arith, 0x00
+sd arith, 0x01
+
+sd tape, 0x00
+ld 0x00, 0x01
+sd arith, 0x00
+gt 0x00, tape
+
+gt 0x01, arith
+sd arith, 0x01
 ; [i[2], i[2] + i[1], (i[0] + i[1]) ^ i[2], (i[3] + i[0]) ^ i[1]]
 ; - add i[2] + i[1], pop from arith | store i[2] + i[1] in arith
 
+gt 0x01, tape
+sd arith, 0x01
 ; [i[2], i[3], (i[0] + i[1]) ^ i[2], (i[3] + i[0]) ^ i[1]]
 ; - pop i[3] from tape use previous i[1] + i[2] from arith to xor
 
+
+sd tape, 0x00
+ld 0x00, 0x03
+sd arith, 0x00
+gt 0x00, tape
+
+gt 0x01, arith
 ; [i[2], (i[2] + i[1]) ^ i[3], (i[0] + i[1]) ^ i[2], (i[3] + i[0]) ^ i[1]]
 ; - use previous i[1] + i[2] from arith to xor and pop
 
+sd arith, 0x00
+gt 0x00, tape
 ; [i[1], (i[2] + i[1]) ^ i[3], (i[0] + i[1]) ^ i[2], (i[3] + i[0]) ^ i[1]]
 ; - put i[2] in arith, pop i[1] from tape
 
+sd arith, 0x00
+
+sd tape, 0x00
+ld 0x00, 0x01
+sd arith, 0x00
+gt 0x00, tape
+
+gt 0x00, arith
 ; [i[2] + i[1], (i[2] + i[1]) ^ i[3], (i[0] + i[1]) ^ i[2], (i[3] + i[0]) ^ i[1]]
 ; - put i[1] in arith, add, pop
 
+sd arith, 0x00
+sd arith, 0x01
+
+
+sd tape, 0x00
+ld 0x00, 0x03
+sd arith, 0x00
+gt 0x00, tape
+
+gt 0x00, arith
+
 ; [i[3], (i[2] + i[1]) ^ i[3], (i[0] + i[1]) ^ i[2], (i[3] + i[0]) ^ i[1]]
 ; - xor i[2] + i[1] with (i[2] + i[1]) ^ i[3]
 
+sd arith, 0x00
+gt 0x00, tape
+
 ; [i[2], (i[2] + i[1]) ^ i[3], (i[0] + i[1]) ^ i[2], (i[3] + i[0]) ^ i[1]]
 ; - put i[3] in arith, pop i[2] from tape
 
+sd arith, 0x00
+
+sd tape, 0x00
+ld 0x00, 0x01
+sd arith, 0x00
+gt 0x00, tape
+
+gt 0x00, arith
+
 ; [i[2] + i[3], (i[2] + i[1]) ^ i[3], (i[0] + i[1]) ^ i[2], (i[3] + i[0]) ^ i[1]]
 ; - put i[2] in arith, add to i[3]
 
 
+sd arith, 0x00
+gt 0x00, tape
 ; [i[0], (i[2] + i[1]) ^ i[3], (i[0] + i[1]) ^ i[2], (i[3] + i[0]) ^ i[1]]
 ; - put i[2] + i[3] in arith, pop i[0] from tape
 
+sd arith, 0x00
+
+sd tape, 0x00
+ld 0x00, 0x03
+sd arith, 0x00
+gt 0x00, tape
+
+gt 0x00, arith
 ; [(i[2] + i[3]) ^ i[0], (i[2] + i[1]) ^ i[3], (i[0] + i[1]) ^ i[2], (i[3] + i[0]) ^ i[1]]
 ; - xor in arith
 
+
+
+sd tape, 0x00
+sd tape, 0x01
+sd tape, 0x02
+sd tape, 0x03
+gt 0x01, tape
+gt 0x02, tape
+gt 0x03, tape
+gt 0x00, tape
+
+
 ; push all to tape
 ; pop all back into mem in correct spot
+; stage 1 done
+
+; :b [a, b, c, d] (, ,)
+sd tape, 0x01
+; :b [a, b, c, d] (b, , )
+sd arith, 0x01
+; :b [a, 0, c, d] (b, 0, +)
+ld 0x01, 0x00
+sd arith, 0x01
+ld 0x01, 0x01
+sd arith, 0x01
+; :b [a, t, c, d] (b, 0, +)
+sd tape, 0x03
+gt 0x01, tape
+; :b t [a, t, c, d] (b, 0, +)
+sd tape, 0x01
+; :b t [a, b, c, d] ()
+gt 0x01, arith
+
+loop:
+; :b t c [a * c, b, 0, d] ()
+sd tape, 0x02
+sd tape, 0x03 ; :b t c d [a, b, c, d]
+sd tape, 0x00
+gt 0x03, tape ; :b t c d [a, b, c, a]
+
+multip:
+
+ sd arith, 0x02
+ ld 0x02, 0x01
+ sd arith, 0x02
+ ld 0x02, 0x02 ; sub 1 from c
+ sd arith, 0x02
+ gt 0x02, arith
+
+ sd arith, 0x00
+ sd arith, 0x03
+ ld 0x00, 0x01 ; a + a
+ sd arith, 0x00
+ gt 0x00, arith
+ sd tape, 0x00 ; :b t c d a+a [a+a, b, c-1, a]
 
 
-; routine for doing overflowing add
-; [a, b]
-; [255 - a, b]
-; 255 - a < b
-;    a = b - [255 - a]
-; 255 - a >= b
-;    a = a + b
-; 
+ ld 0x00, 0x01 ; :b t c d a+a [1, b, c-1, a]
+ le 0x02, 0x00, leavemul - s
+ gt 0x00, tape ; put back a + a
+
+ ; :b t c d [a+a, b, c-1, a]
+ le 0x00, 0x00, multip - s ; jump to mul
+
+leavemul:
+
+gt 0x00, tape
+gt 0x03, tape
+gt 0x02, tape
+; :b t [a * c, b, c, d] ()
+
+sd arith, 0x00
+sd arith, 0x01
+
+sd tape, 0x00
+ld 0x00, 0x01
+sd arith, 0x00
+gt 0x00, tape
+
+gt 0x00, arith
+; :b t [a * c + b, b, c, d] ()
+
+modulo:
+  jl 0x00, 0x03, leavemod - s
+
+  sd arith, 0x00
+  sd arith, 0x03
+
+  sd tape, 0x00
+  ld 0x00, 0x02
+  sd arith, 0x00
+  gt 0x00, tape
+
+  gt 0x00, arith
+  
+  le 0x00, 0x00, modulo - s
+
+leavemod:
+
+; :b t [(ac + b) % d, b, c, d] ()
+
+gt 0x01, tape
+; :b [(ac + b) % d, t, c, d] ()
+sd arith, 0x01
+ld 0x01, 0x01
+; :b [(ac + b) % d, 1, c, d] (t, )
+sd arith, 0x01
+ld 0x01, 0x02
+sd arith, 0x01
+; :b [(ac + b) % d, 1, c, d] (t, 1, -)
+gt 0x01, arith
+; :b [(ac + b) % d, t-1, c, d] ()
+sd tape, 0x02
+ld 0x02, 0x00
+le 0x01, 0x02, finish - s
+
+gt 0x02, tape
+; if t - 1 leq 0 -> jump out
+; else -> continue
+sd arith, 0x02
+; :b [(ac + b) % d, t-1, c, d] (c, ,)
+ld 0x02, 0x00
+sd arith, 0x02
+ld 0x02, 0x01
+sd arith, 0x02
+; :b [(ac + b) % d, t-1, 0, d] (c, 0, +)
+gt 0x02, tape
+; : [(ac + b) % d, t-1, b, d] (c, 0, +)
+sd tape, 0x02
+; :b [(ac + b) % d, t-1, b, d] (c, 0, +)
+sd tape, 0x01
+; :b t-1 [(ac + b) % d, t-1, b, d] (c, 0, +)
+sd tape, 0x02
+; :b t-1 b [(ac + b) % d, t-1, b, d] (c, 0, +)
+gt 0x02, arith
+; :b t-1 b [(ac + b) % d, t-1, c, d] ()
+gt 0x01, tape
+; :b t-1 [(ac + b) % d, b, c, d] ()
+le 0x03, 0x03, loop - s
+; if d leq d -> jump loop
+
+finish: 
+gt 0x02, tape
+gt 0x01, tape
 
 
-
-le 0x00, 0x01, label
 sd disp, 0x00
-label: 
-    sd disp, 0x01
+sd disp, 0x01
+sd disp, 0x02
+sd disp, 0x03
+;   finish things
+
+
+
+
+
+
+;le 0x00, 0x01, label
+;sd disp, 0x00
+;label: 
+;    sd disp, 0x01
diff --git a/src/device.rs b/src/device.rs
index f72b0ec..2c7fa18 100644
--- a/src/device.rs
+++ b/src/device.rs
@@ -6,19 +6,21 @@ pub trait Device {
     fn send_byte(&mut self, byte: u8);
 }
 
-
 pub struct Reader<'a> {
     pub reader: &'a mut dyn Read,
 }
 
-
 impl Device for Reader<'_> {
     fn get_byte(&mut self) -> u8 {
-            self.reader.bytes().next().and_then(|result| result.ok()).map(|byte| byte as u8).unwrap()
+        self.reader
+            .bytes()
+            .next()
+            .and_then(|result| result.ok())
+            .map(|byte| byte as u8)
+            .unwrap()
     }
 
-    fn send_byte(&mut self, _byte: u8) {
-    }
+    fn send_byte(&mut self, _byte: u8) {}
 }
 
 pub struct Tape {
@@ -28,13 +30,12 @@ pub struct Tape {
 
 impl Device for Tape {
     fn get_byte(&mut self) -> u8 {
-        let i = self.tape[self.index];
         if self.index == 0 {
             self.index = 25000;
         }
         self.index -= 1;
         self.index %= 25000;
-        i
+        self.tape[self.index]
     }
 
     fn send_byte(&mut self, byte: u8) {
@@ -42,7 +43,6 @@ impl Device for Tape {
         self.index += 1;
         self.index %= 25000;
     }
-
 }
 
 pub struct Writer<'a> {
@@ -67,10 +67,10 @@ pub struct Arithmetic {
 impl Device for Arithmetic {
     fn get_byte(&mut self) -> u8 {
         match self.buffer[2] {
-            1 => self.buffer[0] + self.buffer[1],
-            2 => self.buffer[0] - self.buffer[1],
+            1 => self.buffer[0].overflowing_add(self.buffer[1]).0,
+            2 => self.buffer[0].overflowing_sub(self.buffer[1]).0,
             3 => self.buffer[0] ^ self.buffer[1],
-            _ => 0
+            _ => 0,
         }
     }
 
@@ -80,5 +80,3 @@ impl Device for Arithmetic {
         self.index %= 3;
     }
 }
-
-
diff --git a/src/main.rs b/src/main.rs
index 85b7777..e14aba4 100644
--- a/src/main.rs
+++ b/src/main.rs
@@ -3,57 +3,76 @@
 mod device;
 mod vm;
 
-use device::{Reader, Arithmetic, Tape, Writer};
-use std::io::Cursor;
+use device::{Arithmetic, Reader, Tape, Writer};
 use rand::prelude::random;
-use std::io::Read;
+use std::fs;
 use std::io;
-use vm::{State, vm_run};
-
+use std::io::Cursor;
+use std::io::Read;
+use vm::{vm_run, State};
 
 fn main() {
+    let program: Vec<u8> = io::stdin()
+        .bytes()
+        .map(|res| res.unwrap_or_else(|_| std::process::exit(1)))
+        .collect();
 
-     let program: Vec<u8> = io::stdin().bytes().map(|res| res.unwrap_or_else(|_| std::process::exit(1))).collect();
-
-
-     for _ in 0..10 {
-         let iv: [u8; 4] = random();
-         let mut input = Cursor::new(iv);
-         let mut output = Cursor::new(vec![0,0,0,0]);
-
-         let state = State {
-             ip: 0usize,
-             mem: [0; 4],
-             program: program.clone(),
-             devices: vec![Box::new(Reader { reader: &mut input } ), 
-                           Box::new(Writer { writer: &mut output } ), 
-                           Box::new(Arithmetic { buffer: [0; 3], index: 0usize }),
-                           Box::new(Tape { tape: [0; 25000], index: 0usize }),]
-         };
-         vm_run(state);
-         let ov: [u8; 4] = output.into_inner().try_into().unwrap();
-         println!("{:?}", iv);
-         println!("{:?}", encrypt(iv));
-         println!("{:?}", ov);
-         if ov != encrypt(iv) {
-             std::process::exit(1);
-         }
-     }
+    for _ in 0..10 {
+        let iv: [u8; 4] = random();
+        let mut input = Cursor::new(iv);
+        let mut output = Cursor::new(vec![0, 0, 0, 0]);
 
+        let state = State {
+            ip: 0usize,
+            mem: [0; 4],
+            program: program.clone(),
+            devices: vec![
+                Box::new(Reader { reader: &mut input }),
+                Box::new(Writer {
+                    writer: &mut output,
+                }),
+                Box::new(Arithmetic {
+                    buffer: [0; 3],
+                    index: 0usize,
+                }),
+                Box::new(Tape {
+                    tape: [0; 25000],
+                    index: 0usize,
+                }),
+            ],
+        };
+        vm_run(state);
+        let ov: [u8; 4] = output.into_inner().try_into().unwrap();
+        if ov != encrypt(iv) {
+            std::process::exit(1);
+        }
+    }
+    println!(
+        "{}",
+        fs::read_to_string("flag.txt").unwrap_or_else(|_| {
+            println!("Couldn't open flag file, please contact an admin if on the server.");
+            std::process::exit(1)
+        })
+    );
 }
 
-
+#[inline(never)]
 fn encrypt(block: [u8; 4]) -> [u8; 4] {
     let mut block = block;
 
+    #[allow(clippy::clone_on_copy)]
+    let mut b = block.clone();
+
     let (i, _) = block[0].overflowing_add(block[3]);
-    block[1] ^= i;
+    b[1] ^= i;
     let (i, _) = block[1].overflowing_add(block[0]);
-    block[2] ^= i;
+    b[2] ^= i;
     let (i, _) = block[2].overflowing_add(block[1]);
-    block[3] ^= i;
+    b[3] ^= i;
     let (i, _) = block[3].overflowing_add(block[2]);
-    block[0] ^= i;
+    b[0] ^= i;
+
+    block = b;
 
     for _ in 0..block[3] {
         let (mul, _) = block[2].overflowing_mul(block[0]);
diff --git a/src/vm.rs b/src/vm.rs
index 0aab1f6..1e18191 100644
--- a/src/vm.rs
+++ b/src/vm.rs
@@ -5,7 +5,7 @@ pub struct State<'a> {
     pub ip: usize,
     pub mem: [u8; 4],
     pub program: Vec<u8>,
-    pub devices: Vec<Box<dyn Device + 'a>>
+    pub devices: Vec<Box<dyn Device + 'a>>,
 }
 
 impl State<'_> {
@@ -25,10 +25,10 @@ impl State<'_> {
 
     #[inline(never)]
     fn next_dword(&mut self) -> u32 {
-        let result = (u32::from(self.program[self.ip]) << 24) | 
-            (u32::from(self.program[self.ip + 1]) << 16) | 
-            (u32::from(self.program[self.ip + 2]) << 8) | 
-            (u32::from(self.program[self.ip + 3]));
+        let result = (u32::from(self.program[self.ip]) << 24)
+            | (u32::from(self.program[self.ip + 1]) << 16)
+            | (u32::from(self.program[self.ip + 2]) << 8)
+            | (u32::from(self.program[self.ip + 3]));
         self.ip += 4;
         result
     }
@@ -43,10 +43,8 @@ pub enum Opcode {
     Get,
     Less,
     LessEq,
-
 }
 
-
 impl From<u8> for Opcode {
     fn from(v: u8) -> Self {
         match v {
@@ -57,18 +55,16 @@ impl From<u8> for Opcode {
             0x21 => Opcode::Get,
             0x50 => Opcode::Less,
             0x51 => Opcode::LessEq,
-            _ => Opcode::Halt
+            _ => Opcode::Halt,
         }
     }
 }
 
-
 pub fn vm_run(mut state: State) -> i32 {
     loop {
         if state.ip >= state.program.len() {
-            return 1
+            return 1;
         }
-        
         match Opcode::from(state.program[state.ip]) {
             Opcode::Nop => {
                 state.ip += 1;
@@ -100,7 +96,7 @@ pub fn vm_run(mut state: State) -> i32 {
                 state.ip += 1;
                 let mem1 = state.mem[state.next_byte() as usize];
                 let mem2 = state.mem[state.next_byte() as usize];
-                let jump = state.next_byte() as usize;
+                let jump = state.next_word() as usize;
                 if mem1 < mem2 {
                     state.ip = jump;
                 }
@@ -109,12 +105,11 @@ pub fn vm_run(mut state: State) -> i32 {
                 state.ip += 1;
                 let mem1 = state.mem[state.next_byte() as usize];
                 let mem2 = state.mem[state.next_byte() as usize];
-                let jump = state.next_byte() as usize;
+                let jump = state.next_word() as usize;
                 if mem1 <= mem2 {
                     state.ip = jump;
                 }
             }
-
         }
     }
 }