diff --git a/2021/src/common.rs b/2021/src/common.rs
index 487b8eb..600cd04 100644
--- a/2021/src/common.rs
+++ b/2021/src/common.rs
@@ -90,7 +90,14 @@ where
     let mut buffer = Vec::new();
     input.read_to_end(&mut buffer).unwrap();
 
-    let (_, output) = parser(&buffer).finish().unwrap();
-
-    output
+    match parser(&buffer).finish() {
+        Ok((_, output)) => output,
+        Err(err) => {
+            panic!(
+                "Failed to parse input with error {:?} at \"{}\"",
+                err.code,
+                String::from_utf8_lossy(err.input)
+            );
+        }
+    }
 }
diff --git a/2021/src/day24.rs b/2021/src/day24.rs
index 113ba49..6646653 100644
--- a/2021/src/day24.rs
+++ b/2021/src/day24.rs
@@ -1,9 +1,159 @@
+//! Very input-specific reverse-engineered solution
+//!
+//! # General implementation
+//!
+//! The code in the examples is a series of 14 times this:
+//!
+//! ```txt
+//! inp w       -> read digit
+//! mul x 0
+//! add x z
+//! mod x 26    -> x = z % 26
+//! div z $A    -> pop Z (see below)
+//! add x $B
+//! eql x w     -> x = ((z + $B) == w)
+//! eql x 0     -> x = ((z + $B) != w)
+//! mul y 0
+//! add y 25
+//! mul y x
+//! add y 1     -> if x { 26 } else { 1 }
+//! mul z y     -> if x { z *= 26 } (push z, see below)
+//! mul y 0
+//! add y w
+//! add y $C    -> y = w + $C
+//! mul y x
+//! add z y     -> if x { z += w + $C }
+//! ```
+//!
+//! `$A` is either `1` or `26` which we can translate to a bool `$A == 26` for convenience. This
+//! simplifies to the following rust.
+//!
+//! ```
+//! fn validate<const A: bool, const B: i32, const C: i32>(mut z: i32, digit: i32) -> i32 {
+//!     let x = (z % 26 + B) != digit;
+//!     if A {
+//!         z /= 26;
+//!     }
+//!
+//!     if x {
+//!         z = 26 * z + digit + C;
+//!     }
+//!
+//!     z
+//! }
+//! ```
+//!
+//! In human terms, `z` is used to hold a base 26 number. When `$A` is `true`, we pop off the least
+//! significant digit by dividing by 26. Then, depending on whether `(z + $B) % 26` is equal to our
+//! digit, we push `digit + $C`. Ideally, we should pop as often as we push in order to arrive at `z
+//! == 0` in the end. The input contains 7 pops, so we want each of those to not push.
+//!
+//! To solve this problem, we use a backtracking memoizing algorithm, where we cancel every sequence
+//! that fails to pop at some point. A pop is failed whenever we execute a validation pop where we
+//! can pop if `x` happened to be set to `0` at the time of the check. We can memoize this over the
+//! running value of `z`.
+//!
+//! This implementation probably doesn't work on other people's input; to fix it, you'll want to
+//! update the `parse_step` function. Good luck with that.
+use std::collections::HashMap;
 use std::io::Read;
 
-pub fn part1(_input: &mut dyn Read) -> String {
-    todo!()
+use nom::branch::alt;
+use nom::bytes::complete::tag;
+use nom::character::complete::newline;
+use nom::combinator::map;
+use nom::multi::separated_list1;
+use nom::sequence::delimited;
+use nom::sequence::preceded;
+use nom::sequence::tuple;
+use nom::IResult;
+
+use crate::common::read_input;
+
+type Cache = HashMap<(usize, i32), Option<i64>>;
+
+#[derive(Debug)]
+struct Step(bool, i32, i32);
+
+impl Step {
+    fn evaluate(&self, digit: i32, z: i32) -> Option<i32> {
+        if self.0 {
+            (z % 26 + self.1 == digit).then(|| z / 26)
+        } else {
+            Some(z * 26 + digit + self.2)
+        }
+    }
 }
 
-pub fn part2(_input: &mut dyn Read) -> String {
-    todo!()
+fn parse_step(input: &[u8]) -> IResult<&[u8], Step> {
+    use nom::character::complete::i32;
+
+    let parse_pop = preceded(
+        tag("inp w\nmul x 0\nadd x z\nmod x 26\ndiv z "),
+        alt((map(tag("1"), |_| false), map(tag("26"), |_| true))),
+    );
+
+    let parse_a = preceded(tag("\nadd x "), i32);
+
+    let parse_b = delimited(
+        tag("\neql x w\neql x 0\nmul y 0\nadd y 25\nmul y x\nadd y 1\nmul z y\nmul y 0\nadd y w\nadd y "),
+        i32,
+        tag("\nmul y x\nadd z y"),
+    );
+
+    map(tuple((parse_pop, parse_a, parse_b)), |(pop, a, b)| {
+        Step(pop, a, b)
+    })(input)
+}
+
+fn parse_input(input: &[u8]) -> IResult<&[u8], Vec<Step>> {
+    separated_list1(newline, parse_step)(input)
+}
+
+fn optimize(
+    current: usize,
+    steps: &[Step],
+    digits: &[i32],
+    z: i32,
+    cache: &mut Cache,
+) -> Option<i64> {
+    if current >= steps.len() {
+        return (z == 0).then(|| 0);
+    }
+
+    if let Some(&memoized) = cache.get(&(current, z)) {
+        return memoized;
+    }
+
+    let result = digits.iter().find_map(|&digit| {
+        let z = steps[current].evaluate(digit, z)?;
+        let result = optimize(current + 1, steps, digits, z, cache)?;
+
+        Some(result + digit as i64 * 10i64.pow(13 - current as u32))
+    });
+    cache.insert((current, z), result);
+
+    result
+}
+
+fn parts_common(input: &mut dyn Read, digits: &[i32]) -> String {
+    let steps = read_input(input, parse_input);
+
+    let mut cache = Cache::new();
+
+    optimize(0, &steps, digits, 0, &mut cache)
+        .unwrap()
+        .to_string()
+}
+
+pub fn part1(input: &mut dyn Read) -> String {
+    let digits = [9, 8, 7, 6, 5, 4, 3, 2, 1];
+
+    parts_common(input, &digits)
+}
+
+pub fn part2(input: &mut dyn Read) -> String {
+    let digits = [1, 2, 3, 4, 5, 6, 7, 8, 9];
+
+    parts_common(input, &digits)
 }