diff --git a/yescrypt/README.md b/yescrypt/README.md
index b9d08b4c..e0151d46 100644
--- a/yescrypt/README.md
+++ b/yescrypt/README.md
@@ -15,6 +15,8 @@ yescrypt is a variant of the [scrypt] password-based key derivation function and
 [Password Hashing Competition]. It has been adopted by several Linux distributions for the system
 password hashing function, including Fedora, Debian, Ubuntu, and Arch.
 
+The algorithm is described in [yescrypt - a Password Hashing Competition submission][paper].
+
 ## ⚠️ Security Warning
 
 The implementation contained in this crate has never been independently audited!
@@ -64,3 +66,4 @@ dual licensed as above, without any additional terms or conditions.
 [yescrypt]: https://www.openwall.com/yescrypt/
 [scrypt]: https://en.wikipedia.org/wiki/Scrypt
 [Password Hashing Competition]: https://www.password-hashing.net/
+[paper]: https://www.password-hashing.net/submissions/specs/yescrypt-v2.pdf
diff --git a/yescrypt/src/flags.rs b/yescrypt/src/flags.rs
index 99e22ead..8fe30472 100644
--- a/yescrypt/src/flags.rs
+++ b/yescrypt/src/flags.rs
@@ -10,6 +10,12 @@ const MODE_MASK: u32 = 0x3;
 const RW_FLAVOR_MASK: u32 = 0x3fc;
 
 /// Flags for selecting the "flavor" of `yescrypt`.
+///
+/// A bitmask allowing to enable the individual extra features of yescrypt, e.g:
+///
+/// - [`Flags::RW`]: yescrypt’s native mode. Call [`Flags::default`] to fully configure this.
+/// - [`Flags::WORM`]: conservative enhancement of classic scrypt. Mutually exclusive with `RW`.
+/// - [`Flags::EMPTY`]: requests classic scrypt.
 #[derive(Clone, Copy, Debug, Eq, PartialEq)]
 pub struct Flags(pub(crate) u32);
 
diff --git a/yescrypt/src/params.rs b/yescrypt/src/params.rs
index 68e0af19..1bd1949c 100644
--- a/yescrypt/src/params.rs
+++ b/yescrypt/src/params.rs
@@ -20,6 +20,9 @@ pub struct Params {
     pub(crate) flags: Flags,
 
     /// `N`: CPU/memory cost (like `scrypt`).
+    ///
+    /// yescrypt, including in scrypt compatibility mode, is defined only for values of N that are
+    /// powers of 2 (and larger than 1, which matches scrypt’s requirements).
     pub(crate) n: u64,
 
     /// `r`: block size (like `scrypt`).
@@ -28,10 +31,14 @@ pub struct Params {
     /// `p`: parallelism (like `scrypt`).
     pub(crate) p: u32,
 
-    /// special to yescrypt.
+    /// Controls yescrypt’s computation time while keeping its peak memory usage the same.
+    ///
+    /// `t = 0` is optimal for achieving the highest normalized area-time cost for ASIC attackers.
     pub(crate) t: u32,
 
-    /// special to yescrypt.
+    /// The number of cost upgrades performed to the hash so far.
+    ///
+    /// `0` means no upgrades yet, and is currently the only allowed value.
     pub(crate) g: u32,
 
     /// special to yescrypt.
diff --git a/yescrypt/src/pwxform.rs b/yescrypt/src/pwxform.rs
index 49629823..520d7ef9 100644
--- a/yescrypt/src/pwxform.rs
+++ b/yescrypt/src/pwxform.rs
@@ -1,17 +1,31 @@
-//! pwxform: parallel wide transformation
+//! pwxform stands for "parallel wide transformation", although it can as well be tuned to be as
+//! narrow as one 64-bit lane.
+//!
+//! It operates on 64-bit lanes which are designed to be grouped into wider "simple SIMD" lanes,
+//! which are in turn possibly grouped into an even wider "gather SIMD" vector.
 
 use crate::{
     salsa20,
     util::{slice_as_chunks_mut, xor},
 };
 
-// These are tunable, but they must meet certain constraints.
+/// Number of 64-bit lanes per "simple SIMD" lane (requiring only arithmetic and bitwise operations
+/// on its 64-bit elements). Must be a power of 2.
 const PWXSIMPLE: usize = 2;
+
+/// Number of parallel "simple SIMD" lanes per "gather SIMD" vector (requiring "S-box lookups" of
+/// values as wide as a "simple SIMD" lane from PWXgather typically non-contiguous memory
+/// locations). Must be a power of 2.
 const PWXGATHER: usize = 4;
+
+/// Number of sequential rounds of pwxform’s basic transformation. Must be a power of 2, plus 2
+/// (e.g. 3, 4, 6, 10).
 const PWXROUNDS: usize = 6;
+
+/// Number of S-box index bits, thereby controlling the size of each of pwxform’s two S-boxes
+/// (in "simple SIMD" wide elements).
 const SWIDTH: usize = 8;
 
-// Derived values.  Not tunable on their own.
 const PWXBYTES: usize = PWXGATHER * PWXSIMPLE * 8;
 const PWXWORDS: usize = PWXBYTES / size_of::<u32>();
 const SMASK: usize = ((1 << SWIDTH) - 1) * PWXSIMPLE * 8;
@@ -28,9 +42,13 @@ pub(crate) struct PwxformCtx<'a> {
 }
 
 impl PwxformCtx<'_> {
-    /// Compute `B = BlockMix_pwxform{salsa20/2, ctx, r}(B)`.
+    /// Compute `B = BlockMix_pwxform{salsa20/2, ctx, r}(B)`. Input `B` must be 128 bytes in length.
     ///
-    /// The input `B` must be 128r bytes in length.
+    /// `BlockMix_pwxform` differs from scrypt’s `BlockMix` in that it doesn’t shuffle output
+    /// sub-blocks, uses pwxform in place of Salsa20/8 for as long as sub-blocks processed with
+    /// pwxform fit in the provided block B, and finally uses Salsa20/2 (that is, Salsa20 with only
+    /// one double-round) to post-process the last sub-block output by pwxform (thereby finally
+    /// mixing pwxform’s parallel lanes).
     pub(crate) fn blockmix_pwxform(&mut self, b: &mut [u32], r: usize) {
         // Convert 128-byte blocks to PWXbytes blocks
         // TODO(tarcieri): use upstream `[T]::as_chunks_mut` when MSRV is 1.88