diff --git a/README.md b/README.md
index ab1df7c..008ffeb 100644
--- a/README.md
+++ b/README.md
@@ -51,6 +51,8 @@ Drop your own `.md` files in `notes/` and ask again.
 | `cli.py` | Question in → answer + citations out |
 | `notes/` | Sample teaching notes (git, Ollama, RAG, pytest) |
 
+See [how scoring works](docs/how-scoring-works.md) for a hand-worked cosine example.
+
 ## Change one thing
 
 1. Add `notes/my-topic.md` and ask a question only that file can answer  
diff --git a/docs/how-scoring-works.md b/docs/how-scoring-works.md
new file mode 100644
index 0000000..52081b1
--- /dev/null
+++ b/docs/how-scoring-works.md
@@ -0,0 +1,67 @@
+# How scoring works
+
+The scorer compares token frequencies, not meanings. Here is a tiny example you
+can check by hand against [`retrieve.py`](../src/notes_rag/retrieve.py).
+
+## 1. Turn text into tokens
+
+Query: `Git remote?` Two Markdown notes:
+
+```markdown
+# Git
+git remote
+```
+
+```markdown
+# Remote
+remote remote
+```
+
+The title **and** body are scored. Text is lowercased, then `[a-z0-9]+` matches
+runs of ASCII letters or digits. Punctuation separates tokens; it is not kept.
+The query becomes `[git, remote]`, note A `[git, git, remote]`, and note B
+`[remote, remote, remote]`.
+
+## 2. Count and normalize
+
+Term frequency is each token's count divided by the total number of tokens.
+Using the coordinate order `(git, remote)`:
+
+| Text | Counts | Term-frequency vector |
+| --- | --- | --- |
+| Query | `(1, 1)` | `(1/2, 1/2)` |
+| A: Git | `(2, 1)` | `(2/3, 1/3)` |
+| B: Remote | `(0, 3)` | `(0, 1)` |
+
+Word order does not affect these vectors. Repeating only `remote` still gives
+note B a frequency of 1 for that token.
+
+## 3. Compute cosine similarity
+
+Multiply matching coordinates and add them for the dot product. The norm is
+the square root of the sum of squared coordinates. Then:
+
+```text
+score = dot(query, note) / (norm(query) * norm(note))
+
+norm(query) = sqrt(1/4 + 1/4) = sqrt(2)/2
+
+A: dot = (1/2 * 2/3) + (1/2 * 1/3) = 1/2
+   norm = sqrt(4/9 + 1/9) = sqrt(5)/3
+   score = (1/2) / (sqrt(2)/2 * sqrt(5)/3) = 3/sqrt(10) = 0.948683...
+
+B: dot = (1/2 * 0) + (1/2 * 1) = 1/2
+   norm = sqrt(0 + 1) = 1
+   score = (1/2) / (sqrt(2)/2 * 1) = 1/sqrt(2) = 0.707107...
+```
+
+## 4. Rank and keep positive hits
+
+Scores are sorted highest first. With `k=2`, A comes before B; the CLI displays
+their scores as `0.949` and `0.707`. With `k=1`, only A remains.
+The scorer takes the first `k` entries, then keeps only scores greater than 0.
+A query such as `pytest` shares no tokens with either note: both dot products
+are 0, so neither note is returned, even with `k=2`.
+
+Raising a positive `--k` lets more lower-ranked, positive-score notes into the
+answer's context; it does not change their scores or admit zero-score notes.