diff --git a/AGENTS.md b/AGENTS.md
index 705a047..0b51199 100644
--- a/AGENTS.md
+++ b/AGENTS.md
@@ -682,6 +682,101 @@ of a hundred thousand times, and the timings a checked run prints are labelled
 as meaningless. **Do not add memory-check suites**: a new path worth checking
 belongs in a benchmark, where it gets both.
 
+## Testing a Tutorial
+
+`ctest -R docs_examples` proves every listing in `docs/` still compiles and
+still matches the file it was quoted from. It cannot prove the chapter *teaches*
+anything: a document can be composed entirely of verified excerpts and still be
+unfollowable, because what a reader needs is the glue between them.
+
+**So test a tutorial by making somebody follow it.** Hand it to a subagent on a
+weaker model with no context beyond the chapter, and have them build the thing it
+describes. The weaker model is the point -- it will not paper over a gap with
+knowledge the document did not give it, which is exactly the failure mode a
+capable reviewer has.
+
+### The rules that make the result mean something
+
+- **Give them what a real reader has, and nothing more.** The chapter, the
+  library source, its public headers, and the art. **Not the finished program.**
+  Copy the tree with `examples/` and `docs/` removed rather than telling them not
+  to look -- a rule they can break is not a control.
+- **Headers are fair game.** The manual defers signatures to Doxygen and a real
+  user has `include/akgl/*.h` in front of them. Forbidding those tests a reader
+  who does not exist.
+- **They must compile it and run it.** This is the whole thing. A design nobody
+  built proves nothing, and the defects that matter most do not fail to compile.
+- **Verify their work yourself.** Run the binary they produced. Look at the
+  screenshot they took. Do not take a report's word for what it built --
+  see "Reading the report" below.
+- **Do not make them author assets by hand.** A real reader draws a map in Tiled;
+  hand-writing a `.tmj` tests nothing about the chapter and eats the whole
+  session. Give them `docs/tutorials/assets/`. The asset *formats* are already
+  covered by `docs_examples`, which runs the chapter's own JSON blocks through
+  `akgl_sprite_load_json` and `akgl_character_load_json`.
+- **Give them a screenshot helper**, marked as scaffolding and not part of the
+  tutorial, so there is a picture to check. Screen output is evidence; an exit
+  status of 0 is not.
+
+### Setting the sandbox up
+
+One directory per reader, because two of them building at once in the same tree
+collide:
+
+```sh
+rsync -a --exclude='.git' --exclude='build' --exclude='examples' --exclude='docs' \
+      . "$SB/reader/libakgl/"
+cp docs/20-tutorial-sidescroller.md "$SB/reader/TUTORIAL.md"
+cp -r docs/tutorials/assets/sidescroller "$SB/reader/art"
+mkdir -p "$SB/reader/game"
+```
+
+A consumer using the CMake the chapter itself teaches --
+`add_subdirectory(../libakgl libakgl)` plus the documented
+`target_link_libraries` line -- configures and builds in about a minute from
+cold, and seconds after that. Tell them to use
+`cmake --build build --target <theirs> --parallel`, or they will also build
+every test suite in the tree.
+
+**Prove the path works before you hand it over.** Write a throwaway consumer
+that opens a window and takes a screenshot, build it, run it, and delete it. A
+reader who cannot build is a reader who finds nothing, and the fault will be
+yours -- the screenshot helper's first draft included `akgl/renderer.h` for
+`akgl_renderer`, which is declared in `akgl/game.h`, and would have cost them
+the session.
+
+### Reading the report
+
+**Every finding is a hypothesis until you check it.** A weaker model reports its
+own mistakes as documentation defects with total confidence, and both kinds are
+worth having -- but only one is worth acting on.
+
+- Reproduce the failure before believing it. One reader reported the dialogue
+  freeze broken; it had drawn its own map with the NPC out of reach and never
+  used the one it was given. Running the reference showed the player moving zero
+  pixels through the whole freeze window.
+- A rejected finding usually still leaves something. That same map failure was
+  not a code defect, but it did show the chapter stated `TALK_RANGE` without
+  saying what it means when you are placing NPCs -- and that a too-distant NPC
+  produces *success*, not an error.
+- Fix the document, not the reader. If they guessed a header name, the chapter
+  never named it.
+
+### What this catches that review does not
+
+Four read-only passes over the two tutorials in `docs/` found real gaps and
+missed all three of these, which the first build-and-run found immediately:
+
+| Defect | Why reading missed it |
+|---|---|
+| libakstdlib's header was never named -- `aksl_*` functions, `<akstdlib.h>` file | Every reviewer already knew, or did not have to write the `#include` |
+| The asset directory defaulted to `"."` instead of the compiled-in macro | Prose said "falls back", and nobody had to run it from another directory |
+| A published example output the chapter could not produce | The capture happened after teardown released the pool. Nothing fails to compile |
+
+The third is the shape to remember: **a tutorial that prints an expected result
+is making a claim, and a claim nobody executed is a claim that is probably
+wrong.**
+
 ## Commit & Pull Request Guidelines
 
 Recent commits use concise, imperative summaries such as `Fix a scale bug...` and often explain related changes in one sentence. Keep each commit scoped and describe user-visible behavior. Pull requests should summarize the change, identify affected modules, list the CMake/CTest commands run, and link relevant issues. Include screenshots only for rendering, map, or `charviewer` changes.
diff --git a/docs/MAINTENANCE.md b/docs/MAINTENANCE.md
index f11e6b4..70a5491 100644
--- a/docs/MAINTENANCE.md
+++ b/docs/MAINTENANCE.md
@@ -4,6 +4,13 @@ This file documents the harness that keeps `docs/` honest. It is not a chapter 
 a reader learning libakgl never needs it — and it is checked by the same harness
 it describes, so the file that publishes the convention is held to it.
 
+**The harness checks that every listing is true. It cannot check that a chapter
+teaches.** For that, the tutorials get a second test: the chapter is handed to a
+subagent on a weaker model with the library, the headers and the art but *not*
+the finished program, and that reader has to build and run the game. See
+`AGENTS.md`, "Testing a Tutorial". Four read-only reviews of chapters 20 and 21
+missed three defects the first build-and-run found in minutes.
+
 ## Documentation examples
 
 **Every fenced block in `docs/*.md` is compiled, linked, run, or matched against