diff --git a/kidlisp/PIECE-IL.md b/kidlisp/PIECE-IL.md
index dbd23a7fb6..d6a97ce2bb 100644
--- a/kidlisp/PIECE-IL.md
+++ b/kidlisp/PIECE-IL.md
@@ -247,3 +247,26 @@ output stay in the closure compiler. Kernels take the parts that are
 already maps over data, which the census says is most of the arithmetic in
 the two big pieces: projection, star fields, the skirt's bands, the
 soldier's boxes.
+
+## 9. The frame as one buffer: drawing on the GPU
+
+A compiled piece that says `; @gpu` does not call the software rasterizer.
+Its drawing heads record into a frame buffer (`lib/gpu-frame.mjs`): clear,
+line, box, oval, tri and shape, each with its colour, one Float32Array a
+frame, handed to bios in a single transferred message. bios tessellates it
+to triangles with per-vertex colour and draws it with WebGPU in one pass on
+its own canvas (`lib/gpu-frame-renderer.mjs`), then composites the CPU
+buffer over it as a texture when the frame asks, so `write` and anything
+the path does not take still show. `ink` keeps running on the CPU for its
+colour parsing and the resolved colour is read back; `wipe` clears both.
+The worker probes bios once; without a renderer the piece draws on the CPU
+as before, same stream.
+
+Measured 2026-10-11, phone size / 1280×720: Fía 36 / 21 fps from 31 / 20
+on the CPU path; the shooter 25 / 24 from 22 / 22. The pictures match. The
+format is the contract: a native host can consume the same buffer.
+
+What is left after this is evaluation and the runtime's own per-frame
+work. Kernels (§8) take the data-parallel arithmetic; the whole program to
+Wasm takes the rest, and a host that keeps kernel outputs on the GPU and
+draws from them never pays a readback at all.
diff --git a/kidlisp/reports/whistlegraph-benchmark-2026-10-10.md b/kidlisp/reports/whistlegraph-benchmark-2026-10-10.md
index f9b7cd237e..50534468a6 100644
--- a/kidlisp/reports/whistlegraph-benchmark-2026-10-10.md
+++ b/kidlisp/reports/whistlegraph-benchmark-2026-10-10.md
@@ -21,6 +21,8 @@ into pools.
 | Fía | compiler v3 + the heart projection as a Wasm kernel | 26,921 | 420 | 28 | 16 |
 | Fía | … and no HUD colouring for the hidden label | 26,921 | 420 | 31 | 20 |
 | Shooter | compiler v3, no HUD colouring | 22,380 | 383 | 22 | 22 |
+| Fía | … and the frame drawn by WebGPU (`; @gpu`) | 26,928 | 421 | 36 | 21 |
+| Shooter | … and the frame drawn by WebGPU (`; @gpu`) | 22,387 | 384 | 25 | 24 |
 
 The compiler rows are the same sources with a `; @compile` first line
 (`lib/kidlisp-compile.mjs`, later the same day): the program compiled once
@@ -48,6 +50,13 @@ is now skipped for a hidden label (339d671608). The star field in
 runtime (35 against 60 fps at phone size): a compute pass has to read its
 results back for a CPU rasterizer, and that costs more than the arithmetic.
 
+The `; @gpu` rows (8754e56644, PIECE-IL.md §9): the compiled program
+records its drawing into one buffer a frame (`lib/gpu-frame.mjs`) and bios
+draws it with WebGPU as one vertex buffer and one draw
+(`lib/gpu-frame-renderer.mjs`), the CPU buffer composited on top for text.
+The frames match the CPU path's picture. What remains is evaluation and the
+runtime's per-frame work around it; the rasterizer is off the clock.
+
 ## What it says
 
 - **Size is a wash.** Prefix arithmetic costs what closures saved. A line