diff --git a/go.mod b/go.mod
index 23d3d7a..37e2ba1 100644
--- a/go.mod
+++ b/go.mod
@@ -5,6 +5,7 @@ go 1.27.2
 require (
 	github.com/go-gfx/gfx v0.41.0
 	github.com/go-images/images v0.3.0
+	github.com/go-images/jpeg v0.3.0
 	github.com/go-ooxml/ooxml v0.7.0
 	github.com/go-ooxml/opc v0.2.1
 	github.com/go-pdfkit/document v0.2.0
@@ -28,7 +29,6 @@ require (
 	github.com/go-browserhttp/browserhttp v0.2.0 // indirect
 	github.com/go-gfx/oksvg v0.2.0 // indirect
 	github.com/go-images/gif v0.2.0 // indirect
-	github.com/go-images/jpeg v0.3.0 // indirect
 	github.com/go-images/jpeg2000 v0.14.0 // indirect
 	github.com/go-images/png v0.2.0 // indirect
 	github.com/go-opentype/fonts v0.14.0 // indirect
diff --git a/redactimage.go b/redactimage.go
index cac5df2..a7f2e3d 100644
--- a/redactimage.go
+++ b/redactimage.go
@@ -5,8 +5,21 @@ import (
 	"fmt"
 	"image"
 	"image/draw"
-	"image/jpeg"
 
+	// ⛔ THE FLEET'S DECODER, NOT THE STANDARD LIBRARY'S. What this promises to
+	// write back is "what a reader would have seen", and the reader is
+	// go-pdfkit/render, which decodes /DCTDecode with this fork: Go's own
+	// image/jpeg with a four-component picture's chroma upsampled the way
+	// libjpeg does rather than by repeating each sample. Decoding here with a
+	// different decoder would make that promise true of a picture nobody was
+	// shown.
+	//
+	// ⛔ AND IT IS NOT ABOUT THE BINARY. This was first written claiming it
+	// saved linking a second decoder; `go list -deps` says otherwise, before
+	// and after. image/jpeg stays, because go-gfx/gfx/codec, go-images/images
+	// and go-pdfkit/html2pdf each import it — go-images/images imports BOTH.
+	// That is a wider divergence than this file, filed rather than fixed here.
+	"github.com/go-images/jpeg"
 	"github.com/go-pdfkit/extract"
 	"github.com/go-pdfkit/reader"
 )
diff --git a/redactjpeg_test.go b/redactjpeg_test.go
index 575cd4e..7336326 100644
--- a/redactjpeg_test.go
+++ b/redactjpeg_test.go
@@ -130,7 +130,14 @@ var bottomQuarter = Redaction{Page: 1, Quads: []Quad{{
 //
 // ⛔⛔ THE KEPT SAMPLES ARE COMPARED AGAINST THIS TEST'S OWN DECODE OF THE
 // ORIGINAL JPEG, BYTE FOR BYTE. That is the whole claim the lossless path
-// makes. Asserting a colour instead would be asserting what an encoder does
+// makes.
+//
+// ⛔ AND THE TWO DECODERS ARE NOT THE SAME ONE. This file encodes and decodes
+// with the standard library's image/jpeg; redactimage.go decodes with
+// go-images/jpeg, the fork go-pdfkit/render shows the reader. So the equality
+// below is also the statement that the two agree on a three-component picture
+// — and the day they stop agreeing, this goes red and names the row and column
+// where. Importing the fork here as well would have thrown that away. Asserting a colour instead would be asserting what an encoder does
 // with it, which is a different subject and a moving one; and asserting "it
 // still looks about the same" would pass for a picture re-encoded into mud —
 // which is what re-encoding as a JPEG would have done to every block of it,