diff --git a/DECISIONS/2026-11-02-pdf-library.md b/DECISIONS/2026-11-02-pdf-library.md
new file mode 100644
index 0000000..479830c
--- /dev/null
+++ b/DECISIONS/2026-11-02-pdf-library.md
@@ -0,0 +1,48 @@
+# Decision: PDF generation library
+
+Status:  Decided
+Date:    2026-11-02
+Review:  On trigger (see below), not scheduled
+
+## Choice
+
+**ReportLab** for all server-side PDF generation, starting with the
+Migration Verification Report (P1).
+
+## Reasoning
+
+Each trade-off was weighed against the project's actual constraints:
+
+- "Air-gap install" — ReportLab is one pure-Python wheel; WeasyPrint
+  needs Pango, Cairo and GDK-PixBuf system libraries vendored for
+  every target OS. #60 (offline install, due 15 Nov) is simpler
+  with ReportLab.
+- "V&V evidence" — ReportLab produces byte-identical PDFs across
+  runs. A report's SHA-256 is stable and quotable in audit logs.
+  WeasyPrint output can shift by a pixel between runs due to font
+  hinting and CSS layout; fine for most contexts, weaker for V&V.
+- "Deterministic toolchain" — matches the discipline this project
+  sells. A declared, provable output format is a stronger artefact
+  than a near-identical one.
+
+## Trade-offs accepted
+
+- Slower visual iteration on report templates (edit Python,
+  regenerate PDF, review).
+- Verbose templates: multi-section reports are hundreds of lines
+  of ReportLab primitives.
+- No HTML/CSS reuse with the frontend. If a frontend ever renders
+  a Migration Verification Report preview, the on-screen view and
+  the PDF will be two separate implementations.
+- Less accessible to a non-Python collaborator than HTML/CSS.
+
+## Reopen trigger
+
+Reconsider this decision if any of the following becomes true:
+
+- Reports become visually dynamic and need frequent layout iteration
+- The report format must match a frontend screen exactly
+- The air-gap constraint is dropped
+- A non-Python contributor needs to modify report templates regularly
+
+Tracked in #67.
\ No newline at end of file