diff --git a/docs/architecture.md b/docs/architecture.md
index ebb1239f..309fe0be 100644
--- a/docs/architecture.md
+++ b/docs/architecture.md
@@ -105,8 +105,15 @@ owns the only round-1 threshold, and folding the two together would delete a com
 leaving both names in place.
 
 The stages are pinned by `tests/test_round1_preservation.py` against a frozen trace generated
-from an independent transcription of the pre-refactor matcher. The PDF path is **not** migrated;
-ADR 0020 shares the boundary rule with it, not the implementation.
+from an independent transcription of the pre-refactor matcher.
+
+**The PDF path runs the same four stages, and reached them separately.** It has its own retrieval,
+evidence, assignment and classification stages in `diff_pdf.py`, its own byte-identity gate
+(`tests/test_pdf_canonical_baseline.py`), and its own boundary tests. What the two paths share is
+the *rule* — retrieval controls consideration, assignment controls correspondence — not an
+implementation, and ADR 0020 deliberately does not unify them: an XML observation is a parsed tree
+node and a PDF observation is a reconstructed block, so a shared matcher would have to pick one
+representation and lose the other.
 
 ### Why the two paths exist at all
 
diff --git a/docs/decisions/0020-matching-stages.md b/docs/decisions/0020-matching-stages.md
index 14e96eac..7107a056 100644
--- a/docs/decisions/0020-matching-stages.md
+++ b/docs/decisions/0020-matching-stages.md
@@ -245,11 +245,15 @@ the contents of one stage change.
 - **Let classification re-consult similarity.** Rejected with a carve-out: classification
   legitimately asks *how much* the corresponding texts differ, which is what
   `move.body_unchanged` records and needs no score. What it may not do is threshold a
-  correspondence score. On the XML side that separation now holds: `diff_bill.diff_bills` is
-  orchestration over the four stages, and the only round-1 cutoff lives in
-  `apply_similarity_assignment_rule`, an assignment act that runs before classification.
-  `diff_pdf._hunk_for_paired_blocks` still thresholds inside classification; the PDF track is
-  not migrated, and this record does not unify the two implementations.
+  correspondence score. Both pipelines now hold that separation. On the XML side
+  `diff_bill.diff_bills` is orchestration over the four stages, and the only round-1 cutoff lives
+  in `apply_similarity_assignment_rule`, an assignment act that runs before classification. On the
+  PDF side `diff_pdf._hunk_for_paired_blocks` used to call a pair *moved* when the anchors differed
+  and the bodies cleared `MOVE_THRESHOLD` — a threshold over correspondence evidence, inside
+  classification. It no longer receives a similarity at all: the rule moved unchanged into
+  `pdf_round1_move_basis`, an assignment stage, and the function is now the `modified` emitter.
+  The two pipelines share this boundary rule; they do **not** share an implementation, and this
+  record does not unify them.
 - **Keep `Correspondence` pair-shaped, and revisit if consolidation proves common.** Cost
   asymmetry: capability costs a type permitting N sides, while not having it turns any later
   change into a migration through every consumer of the matcher's output.