diff --git a/harness/reviews/state-label-evidence-20260924.md b/harness/reviews/state-label-evidence-20260924.md
new file mode 100644
index 00000000..3ff64837
--- /dev/null
+++ b/harness/reviews/state-label-evidence-20260924.md
@@ -0,0 +1,39 @@
+# State label evidence readback
+
+Codex-ytr83r reviewed the published State group captures for issues36/37 after
+network access recovered. This is evidence review, not a layout fix or acceptance
+of the integrated prerelease. Implementation remains deferred behind issue37.
+
+Source evidence: commit `bed7f825fd00ec32299032a664a5f5d753aad86d`,
+`harness/experiments/state-multiselect-20260923/native/evidence.zip`.
+The retrieved archive is 3,384,779 bytes, SHA-256
+`1d08c1bc6af4856a569ddd320ddf5403a302b8d9437750663a498541e489d663`,
+Git blob `e1843b5772653b2bff3308b544400cccbbfffbfc`. One bounded contents-API
+retrieval completed in 2.09 seconds; no Git fetch or application execution was
+needed for this review.
+
+The archive index was checked for duplicate, traversal, absolute and symlink
+entries. Only these members were extracted and checked against the published
+manifest. A fresh local critic and the parent each viewed all three original
+1442 by 932 JPEGs:
+
+| Archive member | SHA-256 |
+| --- | --- |
+| `runs/run-1/raw/05-a-selected-0.jpg` | `a1b3da460a3e3aeaf611321bc2436dd0551463dce0b03a28ef60d509daec671c` |
+| `runs/run-1/raw/12-placement-group-0.jpg` | `ffdd50679ebf1f4ef581e943e4c8edd4e0170d721a38e1b2e4f9090c570d58a4` |
+| `runs/run-1/raw/13-group-moved-0.jpg` | `fa65ffefd695b4b5aef67c22bf23148906182c769e0e512e98a1f7c32435035f` |
+
+At 79% zoom, the A-to-B transition label starts with a legible `Ready` and has
+colliding lower text runs. The collision appears with A selected, with A/B
+selected together, and after the captured group movement. The neighboring A/B
+state titles remain individually legible. The transition text extends across
+A's boundary and occupies the vertical band of its entry action, but that action
+text is farther left: these pixels do not establish glyph overlap with the
+`entry / Record request` text.
+
+The exact collided text cannot be transcribed reliably. Identifying its two
+runs as the multiline title and secondary summary relies on the peer's source
+report; the images independently establish the readability defect. They do not
+establish its algorithmic cause, other zoom/DPI behavior, a correction, or complete
+group-edit correctness. The prerelease's existing multiline State-label limitation
+therefore remains warranted.