1 files changed, 13 insertions, 0 deletions

diff --git a/docs/superpowers/plans/2026-10-20-corrections-and-signature-library.md b/docs/superpowers/plans/2026-10-20-corrections-and-signature-library.md
index 289ffc8..50f9ba7 100644
--- a/docs/superpowers/plans/2026-10-20-corrections-and-signature-library.md
+++ b/docs/superpowers/plans/2026-10-20-corrections-and-signature-library.md
@@ -224,6 +224,19 @@ the worst failure this system has; his "no" removes it rather than lowering a sc
 *inferring* what a stretch of audio is. He was there. So a confirmed signature is trusted at
 once, and the confirmation is never unset — he does not un-hear it next week.
 
+**A label is only actionable while its episode is still mirrored.** Labels deliberately outlive
+the audio they were given on, so the server holds plenty that name episodes rotated out of the
+eight-episode window. Those are simply not applied — there is no audio to harvest or disown. Found
+by pointing a test label at an episode chosen from the show's episode *list*, which holds hundreds
+of entries, when only eight have a `path`. For a daily show the window is about a week; labels
+arrive within minutes of being made, so this does not bite in practice.
+
+**Verified end to end on the live server, 2026-10-27.** A mark vouching for a real detected break
+flowed from `state.json` through the shadow run — matched by feed and guid, overlapped against the
+break, applied to the signature that break produced — and came out as `confirmed: true` in
+`signatures.json`. The probe label and the flag it set were both removed afterwards: that corpus
+holds the listener's statements, not test data.
+
 **A report with a guessed end is trimmed to its first 15 seconds.** An uncorrected end is
 wherever the jump landed, not where the advertising stopped, so its tail may be programme. One ad
 unit is far more audio than a landmark fingerprint needs and is the part most certainly