diff --git a/documentation/audits/DRILL-chaos-night-2026-09-17.md b/documentation/audits/DRILL-chaos-night-2026-09-17.md
index d361e3af..8d9bf0a1 100644
--- a/documentation/audits/DRILL-chaos-night-2026-09-17.md
+++ b/documentation/audits/DRILL-chaos-night-2026-09-17.md
@@ -408,6 +408,13 @@ reconciles)"). **This is not the event path.** A dropped event is a lost *fact*,
 a picture that will be redrawn. No event happened to be raised during the cut, so the event-drop
 behaviour is **still unmeasured** after three rounds of internet cuts.
 
+**And it came back by itself, on schedule.** The very next scheduled report went through —
+**23:23:43Z, „Hub report pushed successfully (15354 bytes)”** — exactly 15 minutes after the
+cycle that failed, and nothing was done to the box to achieve it. The reading was taken at
+23:24:21Z, deliberately *after* the report was due, so it could not be premature. So the whole
+shape of a hub outage is now measured end to end: build → three attempts → give up →
+keep serving → next cycle succeeds → no alarm, no loss.
+
 **The near-miss, which is luck and not design.** Last good report 22:53:43Z; next scheduled
 23:23:42Z; `node_stale` trips at 30 minutes. The gap is **29 m 59 s**. The staleness threshold is
 exactly twice the report cadence, so **a single failed push spends the entire budget** — one second
diff --git a/documentation/audits/evidence-chaos-night-2026-09-17/round-1-notes.txt b/documentation/audits/evidence-chaos-night-2026-09-17/round-1-notes.txt
index 94d1bb9a..52328e4a 100644
--- a/documentation/audits/evidence-chaos-night-2026-09-17/round-1-notes.txt
+++ b/documentation/audits/evidence-chaos-night-2026-09-17/round-1-notes.txt
@@ -196,3 +196,22 @@ mine, with different cures, and it took separating them to fix either.
 
 Household loop through all of this: unit active, **4 sampled apps since the classifier fix, 0
 failures** — the box kept serving its other apps while five were being torn down and rebuilt.
+
+## A LATE GHOST FROM THIS ROUND, recorded 2026-09-17T01:26 CEST (23:26Z)
+Round 1's off-site poll loop reported "completed, exit code 0" TWO HOURS after it stopped working.
+  last real output line: 2026-09-16T21:24:56Z
+  task actually ended:   2026-09-16T23:24:56Z, with
+      "Read from remote host 192.168.0.115: Connection reset by peer"
+      "client_loop: send disconnect: Broken pipe"
+It went silent during round 2's power cut (the box it was polling was abruptly stopped), and the
+SSH session then hung, producing nothing, until TCP reset it two hours later.
+
+Three things worth keeping, all of them this project's recurring classes:
+  1. A HUNG connection and a FINISHED one look identical from outside: both produce no new output.
+     Silence is not completion. The loop's own timestamps are what distinguish them, which is why
+     every poll line carries one.
+  2. The task EXITED 0 while its connection had been reset. Another instance of "exit codes that
+     lie" - the exit status described the local shell, not the remote work.
+  3. A completion notice arriving two hours late could easily be read as a FRESH result for
+     whatever round is running now. It was checked against its own content before being believed.
+It did nothing to the box after 21:24:56Z, so no round is contaminated. Recorded, not hidden.
diff --git a/documentation/audits/evidence-chaos-night-2026-09-17/round-9.txt b/documentation/audits/evidence-chaos-night-2026-09-17/round-9.txt
index 380adaac..699a7c6b 100644
--- a/documentation/audits/evidence-chaos-night-2026-09-17/round-9.txt
+++ b/documentation/audits/evidence-chaos-night-2026-09-17/round-9.txt
@@ -72,3 +72,13 @@ injector is left as it is and this caveat travels with the three rounds that use
 Last good report 22:53:43Z. Next scheduled 23:23:42Z. node_stale trips at 30 minutes.
 The gap is 29 m 59 s. One second more and the operator would have been paged for a box that was
 perfectly healthy and had already repaired itself. That is worth a register row.
+
+### THE RECOVERY, confirmed by a reading that waited for its own precondition (23:24:21Z)
+  2026-09-16T23:23:43.762Z [INFO] [report] Hub report pushed successfully (15354 bytes)
+The very next scheduled report went through - exactly 15 minutes after the cycle that failed, and
+about 13 minutes after the link returned. Nothing was done to the box to achieve this.
+The reading was taken at 23:24:21Z, deliberately AFTER the 23:23:42Z due time, so it could not be
+premature. That is the sixth-mistake fix working as intended: the measurement states its own
+precondition instead of me judging the clock by eye.
+So the full shape of a hub outage on this product is now measured end to end:
+  build -> 3 attempts -> give up -> keep serving -> next cycle succeeds -> no alarm, no loss.