diff --git a/documentation/audits/evidence-chaos-night-2026-09-17/household-summary.txt b/documentation/audits/evidence-chaos-night-2026-09-17/household-summary.txt
new file mode 100644
index 00000000..ac198617
--- /dev/null
+++ b/documentation/audits/evidence-chaos-night-2026-09-17/household-summary.txt
@@ -0,0 +1,45 @@
+# The background household — summary for the whole night
+# A loop ran from the drill host every 2 minutes from 21:06:30Z to 00:18:10Z, reading each app's
+# front door and its dashboard, exactly as a household would use the box while all this happened.
+
+## The numbers
+  total probe lines        192
+  first line               2026-09-16T21:06:30Z
+  last line                2026-09-17T00:18:10Z
+  lines flagged as failure 7
+  REAL failures            4 lines = 2 events (see below)
+  apps probed              cloud inventory media paperless paste photos recipes share status travel
+                           vault wiki   (8 to 22 probes each, uneven because the loop cycles names)
+
+## THREE OF THE SEVEN ARE MY OWN BUG, and the log says so in its own words
+  2026-09-16T21:06:30Z recipes    dash  FAILED http=301
+  2026-09-16T21:08:31Z wiki       dash  FAILED http=301
+  2026-09-16T21:10:31Z inventory  dash  FAILED http=301
+At 21:11:25Z the log carries this line, written when I found it:
+  "### CLASSIFIER FIXED: every line above this counted a 301 redirect as a dashboard FAILURE
+   - my error, not the box's"
+A 301 from traefik is the redirect to HTTPS: the correct answer, not a failure. The classifier was
+fixed and the wrong lines were left in place rather than deleted, so the correction is visible.
+
+## THE TWO REAL EVENTS, both caused by an injected accident, both self-healed
+  2026-09-16T21:27:57Z wiki  read  UNREACHABLE   } round 2: the POWER CUT, 20 s into a restore
+  2026-09-16T21:27:57Z wiki  dash  UNREACHABLE   }
+  2026-09-16T23:28:09Z cloud read  UNREACHABLE   } round 10: the HARD RESET, 4 s into a restore
+  2026-09-16T23:28:09Z cloud dash  UNREACHABLE   }
+Both are single-sample outages: the very next probe two minutes later succeeded in each case, and
+no other app was unreachable at those moments. The household saw the box go away twice, for less
+than one probe interval each time, on the two occasions the machine was abruptly stopped.
+
+## WHAT THE HOUSEHOLD DID NOT SEE
+Ten of the twelve rounds left NO mark in this log at all - including the twenty minutes with the
+data drive pulled out (round 11), the three ten-minute network cuts (rounds 7, 8, 9), the full disk
+(round 3) and the docker restart (round 5). The apps whose files were on the pulled drive did stop,
+but the loop's probes fall on a 2-minute cycle per name and did not coincide with their outage.
+That is a limit of the instrument's sampling rate, stated so the silence is not read as proof.
+
+## THE NINTH INSTRUMENT SLIP, recorded like the rest
+My first attempt to list these failures printed NOTHING and said "grep: /root/household.log: binary
+file matches". The count (7) had already printed, so the summary looked complete while the detail
+was silently dropped. Re-read with `grep -a`. The cause is NOT corruption: zero null bytes; one line
+carries a long run of padding spaces. An instrument that can drop its result without saying so is
+not a measurement - the same lesson as every other slip tonight.