Files
felhom.eu/documentation/audits/clock-list-2026-11-11
admin 978666dde3
gates / gates (push) Successful in 2m2s
lists: the Clock list (70 jobs) and the End-dates list (39 things), 9 register rows, 4 decisions
Read-only review of every job on a timer and of everything that ends on a date,
with measurements from the three test boxes. Nothing ends in the next 14 days.
Register 236 -> 245 (R-1379..R-1387), pointer lines on twelve rows, D76-D79 in
STATUS.md. Part C: felhom-agent's clone address holds no token now; the other
clones wait for the operator's yes. R-1035's 4-week check moved to 2026-11-18
(operator read needed).

Claude-Session: https://claude.ai/code/session_0159rPz1ZhFKsS53msqPYxtS
2026-11-11 11:54:42 +01:00
..

The Clock list — 2026-11-11

The third list of the gap review. One block for each job that a clock starts, with eight questions each. Built from the source of four repositories and from read-only measurements on the three test boxes (host and guest). No document lists every timer — the claim in the brief is true (search S19, 0 hits). DooPlex, ep0 and the hub's database are NOT MEASURED, by ruling; their jobs are in the list from source and documents only. Files: TABLE.md (the answers), MATRIX.md (the marks), evidence/, tools/.

Five lines

  1. Jobs found: 77 helper blocks. 70 are in the table (22 in the guest, 21 on the host, 27 on the hub, DooPlex and ep0). 7 are tiny, 2 are end dates (all named in the table's head). No TEST-ONLY job was found.
  2. Jobs no architecture document names: 1 of 70 by the search (the sender of unsent OS-update reports). The other 69 are mentioned somewhere; no document puts them in one list or one order.
  3. Marks (560 answers): A 457 · D 0 · K 9 · N 94.
  4. The worst N: when the guest-power watchdog gives up, it writes one line in the host's journal and tells nobody (CA07). The household's box is then down, and the only outside sign is the hub's "cannot be reached" check. Next worst: a box with a wrong clock refuses every signed job, so it cannot take an agent update (CA09).
  5. The operator decides the two questions at the end of this page.

The clock itself, per box (MEASURED 2026-11-11)

Box Host clock against DooPlex Who keeps the time Zone: host · guest
demo-hp +0.07 s the host's time service, in step Europe/Budapest · UTC
Tester 1 +0.07 s the host's time service, in step Europe/Budapest · UTC
N100 32.6 s behind nobody: the time service is not running (R-1301) Europe/Budapest · UTC

The guest has no clock of its own; it reads the host's. The controller's night slots are in Budapest time. At a start with a wrong time the crash guard reads the clock before any time service runs (CA04). Who notices a wrong clock: the hub alarms only when a tunnel handshake is stamped in the future (CH15). A clock that is behind: nobody.

One night on a box (backup time 02:30; local time; durations from the night of 2026-11-11)

Start Job demo-hp · Tester 1 · N100 Limit Waits for
02:30 Dump of databases and volumes (CC01) 2 min 33 s · 40 s · 18 s 5 min for each dump; none for the leg nothing
03:30 Copy to the second drive (CC02) 39 s · 15 s · 11 s none the clock only (+60 min)
04:15 Off-site push and clean-up (CC03) 3 min 09 s · 59 s · 38 s none the clock (+105 min); skipped if a backup holds the flag
after it App updates (CC07) not seen this night backup time + 5 h the off-site push
04:30 Whole-guest gate opens; controller auto-update, if on (CC11) — — no backup, no app update in flight
05:10 Sweep of an abandoned off-site store (CC19) ran 1 min nothing
05:30 Off-site proof restore (CC09) 5 s · 5 s · 3 s — the off-site flag
06:00 Off-site integrity check (CC08) under 1 s — the off-site flag
02:30 to 05:37 Whole-guest backup, by the host (CA01) 11–13 min · 3 min · under 1 min the agent's cap the gate; then the OS-update leg (CA02) follows it

Every duration is from a small test store. Nobody knows these numbers for a full household disk. Where two jobs can meet: (1) on Tester 1 and the N100 the whole-guest backup started at about 05:30, the minute of the proof restore — measured, and nothing orders them (CC09-5). (2) On demo-hp one whole-guest backup started at 02:30:38, with the dump leg. (3) The auto-update's default time is the minute the gate opens (CC11-5). (4) A long second-drive copy makes the off-site push skip the night with no message (R-1186). (5) On the N100 the whole-guest backup starts 5 minutes later each day (05:18 to 05:37 in four days).

The N answers, by class (one register row each)

Class Row What Count Main members
W1 R-1383 A job does not run, gives up, or loses its state at a restart, and nobody is told 23 guest-power give-up (CA07) · refused signed job (CA09) · hub 05:00 check and 07:10 audit skipped when the hub is down (CH02, CH11, CH16) · operator mail retry lost at a restart, household mail never retried (CH04, CH03) · peer list push to ep0 (CH14) · app-update leg (CC07) · catalog sync (CC12) · PBS verify failure (CA13)
W2 R-1384 A verdict or an "already ran" check trusts a clock that can be wrong 30 the hub trusts the box's time for backup age, off-site age, proof age (CH02, CH07, CH08) · signed jobs (CA09) · stamps from the future block the OS-update leg, the trim, the line test (CA02, CA11, CA12) · crash guard at early boot (CA04) · hub clock jumps (CH03, CH04, CH09, CH16)
W3 R-1385 Two jobs can meet with no lock and no order 12 proof restore and whole-guest backup (CC09) · auto-update and the gate (CC11) · PBS verify beside a backup (CA13) · the timers on DooPlex and ep0 (CH24, CH25, CH26)
W4 R-1386 No time limit and no measured duration 25 the hub's 60-second sweep: one slow check delays all, the liveness check too (CH19, CH01, CH08) · the 07:10 audit's 10 minutes at fleet size (CH11) · the interval jobs on the box (CC12 to CC16, CC20, CA06, CA10)
W5 R-1387 A document and the source disagree on a time 4 the agent unit: Restart=always and a watchdog in the design, on-failure and no watchdog in the unit (CA18) · 13 h or 26 h for the hub-database alarm (CH23) · old "24 h / 02:30" text (CH17)

Known gaps (K, 9 cells): R-1260 the hour-change night on a box (3) · R-1189 a host clock ahead during a backup (2) · R-1186 the silent off-site skip · R-1293 the hub copy on the autumn hour-change day · R-1301 · R-401.

Not made up, or no limit, by ruling (each is an A)

Only the three backup legs are made up after a missed night; the proof restore, the integrity check, the app updates and the abandon sweep wait for the next night · a daily slot that is more than 60 minutes late is skipped, so nothing heavy runs at noon · no OS-update leg on a night with no good backup · the hub's snapshot and prune are made up, its other daily jobs are not · household mails about an outage: one each 7 days · days of 23 and 25 hours on the hub (D24).

Decisions for the operator (also in STATUS.md)

  1. Shall a job that gives up or is skipped tell you? Yes (CC's pick): one hub alarm for the four worst cases — the guest-power watchdog gave up, a hub daily check was skipped, an operator mail was lost, a signed job was refused. Small changes in the agent and the hub. No: the logs stay the only trace; the cost is that you learn of a dead guest from the household. If nothing happens: the same as "No".
  2. Shall the hub check each box's clock? Yes (CC's pick): the hub compares the time in each report with its own and alarms above one minute of error; it then stops trusting a backup time from the future. No: R-1301 (repair the N100's time service) stays the only guard; the cost is that a box with a wrong clock can look healthy and cannot take an agent update. If nothing happens: the N100 stays 32 s behind and nobody is told when it gets worse.

Method verdict (third list)

  • It worked for the map, and the measurement found the meetings. Three read-only helpers gave 77 blocks. The 05:30 meeting of two jobs and the 5-minutes-a-day drift on the N100 are in no helper answer; they came from the boxes.
  • Read by eye: 138 of 560 cells hold the main session's words (every K and N, and 35 A). For 70 of them the pointer was opened by eye; for 68 the helper's pointer stands and a tool opened it. The main session read every helper answer to questions 4 to 7 (280 cells); it did not open each of their pointers. A tool opened all 896.
  • The weak points: the hub's jobs and the DooPlex and ep0 timers are from source only. The controller's log held one night. The "named" column is a machine count with wide search words; it says "mentioned", not "explained".
  • The ninth line: 6 jobs have a hand step in a runbook; 4 have a hand step that no runbook holds (CC03, CA07, CA09, CH02); 1 in part (CA18). The other 59 have no hand step.