wiki: 2026-11-14 audit corrections — withdraw the 'sign not magnitude' / '12 h of lead time' / '18:03 run' claims; verified escalation mechanism (activity-aware magnitude gate) + 29-cycle zero-item consolidation streak
This commit is contained in:
@@ -2644,3 +2644,43 @@ Method: each "differing" copy was matched against the last **25 authoring revisi
|
||||
`agent-estate-wiki/wiki/incidents/2026-10-31-conduit-voice-local-mode-not-applying.md` is **238 lines** with the root cause evidenced (`/v1/pool`, controlled realtime sessions against the Mac socket, title "…**rejects `session.update` in every form**"). The authoring page has been **103 lines at every commit that ever touched it** (`0c826f2` 2026-11-01 00:30, `a0b3734` 04:48, HEAD), and `git log -S'/v1/pool'` finds **nothing** — that evidence never entered the authoring tree. Vault mtime **2026-11-01 02:56**, i.e. written **straight into the vault** ~2¼ h after the authoring commit and ~2 h before `a0b3734` ("fix the Quartz build (YAML title)") normalised the title. **Left untouched and not forward-synced** (§12b: investigate, never clobber). **This needs an owner decision:** either port the vault copy back into `/home/hermes/workspace/agent-estate-wiki` (keeping the Quartz-safe `title`) or declare the shorter authoring page canonical and record the vault copy as historical. Until then the two trees disagree about the Conduit local-voice-mode root cause.
|
||||
|
||||
|
||||
|
||||
## 2026-11-15 01:2x — corrections pass: the Mnemosyne escalation's mechanism (re-derived)
|
||||
|
||||
**Three claims published in the 2026-11-14 entries of this log are WITHDRAWN** — they were reasoned from a
|
||||
mis-attributed log line rather than read from it. Corrected against `~/.hermes/logs/mnemosyne-health-monitor.log`
|
||||
by line number (file = **1103 lines**):
|
||||
|
||||
- **WITHDRAWN: "the 12:00:52 AWST run saw the same hole and logged it ℹ️ informational (`capture delay -589.5 min`),
|
||||
so ~12 h of lead time produced no alarm."** The `-589.5 min` line is at **line 1046**, i.e. *inside the 00:03:35
|
||||
run's own block* (header line 1030; the ❌ is line 1047). The 12:00:52 run (header **line 957**) measured
|
||||
`capture delay 29.4 min` at **line 973** — healthy, exit 0, delivery `suppressed`. The monitor did not misread
|
||||
anything at 12:00.
|
||||
- **WITHDRAWN: "the job exited 0 at 12:00 and 18:03."** Schedule is `0 */12 * * *` — 2026-11-14's only slots are
|
||||
**00:09:12** and **12:01:52**. No 18:03 run exists.
|
||||
- **WITHDRAWN: the "threshold tests the sign, not the magnitude" explanation.** `mnemosyne_freshness_gate.py` fires
|
||||
on magnitude — `delay_min < -CAPTURE_LAG_MIN`, `MNEMOSYNE_CAPTURE_LAG_MIN` default **120** min. The earlier slots
|
||||
were silent because the gate is **activity-aware**: it can only flag a message with no capture row, and every user
|
||||
message up to the 09:03 AWST `ios` turn had one (final capture 09:32:59, +29.4 min later). The conversation was
|
||||
then idle until the **19:22 AWST desktop** turn — the first uncaptured message — which the next slot (00:03:35)
|
||||
flagged at `-589.5 min`. Detection latency after detectability: **4 h 41 m** — cadence, not a bug.
|
||||
|
||||
**What genuinely went unread — the corrected lesson.** `consolidation_log` shows the sleep cycles firing every
|
||||
~31 min and returning **0 items for 29 consecutive cycles** (streak opening `2026-11-14T02:04:42+00:00` =
|
||||
**10:04:42 AWST**; last productive cycles 4622/4621/4620 at `09:33:34` / `09:33:16` / `09:01:40`). Every monitor
|
||||
run printed `Last consolidation: … (0 items)` — including the "healthy" one — but the gate checks consolidation
|
||||
**age** (`> 36 h`) only, never the item count. **An empty-but-current cycle is the failure signature of a dead
|
||||
write path, and nothing watches it.** Candidate fix: an N-consecutive-zero-item-cycle exception in
|
||||
`mnemosyne_freshness_gate.py` (**not implemented — needs Anthony's approval**).
|
||||
|
||||
⚠ **Format trap, verified:** `consolidation_log.created_at` mixes formats — rows 4640–4651 carry `+00:00`, rows
|
||||
4620–4622 are **naive local (AWST)**. Compared as one axis they misplace events by 8 h.
|
||||
|
||||
⚠ **`failure_type` is routing metadata, not diagnosis:** this job's twelve earlier `cron_incidents` rows read
|
||||
`agent` ×3 / `timeout` ×1 / `unknown` ×8; estate-wide the table is 53 `unknown` / 47 `agent` / 15 `timeout` /
|
||||
14 `script` / 5 `auth`. The job is `no_agent`, so `agent` is impossible — read `auth` as unexplained, not as a
|
||||
credential fault.
|
||||
|
||||
Pages corrected: [[incidents/2026-11-14-mnemosyne-capture-stall]] (false section rewritten, `auth` bullet upgraded
|
||||
from suspicion to evidence), [[systems/scheduled-tasks]], [[systems/mnemosyne]], [[index.md]]. The incident itself
|
||||
remains **OPEN / `alerted` / unclosed**; the *cause* of the stall is still **not established**.
|
||||
|
||||
Reference in New Issue
Block a user