# RLV-1 — the digest's dark-routine count is measured from the most recent beacon, so a routine that misses a day and comes back reads "on schedule"; with a weekly email a routine can miss four days of five and never be named. Measured live: the panel has missed three consecutive Thursdays and no email has said so

status: open · raised daily audit 2026-10-29 · first identified daily audit
2026-10-23 (finding 2) and deferred to this queue, where it was never filed ·
class: **G2 loudness on the owner's only channel** · judgement: **YES** ·
effort ~0.25 pd · horizon: **live now — three misses already hidden, and the
next concealment is tomorrow's 2026-10-30 digest**

---

**Plain-language summary for an owner reading one paragraph.** Your Friday
email tells you whether each of the seven cloud routines is running. It works
out the answer by asking "how many scheduled slots have passed since this
routine last reported?" — so the instant a routine reports again, every earlier
miss disappears from the arithmetic. A routine that goes dark on Thursday and
runs on Friday is reported as healthy on Friday. The research panel has now
missed **three consecutive Thursdays** (10-15, 10-22, 10-29) and come back each
Friday, and no email has ever named any of them. Since the only email you read
is weekly, the hole is as wide as the week: four misses out of five weekdays
still read "7/7 reporting on schedule" as long as the fifth day works.

---

## The mechanism, CONFIRMED-BY-EXECUTION

`_gather_routine_liveness` (`src/thales/execution/fleet_digest.py:458`):

```python
fires = _cron_fire_dates(r["cron"], last + timedelta(days=1),
                         today - timedelta(days=1))
row["missed"] = len(fires)
row["since"] = fires[0].isoformat() if fires else None
```

`last` is the most recent beacon commit date. The window therefore *starts*
at the last success, which is precisely the information a recurrence needs to
survive. Run against the production functions for tomorrow's digest
(`cron='0 13 * * 1-5'`, `today=2026-10-30`):

```
last = 2026-10-30 (panel runs Friday)  -> fires = []                -> missed 0
last = 2026-10-28 (panel dark Friday)  -> fires = [2026-10-29]      -> missed 1
_routine_liveness_lines([... missed 0 ...]) -> "routines: 1/1 reporting on schedule"
```

So the defect is not in the rendering and not in `_cron_fire_dates`. It is the
choice of window start, and it is load-bearing exactly when the thing being
detected is intermittent.

Note what is NOT broken, so a builder does not re-fix it: `_slot_unverified`
(the `⏳ awaiting` line, 2026-10-15) correctly reports a slot that has passed
unproven on the day. It raises nothing by design, and on a non-Friday it
renders only into the box's local log, which no reader off that machine can
see. It is not a substitute for this.

## Evidence that this is live and not theoretical

Every weekday from 2026-08-01 to 2026-10-29, beacon committer dates in UTC:

- panel-specific dark days (no other routine affected): **2026-10-15,
  2026-10-22, 2026-10-29** — three, all Thursdays, all followed by a Friday run.
- every other panel dark day belongs to a multi-routine block (08-25→08-27,
  09-07→09-10, 09-22→09-24, 10-12→10-14), each already recorded.
- the digest named none of the three, by the mechanism above.

Recorded with its measurement in
`research/2026-10-29_three_thursdays_dark_and_the_only_email_says_seven_of_seven.md`.

## Done-when

1. The digest's routine-liveness section reports, per routine, the number of
   scheduled fire dates in a **trailing window** (suggest 30 days, declared as
   a module constant beside `_ROUTINE_SLOT_GRACE_MIN`) that produced no beacon
   — not only those since the last beacon. A routine that missed 10-29 and ran
   10-30 is named on 10-30.
2. The existing `missed since <date>` line keeps its meaning (a routine dark
   *right now*), because "stopped" and "intermittent" are different alarms and
   collapsing them loses the one the owner acts on. Suggest: the current line
   for an ongoing outage, a second line counting recurrences for a routine that
   is currently up but missed slots in the window.
3. **No new state.** Recompute from beacon history (`git log --format=%cI`
   over the already-fetched ref), not from a written ledger: the digest is the
   owner's only email and a write path in it is a new way to lose the email.
   `_last_beacon_date` currently returns only the latest date; a sibling helper
   returning the set of dates in the window is the shape, with the same
   fail-soft contract (any failure → `None` → the line is omitted, never an
   exception).
4. Fail-soft proven by test: an unreadable branch, a host with no `git`, and a
   malformed cron each leave the digest rendering and the email sending.
5. Negative controls: (a) a routine with a beacon on every fire date in the
   window produces no recurrence line; (b) a routine dark for the whole window
   is named once, not twice; (c) the 2026-10-15 `awaiting` behaviour is
   unchanged.
6. A test reproducing the real case: beacon dates 10-21, 10-23, 10-26, 10-27,
   10-28, 10-30 against `0 13 * * 1-5` on 2026-10-30 reports **two**
   recurrences in the window (10-22, 10-29) and not zero.

## Scope note for the implementer

Reporting tier only. `fleet_digest.py` is not imported by any trading path, and
its tests run in `fleet-digest.yml`, not the pre-trade gate. The one real risk
is the Friday email itself, which is why done-when 3 and 4 are written as hard
requirements rather than preferences. **Do not merge a change to this path on a
Friday before 15:30Z.**

## What this row does NOT cover

Why the panel misses Thursdays. The cause is unknown, the control plane is not
reachable from a routine, and a routine may not edit a routine. That half is an
owner/steward matter and is stated in the note, not here.

---

## DISMISSED 2026-11-03 — queue triage, overflow under the pre-registered 12-cap

Second of four evictions owed (`open/` stood at 16 against a cap of 12).

**The mechanism reproduces. Its evidence no longer does — and that is why it
ranks below the other fifteen.**

- **Mechanism, verified at HEAD:** `fleet_digest.py:488-490` still computes
  `fires = _cron_fire_dates(r["cron"], last + timedelta(days=1), today - timedelta(days=1))`
  and sets `row["missed"] = len(fires)` at `:493`. The window still starts at
  the last beacon, so a routine that misses a fire date and then pings reads
  zero missed afterwards. `_ROUTINE_SLOT_GRACE_MIN = 120` at `:415`; no
  trailing-window constant exists. (The body's `:458` citation is now the
  `def` line; the code moved to `:488-495`.)
- **The live instance was cured at the root, not papered over.** The row was
  raised 2026-10-29 on "the panel has now missed three straight Thursdays".
  Cause established since: the weekly Claude quota (`rate_limit: rejected
  (seven_day)`) killed the panel on turn 1 every Thursday morning — answered
  as a HANDS item on 2026-10-31 from the routine's own run history, and fixed
  by the 2026-10-30 ×5 cadence cut (#310), which also moved the panel to
  Mondays. The concealment this row describes cannot be reproduced from the
  panel's beacon any more.
- **Its own horizon has gone stale in the confirming direction.** The body
  says "the next concealment is tomorrow's 2026-10-30 digest". That digest
  came and went five days ago; nothing was concealed, because nothing was
  missed.
- **The cut also shrank the exposure.** At ~4 routine runs a week instead of
  ~23, there are far fewer fire dates to miss, so each miss is both rarer and
  more likely to be visible as a gap rather than hidden inside a window.
- **Effort is higher than stated:** ~0.4 pd, not ~0.25 — six done-whens
  including three fail-soft tests and three negative controls.

**What is NOT claimed here:** that the digest can see intermittency. It
cannot, and the day a routine starts flapping this row is right again. It is
dismissed because it is a latent reporting defect with no live instance, on a
queue where four P0/P1 rows with dated fuses are competing for the same slot.

**Reopen conditions (any one):**

1. **An intermittent miss is observed.** Any routine whose beacon shows
   miss → ping → miss across two scheduled fire dates while the digest's
   routine-liveness section reports it on schedule. This is the row's own
   done-when 6 and it is mechanically checkable from `beacon/*` branch
   history plus the digest text of the same day.
2. **A digest reports 0 missed for a routine whose beacon skipped a fire
   date.** The direct falsification, no flapping required — one skipped date
   plus one later ping is enough.
3. **The cadence is raised again.** If any routine returns to a daily or
   near-daily cron, the exposure the ×5 cut removed comes back and this row
   reopens with it.
4. **RLV-1 is folded into DGX-1's work.** Both are `fleet_digest.py`
   liveness-reporting rows; if DGX-1 (also dismissed today) is ever revived,
   this is a ~0.1 pd addition to that change, not a separate row.
