PCS-1 — every written description of the research pipeline now contradicts the live crons, and in 8 of the next 12 months the approval gate runs AFTER the build slot it is supposed to precede
status: open · raised panel run 48 (2026-11-02) · class: G5 registration hygiene / records truth · judgement: YES, narrowly · effort ~0.1 pd for the prose + a test; the schedule half is NOT an agent fix · horizon: decidable immediately; the first instance is this week (2026-11-05)
First line, so nobody mis-files this: the cron change is the OWNER's, not an implementer's.
ops/STEWARD.md§0 limit 9 forbids the steward to create, edit, run or delete any routine, and the panel and implementer are propose-only. The deliverable on the queue road is the prose + the test that stops it rotting again; moving a cron is anops/HANDS.md/NEEDS-OWNERitem and is written up as a recommendation below, not as work.
Plain-language summary. The research pipeline is meant to run panel → triage ranks it → steward approves it → implementer builds it. On 2026-10-30 the owner cut every routine's cadence by 5×. Triage now runs on the 3rd of the month and the implementer on the 5th, but the approval step in the middle runs weekly, on Thursdays. In most months no Thursday falls between the 3rd and the 5th — so the thing triage ranks on the 3rd cannot be approved before that month's only build slot has already come and gone, and it waits until the 5th of the next month. This November is the first instance: the implementer fires 2026-11-05 at 17:02Z and the steward at 23:00Z, six hours later. Separately, five places in the repo still describe the old daily/weekly pipeline in words, and one of them — written in the same change that caused the problem — asserts that the ordering holds.
What is true (EXECUTED today; the schedule read from the live control plane, not from the repo)
1. The live crons. Read with list_triggers (claude.ai control plane — the
repo's table is a copy, so the copy is not evidence here):
| routine | cron (PT) | next fire |
|---|---|---|
| queue triage | 0 9 3 * * | 2026-11-03T17:02:58Z |
| research implementer | 0 9 5 * * | 2026-11-05T17:02:54Z |
| steward (the approval gate) | 0 15 * * 4 | 2026-11-05T23:00:53Z |
The gate fires 6 h 00 m after the build slot it is meant to precede.
2. The adjacency fails in 8 of the next 12 months. A steward slot lies strictly between the 3rd 09:00 PT and the 5th 09:00 PT only when a Thursday falls on the 3rd or 4th. Enumerated independently three times today (the change lens, the chair, the adversarial reviewer), same answer:
| month | 2026-11 | 2026-12 | 2027-01 | 2027-02 | 2027-03 | 2027-04 | 2027-05 | 2027-06 | 2027-07 | 2027-08 | 2027-09 | 2027-10 |
|---|---|---|---|---|---|---|---|---|---|---|---|---|
| gate precedes build? | no | yes | no | yes | yes | no | no | yes | no | no | no | no |
4 of 12 hold; 8 of 12 impose a one-month wait on anything triage ranks.
3. ops/ROUTINES.md:46-47 asserts the opposite, and was written in this
window. Verbatim: "Triage runs on the 3rd and the implementer on the 5th,
two days apart, so the panel → triage → steward-approval → build order
holds." The two routines named are two days apart; the step between them is
not in the gap. In a shop whose own doctrine is that markdown is executable,
a false ordering claim in the authoritative roster is the defect, not a typo.
4. Four more sites still print the pre-cut pipeline (verbatim, verified at
HEAD d77f0a9):
ops/QUEUE_TRIAGE.md:19—TRIAGE (Mondays)ops/RESEARCH_IMPLEMENTER.md:11-12—triage (Mondays)/STEWARD (weekdays)ops/ROUTINES.md:145— "Runs Monday, giving the week to triage before Wednesday's"ops/ROUTINES.md:153—STEWARD (weekdays)ops/STEWARD.md:269— "approved/never holds more than the implementer clears on Wednesday."
5. One live approved item already carries the stale date.
research/queue/approved/fos-3-…md:340 ends "Implementer slot: Wed
2026-11-04." Written 2026-10-28, falsified 2026-10-30. It will in fact be
built 11-05. The panel must never write approved/, so this one is listed
as a site for whoever moves that file to built/ — not as work to do here.
What this is NOT — three narrowings the adversarial pass imposed, recorded so they are not re-derived
- It does NOT block REF-1 before the December gate. The change lens filed
that as the live consequence; it is wrong.
src/thales/backtest/oos_monitor.pyis not onops/STEWARD.md§3's trading-path list (onlybacktest/engine.pyis), so REF-1 is an observability-tier fix the steward may author itself under §0 ("author observability-tier fixes and records as PRs") on 11-05 and merge on 11-12 — nineteen days before the gate read. REF-1 misses 2026-12-01 only if someone routes it to the implementer. - It is NOT a throughput collapse. The implementer clears all of
approved/in one run, one item at a time (ops/RESEARCH_IMPLEMENTER.md§1; the 2026-10-28 run built three — #298, #299, #300). The cost is a one-month latency on newly-approved items, not a queue that cannot drain. - The gate is being used.
approved/fos-3was approved 10-28 and is due 11-05, which is evidence the chain works, not that it stalled. A 5× slower pipeline is what the owner asked for; the extra month from mis-ordering is not.
Mechanism (why it would pay)
The pipeline's whole value is that a ranked item reaches a builder while the
ranking is still fresh. The failure is silent in both directions: the
implementer correctly reports "nothing approved", which is exactly what a
genuinely empty queue looks like, and the stale prose means the next fresh
session reasons from a diagram that has not been true since 2026-10-30. Five of
the fourteen items now in queue/open/ carry "live now" or pre-December
horizons, so the latency lands on real work.
Data plan
None — no market data, no PiT question, no trial. This is a records-and-schedule
defect. Sources: the live control plane (list_triggers), ops/*.md at HEAD,
and git show 1491885 (#310, the cadence cut).
Test design + negative control
Done-when (the deliverable): one pass over the six prose sites above, plus
a test that asserts every written pipeline diagram agrees with the Active
table's crons — the same shape test_docs_consistency already uses to pin the
beacon allowlist against ops/ROUTINES.md, so a future cadence change reds the
suite instead of rotting five files.
Negative control (the one that matters): re-introduce any one of the five stale strings and assert the new test reds on exactly that file; then revert and assert green. A test that passes against the stale text is decoration. Second control: the test must NOT read the live control plane (no network in CI) — it compares prose against the repo's own table, so the row's claim that the table matches the live crons stays a human check, stated as such.
Pre-registered kill criterion
This row dies when either:
- (a) the owner moves the steward or implementer cron so that a steward slot precedes the implementer in ≥ 11 of any 12 consecutive months, and the prose is corrected to match; or
- (b) the owner rules the adjacency irrelevant — e.g. that a one-month wait on newly-approved items is acceptable at this cadence — in which case the prose is corrected to describe a monthly-latency pipeline and the test pins that.
It does not die on "git history records the change", on this row being read, or on the November instance passing without visible harm. Both exits require the five stale strings to be gone, because the prose is the half an agent can actually deliver.
Recommendation to the owner (one line, for the Friday digest)
Move the implementer from the 5th to the 7th–9th of the month. Any of those puts at least one Thursday steward slot after triage's 3rd and before the build in 12 of 12 months, costs nothing, and needs no other change.
Effort · horizon · rank
Effort ~0.1 pd for the prose + test (the cron edit is minutes, but is the
owner's). Horizon immediate; first instance 2026-11-05. Rank: I put
this above fin-4 (which pre-declares itself the next eviction) and below
the two 60-session clock rows — it is cheap, it is true today, and it is the
only row in the tree that makes the pipeline's own description trustworthy.
queue/open/ goes to 15; the 12-item cap is triage's post-sweep move
(ops/QUEUE_TRIAGE.md, "Open-size cap"), not a gate on filing, so this is
filed on merit with its rank stated.
Provenance
Raised by the change-review lens of panel run 48 and heavily narrowed by the
adversarial pass, which killed the REF-1 alarm and folded in the prose
residue of a second candidate (CR-2, triage's counters) after finding two of
that candidate's four legs factually wrong: the counters are computed from
git log in weekly windows, not in runs, so a monthly cadence changes the
sampling rate and not the denominator; and dismissed/auto-approve.md states
at :32-37 that counter (a) reopens "as decide item #1 by pre-commitment,
rewrite or no rewrite", so its pre-committed action is executable, not dead.
Both corrections verified verbatim at HEAD by the chair.
ANNOTATION 2026-11-09 (panel run 49) — the schedule half has now been RULED (keep the crons), and this row's own first predicted instance occurred on schedule
Status unchanged, still open on its prose half. Two dated facts, recorded so neither is re-derived.
1. The first instance happened, exactly as pre-registered. This row
predicted "first instance is this week (2026-11-05)". From
ops/ROUTINES.md:62-63 (this panel cannot read the control plane; the schedules
are taken from the registry, and the 17:02Z observed fire time from panel run
48's marker, which did read it): the implementer fires 0 17 5 * * and the
steward 0 23 * * 4, so on 2026-11-05 the implementer ran around 17:00Z and
the approval gate six hours later at 23:00Z. Nothing triage ranked on 11-03
could be approved before the month's only build slot. The prediction is
confirmed, not merely still pending — and the steward's own 11-05 record
independently describes the same ordering as the thing it works around.
2. The schedule half is DECIDED, and against the row's recommendation. The
row says the cron fix is the owner's (STEWARD §0 limit 9) and recommends moving
the implementer to the 7th–9th. The steward ruled on it
(research/steward/2026-11-05.md:23, verbatim): "PCS-1's schedule half
(the implementer on the 5th often runs before a Thursday steward can approve
what triage ranked on the 3rd). Keep the crons. I work around the ordering
by approving the week before, and the cost is latency, not loss." So the
schedule half is answered — not by a cron change but by a declared workaround —
and what remains open is the prose half: the five sites still printing the
pre-cut pipeline, including approved/fos-3:340's dead "Implementer slot: Wed
2026-11-04".
3. The row's narrowing 1 was NOT taken, and is still available. This row
pre-registered that "REF-1 misses 2026-12-01 only if someone routed it to the
implementer" — because src/thales/backtest/oos_monitor.py is not on
ops/STEWARD.md §3's trading-path list, so REF-1 is observability-tier work
the steward may author itself. REF-1 was approved 2026-11-05 (#323) and routed
to the implementer's 2026-12-05 slot, with the four-day slip accepted in writing
in its approval block. The cheaper path is still open: authored 11-12 and merged
11-19, it would land twelve days before the 12-01 monthly verdict and nineteen
before the earliest gate decision. That is a one-line choice for a steward run,
not a new row, and it is recorded here rather than filed.