DISMISSED — DSP-1 — the Mac dispatch layer half-failed on three consecutive sessions and no machine noticed; considered as a loudness row and REJECTED
status: dismissed at adversarial review, panel run 25 (2026-09-17) · class considered: loudness & safety (G2) · the mechanism claims are all TRUE; the row is killed on severity, duplication and an unverifiable premise, not on facts
What was considered
On 2026-09-02 two launchd agents were added to the owner's Mac to fire the
time-sensitive workflows at their intended market times via gh workflow run,
because GitHub's own schedule: events had been arriving 3–9 h late and three
of five sessions traded after the close (RUNBOOK.md:21-38). The proposal was
that this layer can degrade silently, and has been degrading:
| session | trading dispatch (14:35Z) | skew dispatch (19:00Z) |
|---|---|---|
| 2026-09-14 Mon | none — crons ran ~5 h late (meanrev 19:36Z, vrp 19:37Z, momentum 19:48Z) | none (single skew commit 22:29Z) |
| 2026-09-15 Tue | on time (14:36–14:49Z) | none (single skew commit 22:10Z) |
| 2026-09-16 Wed | 2 of 3 — meanrev 14:40Z, momentum 14:50Z, vrp missing; a vrp dispatch appears 17:24Z | on time (19:23Z + 22:01Z backup) |
Verified against the Actions API as well as the commit timeline:
paper-trading-vrp.yml has a workflow_dispatch at 14:35Z on 09-09, 09-10,
09-11 and 09-15 and none on 09-14 or at 14:3xZ on 09-16;
skew-snapshot.yml has workflow_dispatch runs on 09-03/04/07/08/09/10/11/16
and none on 09-14 or 09-15.
The mechanism claims — all verified TRUE, recorded so nobody re-derives them
ops/launchd/com.thales.dispatch-trading.plist:10-17— one agent dispatches all three trading workflows in order.scripts/dispatch_workflows.sh:73-87—rccomes fromgh workflow run's exit code only, andbeacon successfires iff every call exited 0. The beacon attests SEND, never RUN CREATION.scripts/dispatch_workflows.sh:44-53— the beacon URL is optional and read fromconfig/.env; absent = silently disabled ([ -n … ] || return 0).RUNBOOK.md:81-96— the beacon inventory table has 14 rows and noHEALTHCHECK_URL_DISPATCHrow, directly belowRUNBOOK.md:76-79's own warning that "an unconfigured beacon reads as coverage while providing none." The secret appears exactly once in the repo outside the script (RUNBOOK.md:37) and is absent fromops/launchd/README.md's beacon setup.scripts/detect_runner_loss.py:106,257— the only mechanical run-level sweep filtersevent != "schedule", so the dispatch leg is outside it by construction.src/thales/execution/fleet_digest.py:1157-1189—digest_exceptionshas no timeliness leg, and the DID-NOT-RUN detector fires only when a sleeve has no run record for the day, which the backup cron prevents.
Why it was killed anyway
- Zero realized harm, and the backstop worked every single day. Every session in the window traded in-session: the latest was momentum 19:48Z on 09-14 = 15:48 ET, and vrp's 09-16 make-up ran 17:24Z = 13:24 ET. The layer this row defends is a convenience over a working fallback, not a control. Contrast the regime it was built for (08-27→09-01), where three of five sessions traded after the close. Arguing urgency from a degraded convenience whose backstop held is severity inflation.
- It is BCN-1 twenty-four hours later.
dismissed/bcn-1-liveness-beacon-gated-on-job-success.mdwas dismissed by panel run 24 on 2026-09-16 on materially this argument — a beacon-coverage gap, backstopped by a channel that actually fired, against a 12/12 cap. None of BCN-1's three reopen tripwires is satisfied here. Opening a near-twin the next day is exactly the paddingops/RESEARCH_PANEL.mdnames. - The load-bearing premise is unverifiable here, and the live evidence cuts
against it.
config/.envis absent from the routine container, so whetherHEALTHCHECK_URL_DISPATCHis configured is NOT-VERIFIABLE. The 09-16 17:22:59Z re-dispatch was vrp-only, not the agent's three-workflow batch — i.e. a targeted human action, which is positive evidence that something did reach the owner. A row headlined "nothing notices" cannot rest on a fact the container cannot check while the contrary reading sits in the same evidence. - The expensive half is the weakest half. A repo-side
run_started_at-vs-intended-time detector is ~0.5–1 pd ofgh apiplumbing.dismissed/cron-delay-capture.mdis not a C0 bar on it (a detector is not a research capture stream, and neither of that dismissal's two reopen tripwires has fired), but its third argument stands: the legitimate ops value is already captured free — every audit memo timestamps these delays, and the daily audit found this one by eyeball the same evening.
Kept out of the queue, handed over instead
- One line to the owner, in this run's report: confirm whether
HEALTHCHECK_URL_DISPATCHis configured — the skew agent produced no run on both 09-14 and 09-15, and the trading agent missed vrp on 09-16. If that beacon is not set, three silent half-failures in three sessions is what its absence looks like. - A docs one-liner in passing: add the
HEALTHCHECK_URL_DISPATCHrow toRUNBOOK.md's beacon inventory table with its check spec, so the beacon stops being the one that is on no list to verify.
Neither is a research item and neither needs a slot.
Reopens if
- (a) a session ever trades after the 16:00 ET close with the dispatch layer having failed silently — the harm this row could not demonstrate; or
- (b) the owner confirms
HEALTHCHECK_URL_DISPATCHis not configured, which converts claim 3 from unverifiable to established and leaves the dispatch layer with no liveness signal at all; or - (c) a dispatch is shown to have exited 0 with no run created (the
Mac's
logs/dispatch.logsaysdispatched <wf>for a workflow with no corresponding Actions run) — that is the "beacon that cannot fail" shape, which no amount of backstop argument excuses; or - (d) BCN-1's own reopen conditions fire, since the two rows would then merge into one structural question about the liveness layer.
The data-side consequence of these same three sessions is NOT dismissed: it is
open/cqa-3-capture-qa-blind-to-capture-instant.md, which stands on its own
evidence and its own kill criterion.
ANNOTATION 2026-10-02 — panel run 33: RULING ON LEG (d), which fired on 2026-09-25 and sat unruled for seven sessions. (d) is now the carrier for the whole liveness-layer question, including the Windows runner.
Leg (d) has FIRED. dismissed/bcn-1-liveness-beacon-gated-on-job-success.md's
annotation of 2026-09-25 records BCN-1's reopen conditions (1) and (2) as
fired. Leg (d) here reads: "BCN-1's own reopen conditions fire, since the two
rows would then merge into one structural question about the liveness layer."
This file carried no annotation until now. That is a lapse of this panel's own
expiry watch — ops/RESEARCH_PANEL.md §3 item 5 makes the panel the reader, and
RESEARCH_PANEL.md:20-23 gave it the instruction — so the failure was a reader
that existed and was not run, which is both worse and more fixable than a missing
instruction. Recorded as the dated counter-evidence to RDR-1's framing (see that
file's 2026-10-02 annotation).
Ruling: (d) is DISCHARGED by merging, not by reopening. The structural question BCN-1 and DSP-1 both circle is a liveness-layer design question only the owner can answer, and the queue stands at 14 against a cap of 12. Neither row reopens. What follows is the merged question, restated once, with the mechanism the 2026-10-02 Windows-runner batch adds and a kill criterion sharp enough to settle it in one session.
The merged question, as of the tbx-win cutover (scheduled Tue 2026-10-06 18:00 PT)
After the cutover retires the Mac's two dispatch agents, a silently dead Windows box reads green on every channel:
- No session means no
alert()at all —scripts/win/session.py:491-493only fires when a session ran. - The GitHub backup cron trades and pings the same healthchecks check
(
paper-trading.yml:374if: success()), and its early-exit path (:118-144) is a successful run, so it pings even when it did nothing.ops/windows-host/README.md:233-235instructs the owner to copy the same four ping URLs onto the box precisely because they "exist only as GitHub secrets, which can't be read back" — so the one-check-per-sleeve aliasing is deliberate, not an accident to be confirmed. src/thales/execution/fleet_digest.py:147'sno_run = (not _ran) and is_trading_day(today)has no leg field.- The dead-man heartbeat reads equity advancement, which advances whichever leg ran.
- PR #174's dispatch-punctuality detector — the one alarm built for "it ran, but
at the wrong time" — reads
actions/workflows/{wf}/runsonly. The box creates no Actions run, so a box that never ran reads PUNCTUAL: the backup's 15:30Z start is 55 min after the 07:35 PT intent in PDT and 5 min before it in PST, both insideLATE_AFTER_MINUTES = 90. Its docstring's claim that it "stays true if the dispatcher is retired, rewritten, or replaced" holds for a dispatcher swap but not for a replacement that moves execution off Actions.
The attribution data already exists and has zero readers.
scripts/win/session.py:52 defines RUNNER_TRAILER = "Runner: tbx-win" and
:229 writes it as a second -m on every state commit; the GitHub leg commits
with one -m and no trailer (paper-trading.yml:354), and
ops/windows-host/README.md:111 says the trailer exists "so audits can tell
where a commit was made". Trailer-presence is therefore a clean binary partition
of the two executors. A repo-wide grep for Runner: tbx-win / RUNNER_TRAILER
finds only those three sites. So this is a reader-side one-liner, not a design
change — which is why the earlier framing "nothing can tell the primary from the
backup" was an overstatement and is not adopted here.
Precedent that the harm class is real and already measured once: PR #199 shows a hole in exactly this record shortened a pre-registered clock (75 counted vs 77 elapsed) and put two two-day moves into the December gate's bootstrap labelled as one-day returns.
Kill criterion — one session, zero cost, no wriggle room
Within the first 10 live sessions after the cutover, disable \thales\trading for
exactly one session under NORTHSTAR.md §3's drill protocol (sealed envelope
first, dead-man revert, watching layer only — never the trading path). If the
fleet digest, all four healthchecks and #174's detector are all green that day,
the hole is demonstrated and a reader lands (grep the trailer on the day's
state: commits; absent trailer on a day the box was due = primary silent).
If any one of the three goes red, this question is dead and both rows stay
closed for good.
One sequencing item for the owner, not a research question
PR #174's two rot pins (tests/test_execution/test_dispatch_gap_detector.py:432
and :439) read ops/launchd/com.thales.dispatch-trading.plist, which PR #206
renames into ops/launchd/retired/. The two PRs touch disjoint files, so
there is no textual merge conflict and nothing warns; whichever lands second
either reds the pins or silently removes them. Both tests are
observability-marked, so the pre-trade gate is not fail-closed by this —
the fleet-digest workflow is what goes red. Merge order is the owner's to
sequence before Tuesday.
ANNOTATION 2026-10-05 — panel run 34: this file's OWN 10-02 reading of PR #174 is FALSE and is corrected here, and the kill criterion it set can no longer be read — two of its three instruments stopped existing about sixteen hours after it was written.
No row. research/queue/open/ stands at 14 against the pre-registered cap of
12 (ops/QUEUE_TRIAGE.md:86-91), and every remedy below is owner-side: a
policy call on an external dashboard, a merge-order decision on an
owner-authored PR, and a countersigned restatement of a pre-registered
criterion. Recorded as an annotation, per the handling dismissed/bcn-1-…'s
2026-09-25 annotation established: the tripwire fired, the dismissal's stated
justification no longer holds, and the remedy it names is still not worth a
slot.
(A) The 10-02 annotation's claim about PR #174 is wrong, and the error inverts
That annotation reasoned: "the backup's 15:30Z start is 55 min after the 07:35
PT intent in PDT and 5 min before it in PST, both inside
LATE_AFTER_MINUTES = 90" — and concluded that after the cutover a dead box
would read PUNCTUAL to #174's detector.
That computed the configured cron, not the delivered run. Measured from
the Actions API this run, paper-trading.yml's schedule runs on the two most
recent sessions were created at 19:44:27Z (10-01) and 20:03:31Z
(10-02) — 309 and 328 minutes after the 14:35Z intent, against a 90-minute
bar. Both concluded success, not skipped: src/thales/execution/daily.py:1526
returns skipped=True, skip_reason="Already ran today (…)" and
src/thales/cli.py:1102-1112 returns exit 0 for every skip reason except
{"No signals", "Stale price panel"}, so took_a_decision is True on it.
After the cutover the dispatch leg disappears and the backup cron is the only
Actions run per trading workflow per day. So the correct reading is the inverse
of what this file recorded: not "a dead box reads PUNCTUAL" but "a live,
punctual fleet reads LATE — every trading day, for all three workflows",
through heartbeat-monitor.yml into the operator's one alarm channel. The
conclusion holds in PST too (249 minutes). A detector that pages on every
correct day trains the operator to silence the channel, which is the one thing
leg (C) below makes expensive.
The absence half of #174's docstring claim (research/2026-09-17_dispatch_punctuality_detector.md)
is still correct; the backup-cron-as-decisive-run case is what defeats it.
Owner's call: fix the intent constant to the backup's real schedule, or hold
the PR. The secondary rot (its two pins read
ops/launchd/com.thales.dispatch-trading.plist, which PR #206 renames into
retired/) is unchanged from the 10-02 sequencing note and cannot fail-close
trading — both tests are observability-marked and deselected by both runners.
(B) The kill criterion set on 10-02 is unreadable
It read: "If the fleet digest, all four healthchecks and #174's detector are all green that day, the hole is demonstrated and a reader lands."
- The four checks (
thales-momentum,-meanrev,-vrp,-skew) were deleted on healthchecks.io 2026-10-02, recorded inf865a27(#212) at 23:20 PT — this file's criterion was committed in7c47c44at 06:54 PT the same day. A deleted check can never go red, so the criterion's "if any one goes red, this question is dead" escape loses one of its three legs entirely. - "#174's detector" is an unmerged PR which, per (A), would read LATE on presence and green on absence.
Only the fleet digest's no_run (src/thales/execution/fleet_digest.py:140-147)
survives as a readable instrument. Until the criterion is restated, the drill
must not be run — a drill whose instruments are gone cannot return a finding
in either direction.
Two things keep this off the C0 ledger and they belong on the record: the owner's deletion was an independent, legitimate, recorded ops decision ("this isn't a production app", 2026-10-02), not a threshold moved after its data existed; and the vacating biases the drill toward "the hole is demonstrated" — toward work, not complacency. Nobody gained anything. It is a G5 registration-hygiene defect, not a trust breach.
The restatement is the owner's, not an agent's. The one Actions-independent
instrument left is scripts/win/session.py:52's
RUNNER_TRAILER = "Runner: tbx-win", written as a second -m on every state
commit (:229) and still with zero readers. Naming it here is as far as a
routine may go; proposing new criterion wording would be exactly the post-hoc
edit NORTHSTAR.md §3 forbids.
(C) The healthchecks deletion is recorded in ONE document and four readers now assert the opposite
git diff 7c47c44..HEAD -- RUNBOOK.md AUDIT.md ops/DAILY_AUDIT.md ops/ABSENCE.md
is empty. The decision lives solely at ops/windows-host/README.md:236-242,
a Windows-host build doc. Meanwhile:
RUNBOOK.md:83-88lists all four deleted checks as live rows — includingHEALTHCHECK_URL_SKEW"high — a missed capture day is unrecoverable at any price" — in a table whose own guard at:78-81reads "Do not consider this section done until every check in the table exists" and "An unconfigured beacon reads as coverage while providing none — worse than having no beacon." The table indicts itself.RUNBOOK.md:208-210: "The irreducible residual — a correlated outage where GitHub gives us no runners at all — is covered only by the healthchecks.io beacons above." For the three sleeves and the capture that coverage is now zero.ops/DAILY_AUDIT.md:347-348instructs the audit routine "Verify the beacons fired; do not assume they did" — unexecutable for that tier.ops/ABSENCE.md:38: "beacons are the only alarm that reaches a phone on a beach."tests/test_execution/test_ops_alerting.py:530-544(test_every_alerting_workflow_has_a_beacon) passes green — verified,pytest -k beacon→ 6 passed — while certifying four of its six beacons watch nothing. It asserts the presence of a curl step, which survives; the thing it proxies for is gone. Severity ceiling, stated honestly: the whole file isobservability-marked, so it never gated trading and cannot fail-close anything. It is a test whose docstring is now a lie about the one asset it names.
Nothing will catch this mechanically: tests/test_docs_consistency.py:574 pins
only the routine beacon names against the relay allowlist; the 14-row
trading/capture inventory has no guard — the same shape as this file's own
claim 4 about the missing HEALTHCHECK_URL_DISPATCH row.
One owner decision closes all of (C): delete the four rows and restate the correlated-outage residual honestly, or re-add the four checks against the box. ~0.25 pd of doc edits behind a one-line choice. An agent picking either would be a silent scope change on a deliberate owner decision.
Reopens
The original legs (a)–(d) stand as written; (d) remains discharged by merging. Added by this annotation:
- (e) if the four checks are re-added, or the RUNBOOK rows deleted — either resolves (C) and the criterion in (B) becomes restatable;
- (f) if PR #174 merges without its intent constant re-cut against the tbx-win topology — the false-daily-LATE alarm in (A) then exists in production, on the channel (C) establishes is the last one standing;
- (g) if the drill in (B) is run before the criterion is restated.
Panel run 35 append — 2026-10-06: reopen condition (b) has NOT fired, and a merged record says otherwise
Why this append exists. research/2026-10-05_the_in_session_skew_dispatch_failed_silently.md:45-50
(merged as #219) states, as §2 leg 2 of a "four independent layers, all silent"
finding: "The Mac's dispatch beacon is not configured. HEALTHCHECK_URL_DISPATCH
is read from config/.env and absent means silently disabled … Established by
DSP-1 (panel run 25) and unchanged." The merge commit is blunter still:
"HEALTHCHECK_URL_DISPATCH is unconfigured … (DSP-1)".
That is a misattribution, and it inverts this row. Dismissal reason 3
(:66-73) establishes the opposite: "The load-bearing premise is unverifiable
here … config/.env is absent from the routine container, so whether
HEALTHCHECK_URL_DISPATCH is configured is NOT-VERIFIABLE." This row
established that the premise cannot be established from a routine container.
The record cites it as having established the premise.
Evaluated, since evaluating fired tripwires is the panel's job
(ops/RESEARCH_PANEL.md:20-23): (b) has NOT fired. config/.env is absent
from this container too, so the premise is still unchecked by the same
constraint. The nearest thing to an owner confirmation is
ops/windows-host/README.md:234-245 — healthchecks "not used", the four
trading/capture checks deleted 2026-10-02, "the other ten checks … stay", and a
"leave on the Mac" list naming HEALTHCHECK_URL_REVALIDATION and
HEALTHCHECK_URL_BACKUP but conspicuously not _DISPATCH. That is
suggestive by omission; it is a statement about four named checks and ten
survivors, not about the dispatch beacon. Inference from absence is not the
owner confirmation (b) requires.
What this row's other mechanism claims do at HEAD, re-verified:
scripts/dispatch_workflows.sh:44-53 reads the URL from config/.env and
beacon() returns 0 silently when unset ✓. RUNBOOK.md:81-95 has 14 beacon
rows and no HEALTHCHECK_URL_DISPATCH row ✓. So the mechanism is unchanged
and correct; only the state is unknown.
Why it was worth saying rather than shrugging. The forward misfire is concrete: a future panel doing the dismissed-tripwire sweep reads "established … and unchanged" in a merged record, concludes (b) is satisfied, and reopens this row into a 12-capped queue over a premise nobody in a routine container can check. That is precisely what the reopen-condition machinery exists to prevent.
Neither reading rescues the 10-05 record's conclusion, and it does not need
rescuing. The capture was missed at its intended instant and nothing paged —
that is empirical and is confirmed independently (the 10-05 session landed at
20:43 ET; see the panel run 35 append on open/cqa-3-…md). What is affected is
only the strength of the claim that all four layers are silent: legs 1, 3 and 4
are mechanically verified, leg 2 rests on an unverified state stated as fact.
And the question is about to retire itself. PR #206's cutover moves
ops/launchd/com.thales.dispatch-*.plist into ops/launchd/retired/, so the
Mac dispatch layer this beacon would watch stops existing. The cheapest
disposition is one line of owner confirmation before the cutover makes it
unanswerable — after which (b) is not merely unfired but unfireable, and this
row's claim 3 should be struck as moot rather than left reading "unverifiable".
Recorded by panel run 35. (b) NOT fired; this row stays dismissed; no row minted; no queue slot consumed. The daily audit corrects its own record — this is the panel telling it.
ANNOTATION 2026-10-16 (panel run 39) — reopen leg (a) FIRED on 2026-10-13, three days unevaluated; the harm this row could not demonstrate is now on the record, and the drill it proposed need not be run
Plain-language summary. This item was killed because it could not show the harm it worried about: a trading session going ahead after the market closed while the layer that dispatches it had failed quietly. It left a tripwire for exactly that event. On 2026-10-13 the event happened — the mean-reversion sleeve traded an hour after the close, from GitHub's backup leg, while the Windows box that was supposed to run it said nothing. Verified here from commit metadata, independently of the steward's report.
The clause, verbatim:
(a) a session ever trades after the 16:00 ET close with the dispatch layer having failed silently — the harm this row could not demonstrate
The evidence, from commit trailers (not from any routine's prose):
| commit | subject | committed (UTC) | Runner: trailer |
|---|---|---|---|
6dbbc2d | state: meanrev paper run 2026-10-13 | 21:00:17Z = 17:00 ET | (empty) |
f271de3 | state: paper trading run 2026-10-13 | 21:01Z | (empty) |
56d6e5e | state: paper trading run 2026-10-13 | — | Runner: tbx-win |
dc2d319 | state: vrp paper run 2026-10-13 | — | Runner: tbx-win |
1e013f6 | skew: snapshot 2026-10-13 | — | Runner: tbx-win |
An empty Runner: trailer is GitHub's backup leg. So on 10-13 the box ran vrp,
momentum and the skew capture, skipped meanrev and one momentum slot, and
the backup crons picked both up — meanrev at exactly one hour past the close.
"Silently" holds in the precise sense this row meant it. The digest's
no-run detector is no_run = (not _ran) and is_trading_day(today)
(fleet_digest.py:147), which reads False because the sleeve did run —
just from the wrong place, an hour late. This file's own text records that the
Runner: trailer has zero readers. And _alert.yml records-rather-than-
mails on a non-Friday under notifications.email: weekly_digest_only, with the
steward as the designated reader — and the steward was dark on 10-13 and 10-14.
It was found by hand on 10-15, two days late.
The corroborating damage, recorded the same day at CAPTURES.md:422-437:
the same post-close backup leg wrote pit_snapshots 2026-10-12 = 49 rows
and 2026-10-13 = 1 row against an 886-row norm, "because the vendor had
published only part of the day's bars". Both are permanent (write-once made the
box skip each slot the next morning). So leg (a)'s harm is not merely a late
fill — on these two dates the same mechanism cost two irreplaceable capture
sessions (NORTHSTAR.md §1 asset 2, gauge G4).
The bonus, and it is a real saving. This file's section (B) proposed an
injection drill to manufacture exactly this evidence. The natural experiment
has now delivered it at zero cost, with no sealed envelope, no dead-man revert
and no drill-safety protocol to administer (NORTHSTAR.md §3 drill rules).
The drill is unnecessary — the finding it was designed to produce is above.
Disposition — no new row, and the reopen is not ours to execute. Per this
file's own 2026-10-02 precedent, a fired leg gets a dated append here, which is
what this is. dismissed → open is the steward's git mv, never a panel's
(ops/QUEUE_TRIAGE.md:100-102). Two concrete follow-ons for whoever opens it:
(1) the Runner: trailer needs a reader — a session that trades from the
backup leg while the box is silent is the signal, and it is currently written
down and read by nobody; (2) steward ask 3 of 2026-10-15 already asks the owner
to run thales --sleeve meanrev portfolio reconcile and read the box's 10-13
log, so the account of why the box skipped is pending and should land in the
same place.
Reopen condition, restated channel-neutrally (the original leg (a) is
intact and satisfied; this restatement is for the legs that are not, because
weekly_digest_only made "an alert arrives" unfirable on any non-Friday):
reopen also if a session trades from the backup leg on any day where no
human or agent reader is notified by any channel within one trading day —
delivery mechanism irrelevant.