ERN-1 — the earnings-date-as-EXPECTED panel: scout run 1's one PROVISIONAL kill, migrated out of the frozen archive so its tripwire is visible to ls
status: dismissed (PROVISIONAL, inherited) — originally killed by PANEL scout run 1 on 2026-08-17; migrated into queue/dismissed/ by PANEL run 24 (2026-09-16) · class: records hygiene / forward-capture candidate · never occupied a row in queue/open/ · costs no open-queue slot and no trial-registry row · does NOT discharge the scout's 8-week cadence clock (fires ~2026-10-12)
Plain-language summary for an owner reading one paragraph. When the scout
half of the frontier lens ran for the first time, it considered twelve possible
new data captures and killed eleven of them outright. The twelfth — recording
each company's expected earnings announcement date, as of the day we see it —
was killed only provisionally, because the kill rested on "no free endpoint
we can reach with credentials we hold", and the census had missed that this
repository already holds an unused API key that plausibly serves exactly that
endpoint. So the kill was made conditional on one question to you, answerable in
about a minute. That question has never been asked, and 30 days have passed.
Worse, the conditional kill was written only into research/RESEARCH_QUEUE.md
— the frozen pre-migration archive — so after the queue became a directory, no
future panel could find it by listing the tree. This file fixes that. It changes
no verdict and asks for no work; it moves one adjudication to where the memory
system can see it, and gives it a dated default so it cannot rot further.
What the idea is, and why it was not killed outright
Record, every day, the expected (not realized) next-earnings date for each name in the universe — the Johnson–So date-revision class. The un-backfillable part is the revision: a vendor's calendar shows today's best guess, and when the guess moves, the old guess is overwritten and gone. History can be bought; history of expectations cannot.
Honest mechanism sketch, as originally recorded: earnings-date revisions are watched closely by a small set of participants and ignored by most systematic flow, so a date that quietly slips is information about a company's own reporting process arriving before the event it describes. The counterparty is calendar-inattentive flow. That is a thesis, not proof — which is the right standard for buying an option rather than running a trial.
Why the kill is provisional — the unasked question
Verified again this run at HEAD 7fd0d74: FMP_API_KEY is wired into three
workflows — paper-trading.yml:96, monthly-revalidation.yml:58,
retrain-model.yml:48 — and is referenced nowhere in src/ (grep over
the source tree returns nothing). It is vestigial but HELD, and FMP documents an
earnings-calendar endpoint.
The original text (research/RESEARCH_QUEUE.md:2350, frozen archive) therefore
reads: "this dismissal is PROVISIONAL: one-time owner check needed (is the
secret live? does its tier serve the calendar endpoint?) — if yes, (9) converts
to a live scout candidate for the full four-bar test; if no, the kill line
completes as 'FMP key held but dead/insufficient-tier' and stands."
The frontier watch logged it as still-unrun on 08-18, 08-19, 08-20 and 08-21,
then stopped mentioning it. No probe commit, no memo, no research/ file
outside the archive mentions FMP or an earnings calendar.
The structural complaint, which is the reason this file exists. The
2026-08-21 queue-as-directory migration's whole memory guarantee is that a panel
can ls research/queue/dismissed/ and see every adjudication with its tripwire
attached. A conditional kill reachable only through a frozen archive is
exactly the rot ops/RESEARCH_PANEL.md "What matters here" #2 names — a
dismissal that future panels skip instead of evaluate. Same failure class as
BTB-1: a recorded basis the shop's own artifacts cannot check.
The owner's one-line question
Is
FMP_API_KEYstill live, and does its tier serve the earnings-calendar endpoint? (Onecurlagainst the documented endpoint with the held key answers both.)
Reopen condition — with a dated default, so it cannot rot again
- If YES (key live and tier serves the calendar endpoint): this converts to a live scout candidate and must clear the full four-bar test (un-backfillable · ~zero marginal cost · mechanism sketch · activation pin AND sunset) on its own merits. It must additionally clear the same scout run's "exhausted large-cap event space" finding, which is what killed the Alpaca corporate-action announcement feed — a shared reason that does not disappear just because a credential turns out to work.
- If NO, or if UNANSWERED by 2026-10-12 (the scout's own cadence date): the kill line completes as "FMP key held but dead/insufficient-tier" and this dismissal stands permanently, with no further owner action required.
Related and not separately proposable before this resolves (C0): the IPO-calendar-as-expected sibling, recorded riding this same provisional kill in scout run 3 (2026-08-19).
ANNOTATION 2026-10-16 (panel run 39) — the dated default FIRED; this kill is now PERMANENT, and it takes the IPO-calendar sibling with it
Plain-language summary. This item was killed provisionally in 2026-08, with one escape hatch: if the shop's Financial Modeling Prep key turned out to be live and to serve the earnings-calendar endpoint, the idea came back as a live scout candidate. The file set a deadline for finding out — 2026-10-12 — and said that if nobody had answered by then, the kill becomes permanent with no further owner action. Nobody answered. It is permanent.
The clause, verbatim (:80-82):
- If NO, or if UNANSWERED by 2026-10-12 (the scout's own cadence date): the kill line completes as "FMP key held but dead/insufficient-tier" and this dismissal stands permanently, with no further owner action required.
Verified UNANSWERED at HEAD. grep -rl "FMP_API_KEY" returns only
config/.env.example, ops/windows-host/README.md, RUNBOOK.md,
scripts/finalize_afk_session.sh, this file and its public twin, plus the
frozen research/RESEARCH_QUEUE.md. There is no probe commit, no memo, and
no code anywhere in research/ or src/ that tested the key or the
endpoint. The deadline passed with the question untouched.
Why it went unanswered is worth one line, and does not change the outcome.
2026-10-12 was the first day of a four-weekday outage in which six of seven
cloud routines were dark (research/2026-10-15_every_weekday_routine_went_dark_*),
so no panel ran on the date itself. That does not rescue the clause: the
condition was "UNANSWERED by" a date, and a probe that nobody ran in the
eight weeks before it was no more run on 10-12 than on 10-16. The default was
written precisely so that silence resolves it, and silence is what it got.
Consequences, both pre-registered in this file and now both executed:
- The kill line completes as "FMP key held but dead/insufficient-tier" — recorded, permanent, no owner action required.
- "Related and not separately proposable before this resolves (C0): the
IPO-calendar-as-expected sibling, recorded riding this same provisional
kill in scout run 3 (2026-08-19)" (
:84-86). That sibling's blocker has now resolved against it, so it is permanently closed on the same grounds. Re-proposing either without new data is a C0 breach.
Cost of this closure: zero trial rows, zero DSR debt, zero owner time. Per
NORTHSTAR.md §1 asset 3 and gauge G3, "a cheap permanent 'no' (the A3
pattern) is a win", and null results count at full value. This is the
cheapest possible form of one: a question closed by a default pinned two
months before its date, with nothing spent to close it.
No reopen condition. This annotation deliberately leaves none — the file's own text makes the kill permanent, and a permanent kill with a resurrection clause is not permanent. The only route back is genuinely new data (a different vendor, or an un-backfillable expected-earnings panel nobody has found), which would be a new slug and a new four-bar scout test, not a reopen of this one.