Stage EU — the European researcher

One signed number per company, so today's European names can be ranked. No call, no
threshold, no direction label. Identical question and identical output contract to the US
and Japanese stages, which is the point: if three regions are scored the same way, a
European result means something next to the other two.

This stage places no orders and reads no broker. There is no alpaca_trade.py step
here and there must not be one. Alpaca carries none of these ten venues anyway, and
execution is a separate decision to be made later on Europe's own resolved numbers.

Ten markets, one stage, and why

The regulatory spine is genuinely shared — MAR Article 17 ad-hoc disclosure, the
Transparency Directive's annual-plus-half-year floor, the Short Selling Regulation's 0.5%
public threshold — but that is not the reason. The reason is measured: no European
market is a daily market on its own. The UK bottoms out at one eligible name on a
Friday; Germany puts 153 of 259 forward events in November and carries one in June;
France is semi-annual with 94 of 161 in February–March. They peak in different months, so
pooling seasonal calendars is what produces one stream. See
researcher_europe/SUBMARKET.md.

Seven markets were added on 2026-09-19 — Stockholm, Copenhagen, Oslo, Helsinki,
Milan, Madrid and Warsaw — and the cap went from 12 to 20 with them. The measurement
that justifies it is seasonal, and the obvious one says the opposite: over the ten
sessions 2026-09-21 → 10-02 the seven add six names, because late September is the
UK's month. By month of forward vendor events:

month	UK+DE+FR	the new seven
Sep 2026	161	13
Oct 2026	114 — the existing stage's thinnest	456
Nov 2026	309	515
Feb 2027	133	38
Mar 2027	250	70

And the cadence is better: measured median gaps of 91–98 days across the Nordics and
Poland against the UK's 217 and France's 204, so a Nordic name recurs four times a year
where a UK one recurs twice.

The ten are NOT equally instrumented, and this is the first thing to say in any note

eu_market.CAPABILITY is the measured table and the code reads it rather than assuming.

markets	short register	day archive	event_occurred: false
uk	FCA, with history	Investegate, by date to 1999	yes
no	dated event history per issuer	Oslo NewsWeb, true day query, ticker-keyed	yes
fr	AMF, per-holder to 2012	AMF flux, issuer's own category	yes
it	CONSOB xlsx, WAF-retried	eMarket STORAGE, WAF-retried	yes
se dk fi	national, snapshot	Nasdaq Nordic, paged, no date query	only inside the feed's ~12-day window
de	Bundesanzeiger, snapshot	EQS, per issuer only	no
es pl	none reachable	none reachable	no

Norway is now the best-instrumented market in this stage after the UK — it is the
second of the ten whose history carries observed announcement dates rather than a
cadence estimate.

Spain and Poland are the weak legs and must be described as such. A Spanish or
Polish name is hunted with no positioning anchor: its priced_lean_pct falls back to
the 20-day run-up, which is also the free control the whole stage is measured against, so
that name cannot beat the benchmark with anything that uses it. And it can never be
confirmed or killed after the fact. Both were given the eight-try standard that rescued
France and Italy and failed it. If a pooled ρ is quoted without saying how much of it is
es/pl, it is being oversold.

What is different from the other two stages, and it matters

	US (stage E)	Japan (stage J)	Europe (stage EU)
Calendar	Nasdaq vendor feed, 20/20 phantom on one day	JPX, the issuer's own notified date	TradingView vendor feed, 2 of 90 phantom measured against RNS; day archives exist for eight of the ten. Forward-date coverage is very uneven: 0.88 of Finnish and 0.81 of Norwegian issuers carry one, against 0.20 Italian, 0.19 Spanish, 0.13 Polish
Session	mixed bmo/amc	all amc	89% bmo — close(D−1) → close(D)
Option anchor	implied move + 25d skew	none; JPX shorts + 信用倍率	none; eight national short registers, two markets have none at all
Tail	uncapped	値幅制限 truncates	uncapped — maxima 42% / 27% / 52%
Names per day	17–22, all hunted	8–125, capped at 25 by random draw	65 eligible on 2026-10-22; capped at 20
Turnover floor	$200k	¥30m (~$200k)	$200k since 2026-09-19 (was $1m); below $1m the register names 12% of issuers, and anchor_covered carries that
Hunters	one, English	one, Japanese	seven definitions over ten markets, one bilingual pass each

Europe reports before the open. 339 of 379 measured UK results announcements landed
before 08:00 London. So the baseline for a bmo name must be sealed the evening
before the print, and eu_universe.py --date defaults to the next calendar day rather
than today. Getting the session backwards roughly halves the move being predicted.

The option anchor was the hope and it did not materialise. Yahoo carries zero
single-stock chains for .L, .DE and .PA; Eurex's free reference file has no
settlement prices, no open interest and no underlying map. Europe runs in the same
anchor-less regime as Japan, which is the regime that produced ρ=+0.073, p=0.45 over 104
events on the sealed corpus (backtest/FINDINGS.md §33). Say so in every note.

What Europe does have is a better short register, and eight of the ten read. The FCA resolves on 89% of the $1–5m turnover band against 36–44% for
Japan's JPX register, and publishes a per-holder history back to 2012 where JPX's rolls
off. Germany works with a cookie. France works too — www.data.gouv.fr is
intermittent rather than blocked, so eu_positioning.load_fr() retries it for the
resource URL and pulls the file from object-api.infra.data.gouv.fr: 40,696 per-holder
rows since 2012, 74 issuers with an open position, and publication end dates that make
the change a measurement. eu_resolve.py still reports lean_vs_free_control_rho per
market — if it climbs back toward 1.0 for a market that was below it, that register has
stopped resolving. It should read ~1.0 for Spain and Poland by construction, because
those two have no register at all.

The seven added in 2026-09 bring five more registers: Finansinspektionen (SE),
Finanstilsynet (DK at a 0.1% threshold, not 0.5% — a Danish aggregate is not on the
same scale as anyone else's), Finanstilsynet (NO, a dated event history and the best of
the ten), Finanssivalvonta (FI) and CONSOB (IT, behind a retried WAF). Sweden, Denmark,
Finland and Italy publish the current position only, so their short_change_pct_pts
is null until two days are cached — and null there means not known, never no change.

Steps

Resolve paths with python3 scripts/run_paths.py --json. Re-read the clock with
date -u; do not trust the date you were told at startup.

0. Heartbeat, before you spend anything.

python3 scripts/run_log.py --heading "Stage EU — Europe researcher — STARTED" --line "<the plan>"
scripts/publish.sh "stage EU: started for <YYYY-MM-DD>"

One cheap commit, and the only thing that distinguishes a Routine that never fired from
a session killed on its first subagent.

1. Universe. <RUN> is research/<YYYY>/<MM>/<DATE>/europe/.

python3 researcher_europe/scripts/eu_universe.py --date <EVENT-DATE> -o <RUN>/universe.json

<EVENT-DATE> is the day the print lands, which for a bmo name is the day AFTER the
session you are sealing against. It reads the vendor calendar for all ten markets,
drops anything below $200k a day of turnover (normalised to USD off a live FX rate in
six currencies, written into the file — it was $1m until 2026-09-19), and if more than
cap survive (20 since 2026-09-19) takes a random sample seeded by the date.
Report selection.method, eligible, eligible_by_market, hunted and by_market in
the note.

Read selection.market_concentration and put it in the note. The draw is random and
the calendar is seasonal, so a day can be almost entirely one market — 15 of 20 names
were Swedish on 2026-10-22, because October is Sweden's month. That is a correlated
exposure the scorer cannot see, the same shape as the four US names the IEEPA tariff
refunds ranked together. The draw is deliberately not stratified by market, because a
per-market quota is a second selection and this stage has already paid once for a cut the
scorer could not see — so the concentration is reported instead.

The floor matches the US and Japanese stages so all ten markets are cut the same way.
It also means most names now arrive with no positioning anchor: below $1m/day the
national short register names 12% of issuers against 89% in the $1–5m band. Every
baseline carries anchor_covered and the note must say how many names had one.

The draw is random on purpose: any other cut is a second ranking the scorer cannot see,
and the US run has already paid for that once. It matters twice here — Phase 1 measured
sell-side coverage at 5–7 analysts in the $1–5m turnover band against 16–19 above $25m,
so cutting to the "under-read" band would bake this stage's own thesis into its universe
and make it unfalsifiable. The floor is for capacity and anchor coverage; the thesis is
tested afterwards, by band, in eu_resolve.py.

If scheduled_today is 0 on every market, check per_market[*].market_closed — the ten
holiday calendars differ and are not the same as the national public holidays (XETRA
trades on Fronleichnam; Euronext Paris does not close for every jour férié; Stockholm and
Helsinki shut for Midsummer Eve, which moves). Only the UK's calendar is fetched live;
the other nine are hand-entered from each exchange's published trading calendar, so a
wrong entry is possible and shows up as market_closed in the universe file rather than
only in the code. Publish the empty universe, say why, and stop.

per_market[*].off_primary_exchange_filtered counts rows dropped because the vendor's
country scanner pools venues: sweden carries 230 NGM names beside 663 OMXSTO ones, and
poland carries 325 NewConnect names beside 384 GPW ones. NGM names do not take Yahoo's
.ST at all, so they would be screened on somebody else's tape.

2. Seal the baselines. Before any hunter.

python3 researcher_europe/scripts/eu_priced_in.py --universe <RUN>/universe.json --out-dir <RUN>/baselines

Sealed means sealed. Nothing downstream may revise a baseline.

3. One hunter per name, in the market's own agent. Dispatch on each baseline's
submarket field — the mapping is MARKETS[<submarket>]["hunter"], so read it rather
than remembering it:

submarket	agent
uk	unpriced-hunter-uk
de	unpriced-hunter-de
fr	unpriced-hunter-fr
se dk no fi	unpriced-hunter-nordic — tell it which one, it covers four
it	unpriced-hunter-it
es	unpriced-hunter-es
pl	unpriced-hunter-pl

Run them in waves of europe_hunt.wave_size (5), publishing after each wave. Give each
hunter only its own ticker, the window, and the path to its own baseline. Never another
name's baseline, never another hunter's findings, never your own view.

A wave that completes is banked. If a whole wave returns empty, stop and write the
— HALTED section rather than starting another.

4. Score. The scorer is the US one, unchanged and deliberately so:

python3 researcher_us/scripts/edge_score.py --run <RUN>

The ranking key is whatever edge-scores.json reports in ranking_key. Read it; do not
carry a remembered contract into the run.

5. Note. Write <RUN>/europe-note.md, answer first: the ranked table with the market
on every row, then the finding and URL driving the top and bottom name, then the names
that could not be ranked and why.

The note must also say, every time:

• selection.method, by_market, and that the draw was random among the eligible names
• which names carry session_unresolved, because their window was assumed rather than
  known and getting it wrong roughly halves the move
• that options is null in all ten markets and Europe runs in the anchor-less regime
• which markets' short registers resolved, and how many names carry
  anchor_covered: true — a name the register was read for but does not name is a
  truncated zero, not an anchor, and under the $200k floor that is most of them
• how many of the day's names came from Spain or Poland, which have no positioning
  anchor and no way to be confirmed after the fact, and how many from Germany, which
  cannot reach event_occurred: false either
• selection.market_concentration — if one market is most of the day, say so before any
  pooled number
• lean_vs_free_control_rho per market from the previous resolved run if there is one
• that the lean's weights are priors borrowed from the US runs, measured nowhere in Europe
• which names carry an estimated history.basis — a scale, not a record of dates.
  Only UK names (observed_rns) and Norwegian ones (observed_newsweb) are real; the
  other eight markets are cadence estimates, and reading a cadence prior as evidence is
  how TRT got ranked, traded and never reported
• which names sit above conviction_floor, and that over the whole US sample the sign
  was a coin flip below it
• that one day is not a result

6. Publish.

python3 scripts/update_index.py
scripts/publish.sh "stage EU: Europe ranking for <YYYY-MM-DD>"

Resolving, the next session

python3 researcher_europe/scripts/eu_resolve.py --run <RUN>

Confirms each release, measures the correct window per session (and BOTH windows where
the session was unresolved), and reports Spearman against the realised move with a
permutation p, the controls, every lean component separately, per-market statistics and
the language-pass control — which reports 0 names on any run sealed after 2026-09-22,
because the hunters stopped freezing an English-only draft that day. An empty section
there is the report, not a fault.

Eight of the ten markets have a day archive (eu_archive.py): the UK on Investegate,
France on the AMF's info-financiere.gouv.fr flux, Germany on the EQS-News search,
Sweden/Denmark/Finland on the Nasdaq Nordic disclosure feed, Norway on Oslo Børs
NewsWeb and Italy on eMarket STORAGE. France and the Nordics carry the issuer's own
release category, so their results are classified by what was filed rather than by a
headline keyword.

Three things differ by market and the resolved file states each one per run
(confirmation[*].false_reachable, archive_kind, archive_read):

• Germany, Spain and Poland can never reach event_occurred: false. EQS has no
  whole-day query, so a German name is searched by issuer and "not found" cannot be told
  from "the search term missed". Spain and Poland have no archive at all.
• Sweden, Denmark and Finland can reach it only for a recent print. The Nasdaq feed's
  fromDate is accepted and ignored — it returns the most recent rows for any date
  you ask for — so the code never passes it and pages back instead, roughly two days a
  request and about twelve days before it stops being cheap. Beyond that window the
  archive returns None and every row resolves null. So a Nordic run should be
  resolved within about a week.
• Italy is behind a WAF that answers about 7 of 8, and eMarket STORAGE's data_to is
  exclusive — a naive single-day query returns zero rows and looks exactly like a
  silent day.

day() returns None, not [], where a source could not be read. That distinction
is the only thing between "nobody announced anything" and "I could not look", and only
the first may ever support a kill.

The language experiment, and why it stopped

The hunters run ONE pass in English and the local language together, since
2026-09-22. Until that day each ran an English pass first, froze it as pre_local,
and only then searched locally; edge_score.py carried
diagnostics.impact_sum_pre_local beside the key and eu_resolve.py ranked both
against the same realised move, so "searching in German and French earns rank
correlation" was a measured claim rather than a believed one.

It was retired on the operator's instruction, and the reason is not only tokens.
Sequencing the two halves forbade them from informing each other — a German filing is
often the reason to run a particular English query, and an English wire item the reason
to go and find the German original — so the control was being paid for out of the
quality of the research it was measuring.

What that cost is worth stating plainly. Nothing measures the local half any more.
It cannot be recovered from a run after the fact, because a freeze reconstructed later
is not a freeze. The one thing that softens it: no European day had resolved while the
control ran, so spearman_pre_local never produced a number against a real outcome.
What was given up is a future measurement, not a result.

What replaces it is prose. Every hunter emits language_note: one line per thing
the local-language sources carried that the English ones did not, or the single line
'nothing the English sources did not already carry'. Nothing ranks it. The note should
quote it where it is interesting, and should not present it as evidence.

The UK case was degenerate and the asymmetry survives. Its local language is
English, so its local half varies source locality (RNS, Investegate, the domestic
trade press) rather than language, and a UK language_note saying the domestic sources
added nothing is an honest result rather than a lazy hunt.

The Nordic case is weaker than the others and the hunter is told so. Nordic issuers
publish in English as a matter of routine — most releases go out in both languages at
once, which is not true in Germany, France or Italy. A Nordic language_note that reads
'nothing the English sources did not already carry' is the expected answer there more
often than anywhere else in the stage.

eu_resolve.py still reports the old control and will read 0 names. That is the
honest report of a control that no longer runs, not a hunter that forgot to freeze. The
runs sealed before 2026-09-22 keep their diagnostics and the dashboard's Taal tab
still shows them, labelled as history.

Budget

config/pipeline.yaml, europe_hunt. When the day would exceed the cap, shed by
europe_hunt.degrade_order and record what you shed in the run log. Half a run that
finishes beats a full one that gets cut off.

The rule that outranks the rest

Never fabricate a number. Every company-specific figure carries a source URL or is marked
unavailable/null. This matters more, not less, when the source is in a language the
reviewer may not read — a translated paraphrase with no original string is
indistinguishable from an invented one, so every non-English finding quotes the original
beside the translation.

This is research, not investment advice. Keep the disclaimer from config/pipeline.yaml
on the note.