Case G

Resolved

All forecasts
Yarrow

Turns uncertainty into forecasts by modeling reality,
simulating behavior, and learning from outcomes.

Monetary policyEuro areaResolved benchmark

The transfer win:
nine memory-clean
ECB meetings

The question

Does an evidence-pack recipe built on one central bank still work when it is pointed, unchanged, at another?

The benchmark — nine European Central Bank rate decisions, discovered by archive scan and graded mechanically. Each row was priced twice: once from the model's own knowledge, once with a frozen evidence pack built by the same rules used on the Federal Reserve.

Bare panel 0.208 mean error over the nine meetings, no evidence pack
With frozen packs 0.131 the same nine rows · error cut by more than a third
Rows the packs won 6/9 three rows went the other way · all nine are in the table below
Memory screen 72/72 recall cells unknown · screening families A–D, two reads each

Brier error: squared distance between the forecast and what happened. 0 = perfect, 0.25 = coin-flip guessing. Lower is better. Both columns are the panel's median probability that rates would change, scored against the decision the ECB published.

01 / The setup

A recipe, or a coincidence?

The evidence-pack recipe was built on the Federal Reserve. A method that only works where it was tuned is not a method. Transfer is the test.

So the pipeline was pointed at a second central bank with nothing changed: the calendar found by archive scan, the outcomes chain-verified from the ECB's own statements, the packs assembled by the same mechanical rules.

How the benchmark was builtNo step was tuned for the ECB

  1. Pack rules were developed and registered on Federal Reserve decisions.

  2. Nine ECB decision dates found by archive scan, not chosen by hand.

  3. Each row's evidence is dated before its own decision. Same rules, no per-institution tuning.

  4. Outcomes chain-verified from the published decision statements.

One setup fact does the heavy lifting: models can recite Federal Reserve history, but they recall nothing of this ECB span. The memory screen below returns 72 unknowns out of 72 cells, so every row here is a fair test rather than a recollection.

02 / The frozen evidence

One dated line
from one meeting's pack.

Shown for the 5 February 2026 decision. Its pack opens with the previous meeting's statement — a public document, quoted as frozen. How packs are assembled stays private.

01Official
European Central Bank
The Governing Council today decided to keep the three key ECB interest rates unchanged. Its updated assessment reconfirms that inflation should stabilise at the 2% target in the medium term. The new Eurosystem staff projections show headline inflation averaging 2.1% in 2025, 1.9% in 2026, 1.8% in 2027 and 2.0% in 2028. …

This is the line the panel read. Nothing in it says what the Governing Council would do seven weeks later.

Excerpt only. The full frozen evidence pack is preserved, hashed and timestamped in the audit trail; how packs are assembled stays private. Full record available for verification on request.

03 / The panel

Nine rows.
Both readings of each.

Every row the benchmark contains is here, including the three where the evidence pack made the answer worse. Closer to the marked outcome is better.

Bare panel · mean error0.208

Asked from the models' own knowledge

With frozen packs · mean error0.131

Evidence assembled by the Fed recipe

Bare median Packed median What happened Each track runs 0 → 1. The outcome sits at one end; the shorter the gap, the better the read.
Meeting Position on 0–1 Bare Packed Verdict
24 Jul 2025HELD 0.1600.0256 0.3150.0992 PACKS LOSE-0.0736
11 Sep 2025HELD 0.4250.1806 0.1800.0324 PACKS WIN+0.1482
30 Oct 2025HELD 0.4000.1600 0.2400.0576 PACKS WIN+0.1024
18 Dec 2025HELD 0.3850.1482 0.2750.0756 PACKS WIN+0.0726
5 Feb 2026HELD 0.5800.3364 0.2250.0506 PACKS WIN+0.2858
19 Mar 2026HELD 0.5900.3481 0.2300.0529 PACKS WIN+0.2952
30 Apr 2026HELD 0.3550.1260 0.1800.0324 PACKS WIN+0.0936
11 Jun 2026RAISED 0.3150.4692 0.1900.6561 PACKS LOSE-0.1869
23 Jul 2026HELD 0.2850.0812 0.3500.1225 PACKS LOSE-0.0413

Each cell pair is the same panel answering the same question on the same date. The only difference between the two columns is whether the frozen evidence pack was supplied. Packs win 6 of the 8 rows that held; they lose two holds and the June hike, and those three rows are on the table rather than dropped from it.

04 / Why this was rare

Most pipelines are tuned
to the place they were born.

The claim here is narrow and therefore checkable: the same rules, a second institution, a span the models cannot recall, and a bar registered before the data existed.

The rules did not change.

No per-institution tuning, no hand-picked calendar, no reweighting after seeing the outcomes. The ECB benchmark ran the Federal Reserve recipe as written.

One recipe, two institutions
FEDsame rulesECB

The span is unrecallable.

Asked bare, no analyst could state what the Governing Council decided on any of these nine dates. That is what makes the packed column evidence rather than memory.

Recall screen

The bar was set first.

The registration commit predates the benchmark build, so the comparison could not be arranged after the fact. Results were recorded to their own commit.

Commit order
db09b92registeredd8c91e9results

05 / The resolution

All nine outcomes,
from the ECB's own statements.

Each decision is quoted as published and linked to its own release. Eight of the nine carry the identical sentence; the ninth is the one that moved.

One meeting · rates raised

11 Jun 2026
The Governing Council today decided to raise the three key ECB interest rates by 25 basis points.

This is the row the packs lost. The bare panel priced a change at 0.315 and the packed panel at 0.190 — the pack, faithfully reporting a Governing Council that had held seven times running, pointed away from the move.

European Central Bank · 11 Jun 2026

Calibration lab

Price the February decision yourself.

You have read the same line the panel read: the 18 December 2025 statement above, and nothing dated after it. Set your probability that the ECB changed rates on 5 February 2026.

Error is (your probability − outcome)². The realized outcome is no change = 0; lower is better.

50%

Probability input← Drag to set →

1% · certain no change99% · certain change

Outcome
NO — rates unchanged
Your error
—
Packed panel
0.051
Bare panel
0.336

How we kept it honest

The span was screened for memory
before it counted.

A transfer result is only worth reporting if the second institution is genuinely unfamiliar to the models. Every row was asked bare first, with no pack, to see whether the outcome was already in the weights.

Memory screen 0/72 No model recalled any of the nine.

Asked with no evidence pack, all 72 recall cells came back “unknown” — screening families A–D, two reads each. The whole grid is below.

The transferred recipe · mean error 0.2080.131 Bare panel, then the same panel with the packs.

Same nine ECB meetings, same seats. The only thing added was an evidence pack built on the Federal Reserve.

No row was dropped

All nine rows are on the table above: the three the packs lost are displayed with the six they won.

Evidence freeze · per row

Each row's pack is dated before its own decision. The excerpt shown is the 18 Dec 2025 statement, read for the 5 Feb 2026 meeting.

Registered before the benchmark was built

Committed as db09b92 before the benchmark existed, so the bar could not be moved afterwards; results recorded at d8c91e9.

The proof · 9 meetings × screening families A–D × 2 reads72 cells

a valid answer, and it was “unknown” · all 72 cells

screening family Ascreening family Bscreening family Cscreening family D

Asked with no evidence pack: “what did the Governing Council decide at this meeting?” Every cell came back unknown. Two reads per family, so a single lucky refusal cannot carry a row.

Pass bar — set before the run; wording condensed here

An evidence-pack recipe developed on the Federal Reserve, applied without modification to a mechanically discovered ECB benchmark, should reduce panel error on a span the models cannot recall.
Technical receiptRegistration, results and artifact family
Registration
db09b92

Committed before the benchmark build.

Results
d8c91e9

Recorded result set.

Artifact family
research_artifacts/policysim/ecbcal1/

Calendar, outcomes, anchors, recall screen and panels.

Row table
ecb_panels.json

Bare and packed medians for all nine rows.

Outcome chain
ecbcal_outcomes.json

Decision statements and their archived sources.

Recall screen
ecb_recall.json

The 72 bare-recall cells reproduced above.

Pack record
ecbcal_anchors.json · ecbcal-2026-02-05

Frozen evidence for the meeting shown above; page shows an excerpt only.

Verification trail

The separate public trail repository is being prepared. Verification materials are available on request.

Request the trail