Foreign extradition case-law search

1. Establish what is actually being asked: which requesting state, which executing states, which ground.
2. Name the executing states. Never sweep everything by default — see Scope it.
3. Run the search (scripts/xsearch.py), or work the coverage map by hand where no shell is available.
4. Report hits as leads with links, grouped by jurisdiction, never as holdings.
5. Tell the user what did not answer, and what a nil return does and does not mean.

---

What this is for

A recurring problem in extradition defence and prosecution is showing how other executing states treat a given requesting state or a given ground: Rechtbank Amsterdam on prison conditions, German OLG and Bundesverfassungsgericht decisions on assurances, Irish High Court surrender judgments, Italian Cassazione, Polish SAOS. That material sits in a hundred-plus national databases, each with its own interface, language and vocabulary, and no commercial service searches them together.

This skill drives an open-source engine — 153 adapters, MIT licensed — that does.

It is not for the domestic case law of the user's own jurisdiction (use a national database or a commercial service), and it does not read, translate, or interpret the judgments it finds.

---

The rule that governs every output

Every result is a lead, never an authority. The engine reports that a judgment exists and that its metadata matched extradition vocabulary. It does not know what the judgment holds. Before anything from here reaches advice, a pleading or a court:

1. Open the decision and read the passage. In the original language, with a translation if it will be relied on. A machine-matched title is not a ratio.
2. Treat the relevance gate as sorting, not judgment. on_topic means extradition vocabulary appeared in the title or snippet. It is not a finding that the case is about the user's point.
3. Keep "failed" and "nil" apart. A database that errored was unreachable that run: its material is unsearched, not absent. The report separates them; so must the answer.
4. Read a genuine nil return carefully. Publication practice varies enormously — in several states the surrender decision is never published at all. "Nothing found" is not "no such decisions", and must never be reported as though it were.
5. Say what was not reached. Flag anything unverified rather than presenting it with more confidence than it has earned. The user remains professionally responsible for every authority they cite; this skill never discharges that.

---

Running it

One-off install, then search:

python3 scripts/xsearch.py install

# How are executing states treating Romania?
python3 scripts/xsearch.py search --issuing-state RO --countries NL,DE,IE

# A ground, in a state's own language, in its own courts
python3 scripts/xsearch.py search "detentie omstandigheden" --countries NL

# Non-arrest-warrant vocabulary — third-state extradition
python3 scripts/xsearch.py search --issuing-state TR --mode extradition

# Narrowed by date
python3 scripts/xsearch.py search --issuing-state PL --countries NL,DE --since 2023-01-01

Supporting commands: preview AE --countries RO,IT,PL (what a requesting state is actually searched as), sources --countries NL,DE (which databases exist and their status), runs (past searches), show <run-id> (a past report in full).

Options: --mode both|eaw|extradition, --since / --until, --limit N per database, --top N printed, --include-manual, --no-expand.

Without a shell — in an assistant that cannot run Python — do not pretend to sweep. Load reference/jurisdictions.md, identify the databases that cover the executing states in question, and work them one at a time with whatever browsing is available, using the localised vocabulary from preview's logic: the requesting state's name in that jurisdiction's language, plus that jurisdiction's extradition terms. Say plainly that this is a partial search.

---

How to search well

Scope it. An unconstrained sweep hits every database, and one source (pl-saos, Poland) takes about 98 seconds on its own while every other adapter answers in under five. Name the executing states whose practice is actually needed. If Poland is wanted, warn the user about the wait rather than letting a 90-second silence read as a hang.

Search by requesting state, not by keyword, when the question is about a state. --issuing-state RO localises "Romania" into each jurisdiction's language and combines it with that jurisdiction's extradition vocabulary in the source's own query syntax. A free-text search for "Romania" finds far less. Run preview first to show the user what will be asked — and if it warns that babel is missing, stop and install it: without it, localisation silently degrades and every requesting-state search narrows without saying so.

Read the arithmetic, not just the hits. "19 results — 8 on point, 10 off topic" means the sweep worked and that source is noisy. "20 results, 0 on point" usually means the database sorted by date and stemmed loosely, and the run established nothing: widen --limit, or go at it by free text in the local language.

Watch for deep-link sources. 104 of the 153 databases cannot be automated. Normally they are skipped. But if every database selected is deep-link only — several jurisdictions have nothing else — they are returned anyway, and a "hit" is a link to a court's search page, not a decision. Check the ACCESS column in sources before treating a thin result as a nil return.

Then do the real work. Open the judgment. Get it translated if it will be relied on. Test whether the foreign court's reasoning survives on facts like the user's — a refusal resting on 2019 prison-monitoring data does not carry itself into a later year without the intervening material.

---

Every search leaves a report

xsearch.py writes a dated markdown report and a JSONL file per run, indexed in runs.jsonl: what was asked, which databases answered, which failed, and every hit with its relevance verdict. This exists so that months later the user can say where a citation came from without re-running the sweep. Quote the run reference in the case file.

---

Politeness — a condition of use, not a suggestion

The engine queries public court databases on a deliberately low-volume footing: one request per second per host, an honest User-Agent, and no automation at all of sources whose terms bar it. An advisory lock prevents two sweeps at once.

Do not parallelise it, loop it unattended across a list of states, or raise --limit into the hundreds to bulk-collect. Several of these databases will block on far less, and they are a shared resource for every practitioner who does this work. If a user asks for bulk harvesting, decline and explain why.

---

Files

Path	What it holds
scripts/xsearch.py	Installs and drives the engine; applies the relevance gate; writes the report
scripts/quality.py	The relevance gate, carried here so the skill works against any version of the engine
reference/jurisdictions.md	Coverage map — every jurisdiction, its databases, automated or deep-link

Engine: https://github.com/mateiclej-wq/eu-extradition-search (MIT).