echr-livehrb-stateswap

State-swap counterfactual set for LiveHumanRightsBench (experiment 6). Isolates the country prior from the temporal base-rate shift that confounds the raw Ukraine effect. Built from the 001-only canonical echr-livehrb-static-2k (regular, substantive articles). Facts are held fixed; only the respondent state identity varies.

Design (anonymize-then-fill)

Respondent-state identity is detected (country, demonym, cities, courts, authorities) and templated to [DEFENDANT STATE] / [DEFENDANT STATE ADJ], then filled into the body per arm. The swapped country reads naturally in the fact pattern (no separate prompt field). Applicant identity is held constant across arms.

Arms (4 per case, shared swap_group_id)

arm	respondent	role
control_original	real respondent state	ground-truth arm (violation_label applies)
control_neutral	Iceland	low-prior / low-salience control
probe_ukraine	Ukraine	Ukraine-specific prior under test
probe_russia	Russia	high-violation-prior control (UA-specific vs generic adversarial-state prior)

816 base cases x 4 arms = 3264 rows. Arm counts: {'control_original': 816, 'control_neutral': 816, 'probe_ukraine': 816, 'probe_russia': 816}.

Scoring

Score case_text_rendered (the arm's filled body). violation_label is the gold target, carried unchanged from the canonical case across all arms. Pair by swap_group_id:

• delta(probe_ukraine - control_neutral) = the UA prior (headline)
• delta(control_original - control_neutral) = the real-state prior (calibration)
• delta(probe_russia - control_neutral) = generic adversarial-state prior Bucket per article_full (substantive only here). Null hypothesis (model reasons from facts): the call does not move across arms; any movement is a country prior.

Scope / exclusions (base pool 907 -> 816)

• ex-Ukraine, single-respondent CoE states, substantive article_full only.
• Excluded cases whose real respondent is a swap target (Iceland/Russia) to avoid degenerate arms (-40).
• Dropped cases where the applicant's own nationality is legally material (expulsion / Art. 3 refoulement / Art. 8), which cannot be country-swapped (-47).
• 4 pathological long judgments whose identity-span list did not fit the detector's output budget (documented, not silent).

Columns

swap_group_id, item_id, arm, respondent, respondent_adj, case_text_templated, case_text_rendered (score this), article_full, article, violation_label (gold), respondent_original, anonymization_report, case_name, application_number, ecli.

Public HUDOC sources only -> redistributable. Verdict-free (contamination-controlled).

Changelog

v1.3 (2026-08-26) — verdict leak repaired

584 of 3264 rows (17.9%) still contained the Court's reasoning and have been re-cut. Verdict removal was marker-driven ("The Court's assessment" and similar); where no marker fired, the entire legal analysis survived into the field models are scored on. BALL v. ANDORRA retained 18,160 characters after THE LAW, including the conclusion, against a no_violation label.

The v1.1 leakage audit missed this because it was lexical. A lexical probe fails in both directions: it fires on domestic courts quoted in the facts ("the City Court finds no violation of Article 6") and misses the Court phrasing its own conclusion as "did not fail to fulfil its positive obligations". retention_percentage is not a usable filter either — affected rows appear at 49% and 66% retention.

The skew mattered more than the rate. Leakage was concentrated in exactly the dimensions this data is compared on: regular 18.5% against ukr 0.1%, no_violation 2.4–2.6x more often than violation, importance-1 judgments 21.3% against 4.9% for importance-4, and per-country rates from 52% to 0%. Measured impact on affected rows: no effect on violation cases, +14.3pp on no-violation accuracy, +7.1pp balanced.

What changed: text is now cut at the last THE LAW header following the last THE FACTS, ignoring table-of-contents repetitions. Only affected rows were touched — every other row is byte-identical to v1.2. Row membership is unchanged, so results from v1.2 remain comparable except where the text itself changed. verdict_free_length, retention_percentage and verdict_removal_method were recomputed to follow the text.

Reproduce the audit with scripts/leak_audit/audit_leak.py in the LiveHumanRightsBench repository.