Habeas Protocol

At a glance

Empirical findings

All three tribunals score at near-ceiling on every per-ruling primitive, and 2/2 on both architectural system properties. The pattern is robust to two stress tests, and seven executable traces from the corpus demonstrate the protocol's methodological coverage across formula, deferred conditional, bounded discretion, arithmetic and Boolean composition, statutory partial-refusal, and a third-party-jurisdiction gate.

Inferential layer. Bootstrap 95% case-resampling intervals (10000 resamples of the judgments within each tribunal; they describe how much each mean moves with which cases are in the sample, and do not capture grader error): ADGM 1.91 [1.89, 1.94], SICC 1.85 [1.80, 1.90], DIFC 1.72 [1.62, 1.81]. All three pairwise differences exclude zero at α=0.05 — the ranking is statistically supported, not point-estimate noise. Raw values: data/bootstrap_ci.json; computation: scripts/compute_bootstrap_ci.py.

Construct validity (external correlate, recomputed 2026-10-05). No positive external correlate was found, so hypothesis H8 is not satisfied. Per-judgment v0.2 mean against subsequent-citation count: Spearman ρ = −0.04 (n=186, null). Against appeal status: ρ = −0.26 (n=163): the 10 first-instance judgments later referenced by appellate output score lower (1.70) than the rest (1.90). Raw values: data/robustness/external_correlate.json. An earlier release reported +0.32 and +0.12; those came from a rank formula that is invalid for tied data and are withdrawn (paper §4.12).

Of the 188 judgments scored, 39 form an LLM-graded first-pass set (32 DIFC + 7 ADGM, scored by Claude Sonnet 4.5); the remaining 149 entries (69 ADGM + 80 SICC) are scored by deterministic regex heuristics — no LLM in the loop for these 149. Per-entry grader-type and provenance are recorded in coding.grader_type, coding.coder, and grader-type-specific fields (model + prompt SHA for LLM entries; producing-script path for regex entries).

• Grader-type stability (ADGM, the only tribunal with both grader types). The LLM grader (n=7) scores ADGM at 1.93; the regex heuristic-triage (n=16) scores 1.93; the regex heuristic-graded (n=53) scores 1.91. The two graders agree to within 0.02 on the overall mean — the within-corpus evidence that the saturation finding is a property of the tribunal rather than the grading instrument.
• SICC PR4 heuristic limitation. The regex grader produces PR4 = 1.55 for SICC because the four-marker triplet test fails on narrative grounds-of-decision documents. The headline SICC mean of 1.85 uses this regex result and is a known-flawed measurement on PR4. A Claude re-grade of PR4 with a prompt instructed to read narrative form is staged and has not been run.
• Falsification cross-check. A 30-instrument falsification set across five non-court instrument classes (sealed awards, on-chain DAOs, regulator notices, platform adjudicators, UDRP panels) separates sealed awards, on-chain tribunals and platform adjudicators from the courts (18 of 30 instruments score low). Regulator notices (1.92) and specialised panels (1.78, the positive control) score level with the courts (1.83) by design: the rubric measures procedural form, and pedigree does not enter it. These 30 scores are author-assigned class defaults, not per-instrument measurements.
• Cross-family replication. The protocol crosses legal-family boundaries — Singapore common law via the IAA, vs DIFC's own statutes and ADGM's English-law-via-statute — and translates to a civil-law foil under the peer-court comparison set. The protocol is not court-specific.
• Architectural system properties. All three tribunals score 2/2 on separation of powers (SP1) and a reviewable appeal path (SP2). These are the structural pre-conditions for plugging software into the bench.
• Methodological coverage. Seven executable traces from the corpus show the protocol covers (i) static-rule arithmetic, (ii) deferred conditionals, (iii) rule-bounded human judgment, (iv) arithmetic composition over substantive findings, (v) Boolean composition over contractual interpretation, (vi) NY-Convention partial refusal under Singapore IAA s 31, and (vii) a third-party-jurisdiction gate under Norwich Pharmacal + Bankers Trust + RDC 28.52.

The tribunals already exist; what is missing is the computational layer.

The Atlas — 188 fingerprints

One sigil per judgment. Each fingerprint below is generated deterministically from the primitive scores of a single ruling in the coded corpus. Same scores → same shape; different scores → different shape. Six concentric rings encode PR1–PR6, two outer arcs encode SP1–SP2, and a hash-seeded rosette gives every case ID its own face.

The six per-ruling primitives

Properties of any individual ruling, scored 0 (absent) / 1 (partial) / 2 (fully implemented). v0.2 of the framework. Definitions live in data/primitives.json.

System properties

Architectural facts about the tribunal as a whole, not properties of individual rulings. Scored once per institution. The score-0 row is what you avoid by anchoring at DIFC or ADGM rather than at an ad-hoc Web3 arbitration project.

Seven working traces

Each trace lifts a real rule from the corpus into Catala source plus a Python evaluator and runs it against the case's event log. The seven span the full methodological spectrum: formula, deferred conditional, bounded discretion, arithmetic composition, Boolean composition, partial statutory refusal under Singapore IAA s 31, and a third-party-jurisdiction gate (Norwich Pharmacal + Bankers Trust). Three tribunals, three legal families, one engine. Trace #3 is the honest one — it shows what the rules cannot fully decide.

Trace viewer

Pick a trace. The left column is the rule as Catala source — the formal specification. The middle column is the event log — what happened, with the facts the human judge had to determine. The right column is the output — what the predicate produces when run against those facts, with assertions checked against the court's ruling.

All judgments