Minotaur-Brief

A pre-registered benchmark for long-form federal-court advocacy across five doctrinal architectures (Title IX deliberate indifference, Title IX retaliation, ADA Title II / Rehabilitation Act Section 504 accommodation, Pennsylvania common-law fraud, Pennsylvania common-law civil conspiracy) and two procedural postures (offensive partial-summary-judgment motion on Counts I–III; opposition to defendants' summary-judgment motion on Counts IV–V), at four context-noise levels including a cross-count condition.

Source: Mullins v. Duquesne Univ., No. 2:25-cv-01366 (W.D. Pa.). Released materials apply a role-based pseudonymization protocol to third-party individuals named in the docket; see §5.1 of the paper.

Pseudonymization v2 (2026-06-10): this revision completes the protocol: residual third-party name occurrences in prompts/exhibit_pool.json (all-caps exhibit headers, first names beside role labels, signature blocks) are now pseudonymized, and the treating physician's private-practice contact details are redacted. Third-party names appearing in the public docket are pseudonymized in this dataset, though unredacted in the public record. Named defendants and the plaintiff (the first author) remain named by design.

Contents

Prompts (prompts/)

• prompts/tasks.json — five (count, posture) tasks with required_elements, relevant_cases, relevant_statutes, parties_context, defendant_context, doc_type, and sumf fields.
• prompts/ground_truth.json — 119 (element, exhibit) mappings across 22 elements and 5 counts, plus designated cross-count pairings at L4.
• prompts/exhibit_pool.json — 38 verified documentary exhibits.
• prompts/context_ladder.json — four context-noise level definitions (L1_focused, L2_exhibit_noise, L3_full_noise, L4_cross_count).

Rubrics and provenance (rubrics/)

• rubrics/PREREGISTRATION.md — frozen rubrics, scoring scales, pitfall-to-score rules, version history (frozen 2026-05-07 prior to main-run launch).
• rubrics/doctrinal_rubric.json — Layer 2 doctrinal element rubric on a 0/1/2 schema (Claude Haiku 4.5).
• rubrics/writing_rubric_v2.md — Layer 2 writing-quality rubric, five 1–5 axes (Claude Sonnet 4.6), grounded in Garner (2014), Aldisert (2017), Edwards (2018).
• rubrics/GROUND_TRUTH_PROVENANCE.md — per-(element, exhibit) provenance covering all 22 elements and 119 mappings.

Scoring code (root)

• gates.py — Layer 1 deterministic gates (element-presence, citation-existence, exhibit precision/recall).
• scorer.py — orchestration and Layer 2/3 invocation.

Metadata

• croissant.json — Croissant metadata.

License

CC BY 4.0 on original contributions (prompts, rubrics, ground-truth maps, scoring code, structured JSON). Underlying federal judicial opinions are public-domain works of the U.S. government and are not covered by this license.

Code

Scoring code at https://github.com/dntstoplaughing-max/minotaur-brief.

Citation

Mullins, D. M. (2026). Minotaur-Brief: A Pre-Registered Benchmark for Long-Form Federal-Court Advocacy Across Doctrines and Procedural Postures. NeurIPS 2026 Datasets and Benchmarks Track (under review).

Disclaimer

Minotaur-Brief measures textual properties of long-form legal advocacy. It does not certify legal correctness or constitute legal advice. The underlying matter remains active at the time of release. See §10 of the paper for the conflict-of-interest disclosure.