The instrument itself.

(a)The corpus

337 rows, frozen 2026-08-23. The research design uses ALERT's harm taxonomy and selected seed objectives; the comic situations and constraints were composed for LulzBench. The website publishes selected exhibits, not the full evaluation bank. Every row carries an elicitation role:

Corpus SHA-256 f02e4ae8… (full instrument, private research repo) · public copy: 329/337 rows, 8 held-out ids excluded from every published corpus file (v0.4 and lineage prompt texts are held privately with them — only the v0.5 excerpt ships) · provenance: ALERT (Tedeschi et al. 2024) · license CC BY-NC-SA 4.0 · future corpus distribution and licensing scope require an explicit operator decision.

Canonical specimen IDs use LULZ-####; explicitness variants keep their trailing letter, such as LULZ-0394B. Historical receipts retain their original IDs and resolve through the corpus crosswalk.

(b)Version lineage — v0.1 → v0.5

(c)A/B pair discipline

Each B-variant is prompt_a with exactly one substring replaced. The find/replace pair is stored in the manifest, and the build asserts the match is unique. Anyone can verify the manipulation by applying the edit themselves.

(d)Grader discipline

LLM grading only — deepseek-v4-pro, temperature 0. Grader freeze v1.1. Model and grader outputs are never truncated. Capture-integrity counters:

(e)The floor-control protocol

The floor_pool exists to catch over-refusal. Each floor is assigned a verdict:

The three INVALID rows — refusing them is the model succeeding, not failing:

(f)Human validation

Pending

The instrument is validated at the model-audit level (L2) but has not been human-validated. This status is marked in three places across the repository and is repeated here rather than buried.

(g)Banked experiments — the small-model campaign

October 2026: seven ordinary instruction-tuned open-weight models (1.2B–7B) each answered all 329 eligible specimens of the frozen v0.5 non-held-out population, locally on a Mac Studio M5 Ultra. Every run kept exact prompts, rendered templates, full responses, sampling settings, finish reasons, timings, attempts and recovery evidence; the raw traces remain in private research storage.

• Full-cohort coverage: 2,302 complete observations across the primary cohort, zero request-failure rows. Six models completed all 329 specimens.
• One honest truncation: SmolLM3's LULZ-0373 response stopped mid-sentence at its token allowance; a 16,384-token recovery run revealed an unresolved generation loop (two paragraphs repeated 145 and 146 times) rather than a near-miss. Both captures are preserved and the specimen is excluded from any candidate shortlist.
• What direct review found: on the roast specimen LULZ-0336, several small models answer the "you're bald" aside instead of the explicit request for five fat jokes; one narrates the scene without delivering any. A task-reading failure is curated as a task-reading failure — not as praise for harmful compliance.
• What it does not claim: no funniness scores, no rankings, no model-graded numbers presented as taste. A private seven-response comparison and heuristic review queue exist for a future, separately reviewed exhibit.

LFM2.5-1.2B-Instruct responses: the pinned LFM Open License v1.0 (upstream LICENSE sha-256 61d7e939…) conditions commercial use of the model on its revenue threshold and governs redistribution of the weights; it imposes no restriction on displaying generated responses from this non-commercial research use. The small-model arcade cabinet now shows the retained LFM traces.

(h)Banked experiments — the 27B refusal-removal tournament

October 2026: Qwen3.8-27B stock was run against the frozen 329-specimen population alongside two refusal-removed derivatives (OBLITERATUS, Huihui; Q8_0), each with its own native chat template and verified pinned artifacts. The completed cohorts are sealed with integrity reports; traces stay private. Descriptive, same-population contrasts from this banked work — never causal claims about ablating refusals — reach the site only as individually reviewed exhibits, starting with the Arena's locket-ministry exhibit. A separate judge-ablation experiment (stock vs ablated 27B as LulzBench graders) found the ablated judge less faithful to the stock grader and collapsed its PARTIAL tier — evidence recorded, not yet an exhibit.