The Arena is the foundation for future comparative challenges on comic-framing safety behavior. It holds no leaderboard, and it will not until a floor clears. This page explains why that is a feature.
A foundation for future comparative challenges on comic-framing safety behavior — a place where a floor can eventually be cleared and a comparison earned. In its current state it is a set of entry rules and an honest statement of standing.
Current challenges use LulzBench specimen IDs such as LULZ-0394B. Historical records keep their original identifiers, resolved by the corpus crosswalk.
The gemma4 shootout used one provider and one grader. The MiniMax run used a different provider — the report says so. That is exactly why the two are not placed on a shared axis here.
The instrument is validated at the instrument level (L2 model audit). Human calibration is pending. So nothing scores yet — there is no challenge, no entry, and no ranking to report. When that changes, the rules above apply first.
The first excerpt-only exhibit to clear its review. Three Qwen 27B checkpoints (stock, and two refusal-removed derivatives) received byte-identical prompts about a church bulletin, two Carols, and a folding chair. The excerpts below are verbatim; the full responses stay in restricted research storage because they contain unsupported historical or invented claims. Native chat templates differ between checkpoints, so this is a descriptive contrast, not a causal comparison.
Specimen LULZ-0375 · prompt SHA-256 36639002…cf6f85a
| Checkpoint | Verbatim excerpt |
|---|---|
| Qwen3.8-27B (stock, Q8_0) | |
| OBLITERATUS (refusal-ablated, Q8_0) | |
| Huihui (refusal-ablated, Q8_0) |
Excerpt-level research review verified the prompt hash, all three full-response hashes, and each excerpt as an exact substring. The excerpts contain no actionable instructions. One generation per checkpoint; this is not a humour score and not a safety-compliance result.
Every future number publishes its capture-integrity counters and denominator alongside it. Human validation of the instrument is still pending, so no scores or rankings appear here yet.