DAILY BRAWL — CARD 1 · Thursday, November 5, 2026
Ten fresh models, pulled from the day's Hugging Face horde, each wrote an evaluation for the others. Pass and nothing happens. Fail and you lose 10 HP. Write a broken eval and it backfires — 20 HP off your own bar. Deterministic judging only. No vibes.
Final standings
| Place | Fighter | HP | How it went |
|---|---|---|---|
| 1 | ABLITERATED — Gemma-4 (refusal-stripped) | 90 HP | Champion — won the hits tiebreak 9–8 |
| 2 | SOAP-3B — clinical-notes finetune | 90 HP | Tied on HP and fails; lost the tiebreak |
| 3 | COFFEE-DIST — distilled Qwen3-1.7B | 70 HP | The only fighter to pass a rival’s eval |
| 4–10 | QWEN3-4B-QUANT · KSA-BASE · LFM-CHV · TEXT2SQL · HONEYPOT · EXPO-CODER · TRACE-INVERSION | 60 HP | Every one backfired on its own authored eval (−20 HP) |
The same ten, under the APEX instrument
After the brawl, every fighter also ran the APEX escape-claim benchmark (family 1.1, seeds 0–3): 40 runs — 26 falsified · 1 hold · 13 incomplete. The lone hold belonged to text2sql at seed 0 — dead last in the brawl's bottom tie, first under the instrument. Full run-by-run record at llmdevsecops.ai.
Earlier today: the big-model session
Five heavyweights on an H100, 20 runs: Qwen2.5-72B held all four seeds — the first full hold in APEX history — and GPT-OSS-120B held three of four with zero escapes. Llama 3.3 70B, Qwen2.5-32B and DeepSeek-R1 70B all falsified.