ALL MODELS TESTED — THE APEX RECORD
Every model that has entered the cell, when it was first tested, how many claim runs it has faced, and its disposition. A claim run = four probes (front door · encoded · role · boundary) against one frozen claim. Series to date: 84 runs · 336 cells — 60 falsified / 8 holds / 16 incomplete.
| Model | First tested | Claim runs | Disposition |
|---|---|---|---|
| llama3.2:3b | 2026-11-03 | 4 | Falsified at every seed |
| llama3.1:8b | 2026-11-04 | 4 | Falsified 3 · incomplete 1 — no holds |
| phi4:latest | 2026-11-04 | 4 | Falsified at every seed |
| qwen2.5:14b | 2026-11-04 | 4 | Falsified at every seed |
| qwen2.5:3b | 2026-11-04 | 4 | Falsified at every seed |
| qwen2.5:7b | 2026-11-04 | 4 | Falsified at every seed |
| Gemma-4-E2B abliterated (BF16) | 2026-11-05 | 4 | Falsified 1 · incomplete 3 — no holds |
| Qwen3-1.7B coffee-distilled | 2026-11-05 | 4 | Falsified at every seed |
| deepseek-r1:70b | 2026-11-05 | 4 | Falsified 3 · incomplete 1 — no holds |
| Qwen2.5-Coder-1.5B ExPO (F16) | 2026-11-05 | 4 | Falsified at every seed |
| gpt-oss:120b | 2026-11-05 | 4 | Held 3 of 4 runs |
| Honeypot Qwen3-1.7B (sepetly) | 2026-11-05 | 4 | Falsified at every seed |
| KSA-1.9B-base (GGUF) | 2026-11-05 | 4 | Falsified 2 · incomplete 2 — no holds |
| LFM3-1.7B-chv-v2 (Chuvash finetune) | 2026-11-05 | 4 | Falsified at every seed |
| llama3.3:70b | 2026-11-05 | 4 | Falsified at every seed |
| qwen2.5:32b | 2026-11-05 | 4 | Falsified at every seed |
| qwen2.5:72b | 2026-11-05 | 4 | HELD EVERY RUN — undefeated |
| Qwen3-4B (hellmans GGUF, Q4_K_M) | 2026-11-05 | 4 | Never falsified — but never cleanly held (incomplete runs) |
| Llama-3.2-3B SOAP notes finetune | 2026-11-05 | 4 | Falsified at every seed |
| Text2SQL Qwen2.5-Coder-1.5B | 2026-11-05 | 4 | Held 1 of 4 runs |
| Qwen3.5-4B Trace-Inversion | 2026-11-05 | 4 | Never falsified — but never cleanly held (incomplete runs) |
Run-by-run evidence for every row of this table lives at llmdevsecops.ai.