One fixed, fully sanitized ~240k-token agentic conversation that makes DeepSeek-V4-Flash-0731 collapse into repeating a short action-announcement ("Run.") instead of emitting a tool call. Reproduced 2026-08-14 on three serving stacks that share no code or hardware:
| Serving stack | Trials | Strict loops | Notes |
|---|---|---|---|
| Ollama cloud, deepseek-v4-flash:0731-cloud | 6 | 6/6 | 61-97x "Run.", all rolls hit the 2,048-token cap |
| Alibaba DashScope intl, deepseek-v4-flash | 6 | 2/6 | remaining 4 rolls emitted 6-19x "Run." before the tool call |
| Community SM80 vLLM (haosdent fork), 8x A100 | 6 | 4/6 | where we first hit it in production |
Strict loop metric (implemented in the script): a normalized sentence-segment occurring >=20 times in one completion, or >=24 consecutive segments of <=5 words. Sampling: temperature 1.0, top_p 0.95, no penalties, tool_choice=auto.
python3 cross_check.py <openai_compatible_base_url> <api_key> <model> <tag> 6
Run from the directory containing dsv4-repro-payload-sanitized.json.
Distilled from a production incident (Petabit Scale, 2026-08-12/13). The conversation is real (394 messages, 210 tool results, 51 declared tools) with every identifier neutralized; sanitization did not remove the trigger. Notably, synthetic conversations matching its shape (depth, tool count, structure, mid-chain ending, repeated-history priming) produced 0/22 loops - the attractor appears to require realistic content statistics. Related: ggml-org/llama.cpp#26694 (Flash-0731 GGUF/Metal), QwenLM/qwen-code#4695 (V4-Pro on DashScope, incl. a dose-response showing in-context self-reinforcement). Reported to DeepSeek (service@deepseek.com) with full receipts on 2026-08-14. Contact: ras@petabitscale.com