{ "captured_at_utc": "2026-11-14T18:10:54.606723+00:00", "host": { "cpu": "ARM Neoverse-N1, 2 vCPU (aarch64)", "mem_total_mb": 11927, "mem_available_mb": 7821, "note": "shared 2-vCPU host also running the blog automation pipeline; CPU contention is the dominant noise source (see limitations)" }, "ollama": { "version": "0.20.3", "endpoint": "http://127.0.0.1:11434/api/generate", "stream": false, "raw": true }, "model": { "name": "qwen2.5:0.5b", "digest": "a8b0c51577010a279d933d14c2a8ab4b268079d44c5c8830c0a93900f1827c67", "quantization": "Q4_K_M per tag listing" }, "request_options": { "temperature": 0, "num_predict": 3, "seed": 42, "keep_alive": "30m" }, "fixture": { "system_block": "25 numbered policy lines (~1.9k chars)", "tool_registry": "6 terse function definitions (~1.7k chars)", "turns": "scripted 4-turn user sequence, history appended monotonically", "prompt_tokens_server_reported": "941-993" }, "response_keys": [ "created_at", "done", "done_reason", "eval_count", "eval_duration", "load_duration", "model", "prompt_eval_count", "prompt_eval_duration", "response", "total_duration" ], "prompt_eval_cached_count_present": false, "sweep_note": "variants run consecutively, not interleaved: Ollama keeps one prefix per loaded slot, so interleaving would itself destroy the reuse under test. Consequence: cross-variant comparison is exposed to host drift." }