{ "context": { "cap_tokens": 261056, "max_model_len": 262144, "targets": [ { "attempted_context_tokens": 32768, "chars_per_token": 3.0, "clamped_tokens": 32768, "e2e_ms": 3225.151300430298, "estimated_prompt_tokens": 32768, "output_tokens": 0, "status": "passed", "target_tokens": 32768, "ttft_ms": 3225.1503467559814, "usage": { "completion_tokens": 64, "prompt_tokens": 27774, "total_tokens": 27838 } }, { "attempted_context_tokens": 128000, "chars_per_token": 3.539, "clamped_tokens": 128000, "e2e_ms": 32699.106693267822, "estimated_prompt_tokens": 127999, "output_tokens": 0, "status": "passed", "target_tokens": 128000, "ttft_ms": 32699.105262756348, "usage": { "completion_tokens": 64, "prompt_tokens": 149016, "total_tokens": 149080 } }, { "attempted_context_tokens": 240000, "chars_per_token": 3.04, "clamped_tokens": 240000, "e2e_ms": 81470.12066841125, "estimated_prompt_tokens": 239999, "output_tokens": 0, "status": "passed", "target_tokens": 240000, "ttft_ms": 81470.12042999268, "usage": { "completion_tokens": 64, "prompt_tokens": 243644, "total_tokens": 243708 } } ] }, "evaluation_protocol": { "budget_semantics": "visible_answer_tokens plus reasoning_headroom_tokens are sent as one max_tokens cap; the endpoint does not hard-partition the channels", "minimum_comparable_repetitions": 3, "minimum_pass_rate": 1.0, "reasoning_headroom_tokens": 4096, "records_finish_reason": true, "records_full_visible_answer": true, "records_reasoning_channel_metadata": true, "repetitions": 3, "version": 3, "visible_answer_tokens": 512 }, "failures": [ { "error": "wrong argument zip: '\"98101\"'", "failure_classes": [ "visible_answer_missing" ], "pass_rate": 0.3333333333333333, "suite": "tool" } ], "identity": { "base_url": "http://127.0.0.1:39038/v1", "candidate_id": "unsloth-gemma4-12b-nvfp4", "config_id": "heavy-256k", "engine": "vllm-0.25.1", "gpu": "RTX-PRO-6000-96GB", "model": "gemma4-heavy-lab", "started_at": "2026-07-16T23:21:08Z" }, "intelligence": { "checks": [ { "attempt_count": 3, "attempts": [ { "attempt": 1, "budget": { "legacy_max_tokens_as_visible": false, "max_completion_tokens": 4608, "reasoning_headroom_tokens": 4096, "visible_answer_tokens": 512 }, "content": "```diff\n--- app.py\n+++ app.py\n@@ -1,2 +1,2 @@\n-timeout = 30\n+timeout = 45\n retries = 2\n```", "content_excerpt": "```diff\n--- app.py\n+++ app.py\n@@ -1,2 +1,2 @@\n-timeout = 30\n+timeout = 45\n retries = 2\n```", "failure_class": null, "finish_reason": "stop", "latency_ms": 3148.501396179199, "reasoning_chars": 691, "reasoning_excerpt": "* Input file content:\n ```python\n timeout = 30\n retries = 2\n ```\n * Task: Change `timeout` to 45, leave `retries` unchanged.\n * Output format: Unified diff.\n\n ", "reasoning_field": "reasoning", "reasoning_tokens": null, "status": "passed", "text_checks": [ { "name": "diff_shape", "passed": true }, { "name": "removes_old_timeout", "passed": true }, { "name": "adds_new_timeout", "passed": true } ], "usage": { "completion_tokens": 312, "prompt_tokens": 57, "prompt_tokens_details": null, "total_tokens": 369 } }, { "attempt": 2, "budget": { "legacy_max_tokens_as_visible": false, "max_completion_tokens": 4608, "reasoning_headroom_tokens": 4096, "visible_answer_tokens": 512 }, "content": "```diff\n--- app.py\n+++ app.py\n@@ -1,2 +1,2 @@\n-timeout = 30\n+timeout = 45\n retries = 2\n```", "content_excerpt": "```diff\n--- app.py\n+++ app.py\n@@ -1,2 +1,2 @@\n-timeout = 30\n+timeout = 45\n retries = 2\n```", "failure_class": null, "finish_reason": "stop", "latency_ms": 3183.795690536499, "reasoning_chars": 691, "reasoning_excerpt": "* Input file content:\n ```python\n timeout = 30\n retries = 2\n ```\n * Task: Change `timeout` to 45, leave `retries` unchanged.\n * Output format: Unified diff.\n\n ", "reasoning_field": "reasoning", "reasoning_tokens": null, "status": "passed", "text_checks": [ { "name": "diff_shape", "passed": true }, { "name": "removes_old_timeout", "passed": true }, { "name": "adds_new_timeout", "passed": true } ], "usage": { "completion_tokens": 312, "prompt_tokens": 57, "prompt_tokens_details": null, "total_tokens": 369 } }, { "attempt": 3, "budget": { "legacy_max_tokens_as_visible": false, "max_completion_tokens": 4608, "reasoning_headroom_tokens": 4096, "visible_answer_tokens": 512 }, "content": "```diff\n--- app.py\n+++ app.py\n@@ -1,2 +1,2 @@\n-timeout = 30\n+timeout = 45\n retries = 2\n```", "content_excerpt": "```diff\n--- app.py\n+++ app.py\n@@ -1,2 +1,2 @@\n-timeout = 30\n+timeout = 45\n retries = 2\n```", "failure_class": null, "finish_reason": "stop", "latency_ms": 3139.131784439087, "reasoning_chars": 691, "reasoning_excerpt": "* Input file content:\n ```python\n timeout = 30\n retries = 2\n ```\n * Task: Change `timeout` to 45, leave `retries` unchanged.\n * Output format: Unified diff.\n\n ", "reasoning_field": "reasoning", "reasoning_tokens": null, "status": "passed", "text_checks": [ { "name": "diff_shape", "passed": true }, { "name": "removes_old_timeout", "passed": true }, { "name": "adds_new_timeout", "passed": true } ], "usage": { "completion_tokens": 312, "prompt_tokens": 57, "prompt_tokens_details": null, "total_tokens": 369 } } ], "id": "unified_diff_timeout_edit", "pass_count": 3, "pass_rate": 1.0, "required_pass_rate": 1.0, "status": "passed", "validator": "deterministic_text_checks" }, { "attempt_count": 3, "attempts": [ { "attempt": 1, "budget": { "legacy_max_tokens_as_visible": false, "max_completion_tokens": 4608, "reasoning_headroom_tokens": 4096, "visible_answer_tokens": 512 }, "content": "The total latency (3000 ms) exceeds the 2500 ms timeout; implement LLM streaming to the TTS to reduce the time-to-first-byte.", "content_excerpt": "The total latency (3000 ms) exceeds the 2500 ms timeout; implement LLM streaming to the TTS to reduce the time-to-first-byte.", "failure_class": null, "finish_reason": "stop", "latency_ms": 5242.51914024353, "reasoning_chars": 1583, "reasoning_excerpt": "* Components: Voice Agent $\\rightarrow$ STT $\\rightarrow$ LLM $\\rightarrow$ TTS.\n * Constraint: Total turn timeout = 2500 ms.\n * Actual Latency: STT (550 ms) + LLM (1800 ms) + TTS (650 ms)", "reasoning_field": "reasoning", "reasoning_tokens": null, "status": "passed", "text_checks": [ { "name": "identifies_budget_overrun", "passed": true }, { "name": "offers_latency_fix", "passed": true } ], "usage": { "completion_tokens": 517, "prompt_tokens": 81, "prompt_tokens_details": null, "total_tokens": 598 } }, { "attempt": 2, "budget": { "legacy_max_tokens_as_visible": false, "max_completion_tokens": 4608, "reasoning_headroom_tokens": 4096, "visible_answer_tokens": 512 }, "content": "The total latency (3000ms) exceeds the 2500ms timeout limit; optimize the LLM response time by using a faster model or implementing streaming.", "content_excerpt": "The total latency (3000ms) exceeds the 2500ms timeout limit; optimize the LLM response time by using a faster model or implementing streaming.", "failure_class": null, "finish_reason": "stop", "latency_ms": 4548.805952072144, "reasoning_chars": 1324, "reasoning_excerpt": "* Components: Voice Agent -> STT -> LLM -> TTS.\n * Total Turn Timeout: 2500 ms.\n * Actual Latency: STT (550) + LLM (1800) + TTS (650) = 3000 ms.\n * Problem: The total latency (3000 ms", "reasoning_field": "reasoning", "reasoning_tokens": null, "status": "passed", "text_checks": [ { "name": "identifies_budget_overrun", "passed": true }, { "name": "offers_latency_fix", "passed": true } ], "usage": { "completion_tokens": 449, "prompt_tokens": 81, "prompt_tokens_details": null, "total_tokens": 530 } }, { "attempt": 3, "budget": { "legacy_max_tokens_as_visible": false, "max_completion_tokens": 4608, "reasoning_headroom_tokens": 4096, "visible_answer_tokens": 512 }, "content": "The total latency (3000ms) exceeds the 2500ms timeout limit; optimize the LLM response time by using a faster model or implementing streaming.", "content_excerpt": "The total latency (3000ms) exceeds the 2500ms timeout limit; optimize the LLM response time by using a faster model or implementing streaming.", "failure_class": null, "finish_reason": "stop", "latency_ms": 4536.951303482056, "reasoning_chars": 1324, "reasoning_excerpt": "* Components: Voice Agent -> STT -> LLM -> TTS.\n * Total Turn Timeout: 2500 ms.\n * Actual Latency: STT (550) + LLM (1800) + TTS (650) = 3000 ms.\n * Problem: The total latency (3000 ms", "reasoning_field": "reasoning", "reasoning_tokens": null, "status": "passed", "text_checks": [ { "name": "identifies_budget_overrun", "passed": true }, { "name": "offers_latency_fix", "passed": true } ], "usage": { "completion_tokens": 449, "prompt_tokens": 81, "prompt_tokens_details": null, "total_tokens": 530 } } ], "id": "parallel_timeout_triage", "pass_count": 3, "pass_rate": 1.0, "required_pass_rate": 1.0, "status": "passed", "validator": "deterministic_text_checks" } ], "status": "passed" }, "run_id": "fast-bakeoff-20260716T232108Z", "schema": "anvil-serving.fast-tier-bakeoff/v1", "score_inputs": { "e2e_p50_ms": 32699.106693267822, "intelligence_pass_rate": 1.0, "operational_fit_notes": [ "endpoint was already loaded; benchmark did not start or stop serves" ], "session_recall_passed": true, "thinking_mode": "enabled", "tool_call_passed": false, "ttft_p50_ms": 32699.105262756348, "usable_context_tokens": 240000, "voice_latency_ms": null }, "selection": { "context_targets": [ 32768, 128000, 240000 ], "endpoint_already_loaded": true, "requests_per_context": 1, "suites": [ "chat", "context", "tool", "session", "intelligence" ] }, "session": { "checks": [ { "attempt_count": 3, "attempts": [ { "attempt": 1, "budget": { "legacy_max_tokens_as_visible": false, "max_completion_tokens": 4608, "reasoning_headroom_tokens": 4096, "visible_answer_tokens": 512 }, "content": "RIVER-918", "content_excerpt": "RIVER-918", "expected": "RIVER-918", "failure_class": null, "finish_reason": "stop", "latency_ms": 819.4069862365723, "reasoning_chars": 257, "reasoning_excerpt": "The user is asking for the session code that was provided in the previous turn.\nThe session code provided was \"RIVER-918\".\nThe user requested to reply with *only* the code.\n\nPlan:\n1. Identify the sess", "reasoning_field": "reasoning", "reasoning_tokens": null, "status": "passed", "usage": { "completion_tokens": 77, "prompt_tokens": 55, "prompt_tokens_details": null, "total_tokens": 132 } }, { "attempt": 2, "budget": { "legacy_max_tokens_as_visible": false, "max_completion_tokens": 4608, "reasoning_headroom_tokens": 4096, "visible_answer_tokens": 512 }, "content": "RIVER-918", "content_excerpt": "RIVER-918", "expected": "RIVER-918", "failure_class": null, "finish_reason": "stop", "latency_ms": 819.1375732421875, "reasoning_chars": 257, "reasoning_excerpt": "The user is asking for the session code that was provided in the previous turn.\nThe session code provided was \"RIVER-918\".\nThe user requested to reply with *only* the code.\n\nPlan:\n1. Identify the sess", "reasoning_field": "reasoning", "reasoning_tokens": null, "status": "passed", "usage": { "completion_tokens": 77, "prompt_tokens": 55, "prompt_tokens_details": null, "total_tokens": 132 } }, { "attempt": 3, "budget": { "legacy_max_tokens_as_visible": false, "max_completion_tokens": 4608, "reasoning_headroom_tokens": 4096, "visible_answer_tokens": 512 }, "content": "RIVER-918", "content_excerpt": "RIVER-918", "expected": "RIVER-918", "failure_class": null, "finish_reason": "stop", "latency_ms": 823.155403137207, "reasoning_chars": 257, "reasoning_excerpt": "The user is asking for the session code that was provided in the previous turn.\nThe session code provided was \"RIVER-918\".\nThe user requested to reply with *only* the code.\n\nPlan:\n1. Identify the sess", "reasoning_field": "reasoning", "reasoning_tokens": null, "status": "passed", "usage": { "completion_tokens": 77, "prompt_tokens": 55, "prompt_tokens_details": null, "total_tokens": 132 } } ], "name": "single_request_multiturn_recall", "pass_count": 3, "pass_rate": 1.0, "required_pass_rate": 1.0, "status": "passed" } ], "status": "passed" }, "source_recipe": { "ref": null, "serve_command": null }, "suites": {}, "thinking": { "chat_template_kwargs": { "enable_thinking": true }, "control_evidence": "C:\\Users\\operator\\ai-code\\anvil-serving\\docs\\findings\\2026-07-16-gemma4-nvfp4-evidence\\nvfp4-thinking-control.json", "control_evidence_sha256": "76122ddb42cbb489bc023164a2693a741215b257d949c18d8663981903c2e99b", "control_mechanism": "chat_template_kwargs", "control_requested": { "enable_thinking": true }, "control_status": "supported", "mode": "enabled", "reasoning_effort": null, "unsupported": false }, "timing": { "chat": { "e2e_p50_ms": 32699.106693267822, "e2e_p95_ms": 81470.12066841125, "output_tokens": 0, "ttft_p50_ms": 32699.105262756348, "ttft_p95_ms": 81470.12042999268 }, "wall_ms": 146391.25323295593 }, "tool": { "checks": [ { "attempt_count": 3, "attempts": [ { "arguments": { "zip": "98101" }, "attempt": 1, "budget": { "legacy_max_tokens_as_visible": false, "max_completion_tokens": 4608, "reasoning_headroom_tokens": 4096, "visible_answer_tokens": 512 }, "content": "", "content_excerpt": "", "failure_class": null, "finish_reason": "tool_calls", "latency_ms": 306.5340518951416, "reasoning_chars": 0, "reasoning_excerpt": "", "reasoning_field": null, "reasoning_tokens": null, "status": "passed", "tool_call_count": 1, "usage": { "completion_tokens": 25, "prompt_tokens": 81, "prompt_tokens_details": null, "total_tokens": 106 }, "valid_tool_call_count": 1, "validation_errors": [] }, { "arguments": null, "attempt": 2, "budget": { "legacy_max_tokens_as_visible": false, "max_completion_tokens": 4608, "reasoning_headroom_tokens": 4096, "visible_answer_tokens": 512 }, "content": "", "content_excerpt": "", "error": "wrong argument zip: '\"98101\"'", "failure_class": "visible_answer_missing", "finish_reason": "tool_calls", "latency_ms": 1197.8278160095215, "reasoning_chars": 286, "reasoning_excerpt": "The user wants to call the `record_weather_zip` tool with the argument `zip=\"98101\"`.\nI should check the tool definition for `record_weather_zip`.\nThe tool `record_weather_zip` takes a `zip` parameter", "reasoning_field": "reasoning", "reasoning_tokens": null, "status": "failed", "tool_call_count": 1, "usage": { "completion_tokens": 115, "prompt_tokens": 81, "prompt_tokens_details": null, "total_tokens": 196 }, "valid_tool_call_count": 0, "validation_errors": [ "wrong argument zip: '\"98101\"'" ] }, { "arguments": null, "attempt": 3, "budget": { "legacy_max_tokens_as_visible": false, "max_completion_tokens": 4608, "reasoning_headroom_tokens": 4096, "visible_answer_tokens": 512 }, "content": "", "content_excerpt": "", "error": "wrong argument zip: '\"98101\"'", "failure_class": "visible_answer_missing", "finish_reason": "tool_calls", "latency_ms": 1192.7664279937744, "reasoning_chars": 286, "reasoning_excerpt": "The user wants to call the `record_weather_zip` tool with the argument `zip=\"98101\"`.\nI should check the tool definition for `record_weather_zip`.\nThe tool `record_weather_zip` takes a `zip` parameter", "reasoning_field": "reasoning", "reasoning_tokens": null, "status": "failed", "tool_call_count": 1, "usage": { "completion_tokens": 115, "prompt_tokens": 81, "prompt_tokens_details": null, "total_tokens": 196 }, "valid_tool_call_count": 0, "validation_errors": [ "wrong argument zip: '\"98101\"'" ] } ], "error": "wrong argument zip: '\"98101\"'", "expected_arguments": { "zip": "98101" }, "expected_function": "record_weather_zip", "name": "openai_tool_call_smoke", "pass_count": 1, "pass_rate": 0.3333333333333333, "required_pass_rate": 1.0, "status": "failed" } ], "status": "failed" }, "voice": { "llm_latency_ms": null, "status": "not_run", "stt_latency_ms": null, "total_turn_latency_ms": null, "tts_latency_ms": null } }