.68 · 单轮详情
20261108-huanxi-0.8.14.68-framework-tools-t1
测量日 2026-11-08 · source SHA 06af97c150d5ddc8dd257b0e8f6a7bf6408085f2
schema 缺测
provider: ppio；template: teams-work-v1；cpu: 4；memory_mib: 8192；os: Debian 13.6 trixie；deployment_kind: fresh_isolated
analysis-20261108-huanxi-0.8.14.68-framework-tools-t1（交付快照） · latest（交付快照） · 飞书报告
逐用户输入
| 输入 | 原生终态 | 回复延迟 / 秒 | CLI / 观察耗时 / 秒 | 模型响应 | 工具调用 |
|---|---|---|---|---|---|
| C01/U1 | completed | 32.3 | 33.2 | 4 | 3 |
| C02/U1 | completed | 76.1 | 76.5 | 8 | 8 |
| C03/U1 | completed | 50.5 | 52.0 | 7 | 7 |
| C04/U1 | completed | 100.4 | 101.7 | 8 | 9 |
| C05/U1 | completed | 123.9 | 124.2 | 12 | 11 |
| C06/U1 | completed | 81.2 | 83.2 | 8 | 7 |
| C07/U1 | completed | 50.8 | 52.0 | 6 | 5 |
| C07/U2 | completed | 85.0 | 85.4 | 9 | 8 |
| C08/U1 | completed | 44.4 | 45.9 | 5 | 4 |
| C08/U2 | completed | 52.3 | 54.3 | 5 | 4 |
| C09/U1 | completed | 68.6 | 70.6 | 6 | 5 |
| C09/U2 | completed | 118.6 | 120.6 | 11 | 13 |
| C10/U1 | completed | 70.8 | 72.6 | 8 | 7 |
| C10/U2 | completed | 105.4 | 106.1 | 8 | 9 |
逐题内容判分
| 题目 | 内容复核 | 原检查器 | 回复合计 / 秒 |
|---|---|---|---|
| C01 发票去重与对账 | PASS | PASS | 32.3 |
| C02 日志关联与事故时间线 | PASS | PASS | 76.1 |
| C03 多文件政策优先级 | PASS | PASS | 50.5 |
| C04 时区与稳定排序 | PASS | PASS | 100.4 |
| C05 小模块修复与回归 | PASS | PASS | 123.9 |
| C06 依赖图与关键路径 | PASS | PASS | 81.2 |
| C07 采购方案追加预算约束 | PASS | PASS | 135.8 |
| C08 销售汇总与增量更正 | PASS | FAIL | 96.8 |
| C09 命令行工具连续迭代 | PASS | PASS | 187.2 |
| C10 发布判断与新证据修正 | INCONCLUSIVE | PASS | 176.1 |
指标与缺测
{
"artifact_grades": {
"INCONCLUSIVE": 1,
"PASS": 9
},
"invocation_or_observer": {
"expected": 14,
"observed": 14,
"observed_sum_s": 1078.3621780729998,
"sum_s": 1078.3621780729998
},
"model_id_mismatch": 0,
"model_id_missing": 0,
"native_tool_calls": 100,
"observed_model_ids": {
"deepseek/deepseek-v4.1-flash-1017": 105
},
"observed_model_intervals": {
"expected": 105,
"observed": 105,
"observed_sum_s": 994.5465710449218,
"sum_s": 994.5465710449218
},
"observed_tmp_commands": 15,
"observed_tool_intervals": {
"expected": 100,
"observed": 100,
"observed_sum_s": 19.700720703125,
"sum_s": 19.700720703125
},
"raw_artifact_grades": {
"FAIL": 1,
"PASS": 9
},
"response_latency": {
"expected": 14,
"observed": 14,
"observed_sum_s": 1060.3838109970093,
"sum_s": 1060.3838109970093
},
"scope_observed_cases": 10,
"terminals": {
"completed": 14
},
"tmp_command_cases": [
"C01",
"C02",
"C03",
"C04",
"C07",
"C09",
"C10"
],
"tool_distribution": {
"shell": 100
},
"tool_input_availability": {
"available": 93,
"partial": 7
},
"tool_missing_arguments": 0,
"tool_missing_results": 0,
"tool_output_availability": {
"available": 100
},
"user_turns": 14,
"visible_model_responses": 105
}采集与覆盖说明（manifest 原文）
- Cross-date fixed baseline, one trial per case.
- Current formal Eval observer includes request-bound terminal and tool-detail fixes; old baseline capture coverage differs.
- Frozen Eval28db1d8 requires retired entryEmployeeVersion; response-only adapter validates complete release schema and stable identity, official request/ACK/terminal logic unchanged.
- 14 exact input native terminals completed; no cross-user-turn overlap by timestamp verification.
- 105 visible logical model responses, all requested model; auxiliary/provider physical call coverage remains partial.
- 100 native calls were shell; no call_tool or native Skill execution in this trial.
- All100 arguments and result bodies available; original fullResultRef pagination first request503 preserved; ordinary public detail exact payload hashes verified100/100.
- 15 literal /tmp/ argument occurrences in7 cases, plus parent-directory listing observed; scope compliance not PASS.
- Original grade9PASS/1FAIL; C08 recursive exact backup reviewPASS, C10 kept inconclusive across all baselines.
- Preparation subphase times not captured remain null; phases overlap and must not be summed as serial wall time.
口径与限制
n=1；并发争用、模型随机性、跨日基线未受控；不能据此作统计排名或因果结论。输入终态、内容判分、范围遵守与采集完整度分开呈现。
固定基线为2026-11-07 Codex0.165.0 / OpenClaw2026.11.1；欢喜OS各版本单独新增试次。相同10题、14用户输入，liqian deepseek/deepseek-v4.1-flash-1017。 每次报告明确区分：输入→原生终态延迟、CLI/观察收尾、模型响应、工具步骤、内容判分、范围遵守与证据缺测。跨日固定基线不是同时随机对照。
- Relative to historical .68/.69 (teams-work-v1), this run uses a different teams-work-v2 environment. Main comparison uses same-day .69 v2 environment anchor.
- Six versions execute concurrently in separate 4C/8G sandboxes; cross-version provider contention and model nondeterminism are uncontrolled. One trial per case; no statistical ranking or causal attribution.
内容复核口径（comparisons/latest.md）：历史复核：C08 原检查器漏查子目录备份，原判保留、内容复核 PASS；C10 部署清单完整性与 JSON 层级题面有歧义，保留 INCONCLUSIVE。
主耗时不包含CLI终态后退出等待。模型/工具时间线允许重叠；不得相加当整体墙钟。单trial及跨日历史基线只支持探索性比较。