.81 · 单轮详情
20261112-huanxi-0.8.14.81-framework-tools-t1
测量日 2026-11-12 · source SHA 75eaaf3427057de3e5a2d08ea5aa187afccf0eee
schema 0.8.2-runtime-tables.133
provider: ppio；template: teams-work-v2；cpu: 4；memory_mib: 8192；os: Debian 13.6 trixie；deployment_kind: fresh_isolated
analysis-20261112-huanxi-0.8.14.77-81（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 | 37.7 | 670.9 | 4 | 3 |
| C02/U1 | completed | 80.8 | 81.7 | 7 | 6 |
| C03/U1 | completed | 53.4 | 55.1 | 6 | 5 |
| C04/U1 | completed | 55.7 | 57.2 | 6 | 5 |
| C05/U1 | completed | 158.2 | 159.4 | 14 | 13 |
| C06/U1 | completed | 47.2 | 49.1 | 5 | 4 |
| C07/U1 | completed | 42.5 | 43.0 | 4 | 3 |
| C07/U2 | completed | 96.6 | 98.3 | 10 | 9 |
| C08/U1 | completed | 39.9 | 41.0 | 4 | 3 |
| C08/U2 | completed | 58.8 | 59.4 | 4 | 3 |
| C09/U1 | completed | 56.6 | 57.4 | 6 | 5 |
| C09/U2 | completed | 104.0 | 104.6 | 9 | 8 |
| C10/U1 | completed | 65.2 | 65.6 | 7 | 7 |
| C10/U2 | completed | 51.8 | 53.4 | 4 | 3 |
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 | 37.7 |
| C02 日志关联与事故时间线 | PASS | PASS | 80.8 |
| C03 多文件政策优先级 | PASS | PASS | 53.4 |
| C04 时区与稳定排序 | PASS | PASS | 55.7 |
| C05 小模块修复与回归 | PASS | PASS | 158.2 |
| C06 依赖图与关键路径 | PASS | PASS | 47.2 |
| C07 采购方案追加预算约束 | PASS | PASS | 139.0 |
| C08 销售汇总与增量更正 | PASS | FAIL | 98.6 |
| C09 命令行工具连续迭代 | PASS | PASS | 160.6 |
| C10 发布判断与新证据修正 | INCONCLUSIVE | FAIL | 116.9 |
指标与缺测
{
"artifact_grades": {
"INCONCLUSIVE": 1,
"PASS": 9
},
"invocation_or_observer": {
"expected": 14,
"observed": 14,
"observed_sum_s": 1596.2426935670007,
"sum_s": 1596.2426935670007
},
"model_id_mismatch": 0,
"model_id_missing": 0,
"native_tool_calls": 77,
"observed_model_ids": {
"deepseek/deepseek-v4.1-flash-1017": 90
},
"observed_model_intervals": {
"expected": 90,
"observed": 90,
"observed_sum_s": 817.4532661132813,
"sum_s": 817.4532661132813
},
"observed_tmp_commands": 1,
"observed_tool_intervals": {
"expected": 77,
"observed": 77,
"observed_sum_s": 25.546012451171876,
"sum_s": 25.546012451171876
},
"raw_artifact_grades": {
"FAIL": 2,
"PASS": 8
},
"response_latency": {
"expected": 14,
"observed": 14,
"observed_sum_s": 948.2676181793213,
"sum_s": 948.2676181793213
},
"scope_observed_cases": 10,
"terminals": {
"completed": 14
},
"tmp_command_cases": [
"C10"
],
"tool_distribution": {
"shell": 77
},
"tool_input_availability": {
"available": 71,
"partial": 6
},
"tool_missing_arguments": 0,
"tool_missing_results": 0,
"tool_output_availability": {
"available": 77
},
"user_turns": 14,
"visible_model_responses": 90
}采集与覆盖说明（manifest 原文）
- 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.
- Frozen Eval 28db1d8f2 and .69 response-only adapter; no OS source changes.
- Eval codeload returned HTTP404; identical frozen 1128-file source snapshot transferred and verified in sandbox. Initial incomplete archive upload retained; reassembled archive SHA256 32d89e797c08be115ae19e0d3a12a1242ad6bc55b54ecd05511a203f9416cb15.
- Old Eval requires optional upgrades; .80 removes TaskFlow upgrades projection (#3724). Public schema never required upgrades.
- C01 observer restart authorized only after native final reply + idle, preserving original intent/ACK/Session/command/operation; no input replay. Native reply timestamp remains latency endpoint; observation recovery delay reported separately.
- Adapter/driver diff: raw/resume-adaptation.diff; SHA256=ac954830f2c52a044f26e21c96b3310607fcac9fb2dd72e2e646018a02485579; adapter SHA256=ff8c2500e4c31223f5e8bcba3cc386e7668a84d422f1b72fc9323652442fa196; resume driver SHA256=342434d4eab58d36d119e563733b73e4635d438adb35f1934530cc532fbac03f
- Content grades, native terminals, availability and public pagination probe recorded separately. Logical model token/call counts do not prove physical attempts or billing.
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。