Skip to content

Report benchmark effort, score curves, and run configurations - #7750

Merged
huangruiteng merged 2 commits into
mainfrom
codex/benchmark-campaign-effort-curves
Nov 1, 2026
Merged

huangruiteng merged 2 commits into
mainfrom
codex/benchmark-campaign-effort-curves

Conversation

@huangruiteng

@huangruiteng huangruiteng commented Oct 31, 2026 •

Copy link
Copy Markdown
Collaborator

Campaign reports could compare terminal scores and run states, but they could not carry the evidence-linked effort, score-over-time, or exact selected-run configuration needed to assess long-run efficiency and comparison quality. This adds optional duration, token, request-count, normalized score-sample, and typed source/scorer/model/budget/feedback-cadence/resource projections while preserving missing values as unknown.

The JSON output includes flattened score/time samples and grouped configuration profiles. CSV includes effort and configuration provenance. Markdown reports coverage-aware totals and one row per observed configuration, so heterogeneous arms stay visible and are never presented as strict causal pairs merely because they share scored cases.

Validation:

  • node --experimental-strip-types --test tests/control_plane_ts/benchmark_campaign_report.test.ts (20/20)
  • Real four-arm EdgeBench adapter refresh: 204 selected runs, 21 observed configuration profiles, 200 score/time samples
  • Targeted LoopX public-boundary scan: 3/3 changed files clean
  • git diff --check

Signed-off-by: huangruiteng <huangrt01@163.com>
huangruiteng added a commit that referenced this pull request Oct 31, 2026
Signed-off-by: huangruiteng <huangrt01@163.com>
@mergify

mergify Bot commented Oct 31, 2026

Copy link
Copy Markdown

Hi @huangruiteng, the DCO Sign-off check did not pass. Please inspect
its details first: checkout, fetch, timeout or infrastructure errors
need their own recovery, not a rewrite of otherwise signed commits.

If the log confirms a missing Signed-off-by trailer, amend the
affected commit with git commit --amend -s; for multiple commits,
use an interactive rebase against the current base from the correct
base-repository remote and sign off each affected commit. Push the
rewritten PR branch with git push --force-with-lease origin HEAD.

Signed-off-by: huangruiteng <huangrt01@163.com>
@huangruiteng
huangruiteng force-pushed the codex/benchmark-campaign-effort-curves branch from 3d164cb to e3d3eb1 Compare October 31, 2026 13:00
@huangruiteng huangruiteng changed the title Report benchmark effort and score curves Report benchmark effort, score curves, and run configurations Oct 31, 2026
huangruiteng added a commit that referenced this pull request Oct 31, 2026
Signed-off-by: huangruiteng <huangrt01@163.com>

@songoow songoow left a comment

Copy link
Copy Markdown
Collaborator

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Reviewer: model_agent | GPT-6 Astra Pro | OpenAI | self_reported

审阅固定在 e3d3eb17ffb15610f1db778a37cca9de37d0240e;基线 04c5799b88f6c4cf12a224b7fa679f272e1e794e。核对了最新讨论与当前 main 21367015cce45294dcf99f54a3211d66f5d842c5,触及的关键生产 owner 在 main 仍与基线一致。

本轮未发现新的具体阻断性代码缺陷;当前为验证证据 hold。 当前失败 test-shard (4)/(2)/(3) 未取得独立可归因诊断;其余pending远程任务不作为单独否决理由。

动机 / Motivation

受影响的是比较多轮 benchmark campaign 的研究人员,他们需要同时看分数、消耗和实际运行条件。 旧报告只能比较终局成绩与状态;两个 arm 即使使用不同模型、预算或反馈节奏,读者也难以从同一报告看出差异和缺失测量。 本次让被选中的 run 携带有证据引用的消耗、随时间评分和配置画像,并在报告里保留缺失覆盖率。 范围是已有离线报告器,不启动实验、不修改评分权威,也不宣称描述性分数差异证明因果收益。

未发现为本次具体增量单独批准的书面验收规范;按原问题、基线现有owner契约与已有讨论判断,不把本PR新增文档当作独立验收依据。

改动思路 / Approach

新观测仍由既有输入校验与报告生成器消费,没有把报告层变成实验执行、资格判定或评分 owner。 当前 PR 的边界是把可选测量完整送达 JSON、Markdown 与 CSV,同时保持已有共同题集比较语义。 我检查了不带新字段的旧输入,再检查只加部分 effort 的同一输入;旧终局比较没有被未知用量、配置分组或时间采样改写。这里不能用表格里的零汇总单独代表测到零:JSON 同时给 known_runs,Markdown 在无覆盖时显示未知。配置画像按实际 typed 字段分组,并显式保留 configuration_supplied=false。

具体改动 / Changes

关键代码讲解

  • validateInput(scripts/benchmark-campaign-report.ts:79)校验可选 effort 至少一个测量已知、计数为非负安全整数,score_samples 时间严格递增且分数处于有效区间,并要求对应 evidence_ref。它没有把采样得分升级为正式终局 score。
  • buildReport(:158)仍使用选定 attempt 和共同题集构造 comparisons;新增 coverage-aware effort、逐 run 的 score_time_series 与配置 profile。未知配置与明确提供的部分配置有独立 supplied 标志,证据引用不会改变配置分组。
  • renderMarkdown(:270)呈现覆盖数、小时和配置维度;renderCsv(:294)新增列并沿用现有 CSV 公式转义。CSV 原 evidence_ref 列改为 score_evidence_ref,我检查了仓内消费者,没有找到依赖原列名的第二个解析 owner;这是需要告知外部手写消费者的可观察格式变化。

验证

实际在 exact head 执行 Node 测试 20/20 passed,包含真正 CLI 写入三种报告和无效输入不产物。另执行基线与 head 的同输入对照:去掉恰好新增的 effort/configuration_profiles/score_time_series 字段后,旧 JSON 全对象相等;加入部分 effort 后 comparisons 仍严格相等、未知 request count 的 known_runs=0。该对照用不可变 base 04c5799,不是根据新实现构造预期。 本PR对其不可变base执行git diff --check,检查通过。

对主干的风险 / Risks

未发现具体阻断性代码缺陷。数据仅来自输入中的 selected run,配置真实性与原始 usage 的采集正确性仍由引用的生产者负责;这份报告不会自行证明同条件实验或完整效率。当前失败 test-shard (4)/(2)/(3) 的日志尚不可取得,无法独立判断失败是否与变更路径有关,所以保留验证 hold;其余 pending job 本身不是拒绝理由,未知失败也不能称作已证实回归。 当前报告只覆盖明确列出的执行与源码证据,未完成的provider、scope矩阵或视觉状态不会标成通过;review结论与合并时的分支保护、完整CI资格分开记录。

我的整体评价 / Verdict

功能边界清楚,局部验证与旧输入对照支持新增观测能力且没有改写原有终局比较。当前结论为证据不足的 hold;最小补充是查明当前三个失败 shard 的具体签名并用相同命令验证当前相关不变量或归因到不可变基线。无需重跑真实付费实验来证明这个离线序列化切片,也不需要为报告器增加新的 kernel owner。

本变更输出离线报告文本,没有更改网页渲染组件;已检查 Markdown 表格转义、CSV 公式防护及 CLI 实际产物。 实际消费端的 Markdown/电子表格渲染未检查;本 PR 没有浏览器布局承诺。

English verdict: REQUEST_CHANGES — exact head e3d3eb1; no new concrete code defect found; current validation gaps remain explicitly unverified.

@huangruiteng
huangruiteng merged commit 301269f into main Nov 1, 2026
33 of 41 checks passed
@huangruiteng
huangruiteng deleted the codex/benchmark-campaign-effort-curves branch November 1, 2026 02:52
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

2 participants