-
Notifications
You must be signed in to change notification settings - Fork 39
docs(research): add harness strength protocol #306
New issue
Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.
By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.
Already on GitHub? Sign in to your account
Open
seekskyworld
wants to merge
1
commit into
openpi-dev:main
Choose a base branch
from
seekskyworld:docs/issue-45-harness-strength-protocol
base: main
Could not load branches
Branch not found: {{ refName }}
Loading
Could not load tags
Nothing to show
Loading
Are you sure you want to change the base?
Some commits from the old base branch may be removed from the timeline,
and old review comments may become outdated.
Open
Changes from all commits
Commits
File filter
Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
There are no files selected for viewing
83 changes: 83 additions & 0 deletions
83
docs/research/OPENPI_HARNESS_STRENGTH_PROTOCOL_2026-08-30.md
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
| Original file line number | Diff line number | Diff line change |
|---|---|---|
| @@ -0,0 +1,83 @@ | ||
| # OpenPI Harness 强度与模型能力研究协议 | ||
|
|
||
| > 状态:validated(研究协议;不代表新的 Benchmark 结果) | ||
| > | ||
| > 创建日期:2026-08-30 | ||
| > | ||
| > 最后核验:2026-08-30 | ||
| > | ||
| > 关联 Issue:[#45](https://github.com/openpi-dev/openpi/issues/45) | ||
|
|
||
| ## 目的 | ||
|
|
||
| 验证模型能力变化时,Harness 脚手架强度的收益是否变化,并区分 OpenPI 的稳定静态成本、动态执行轨迹和实际能力收益。本协议只定义可复现实验和解释边界,不要求 OpenPI 复制 OMP,也不改变默认运行时行为。 | ||
|
|
||
| ## 已核验事实 | ||
|
|
||
| - Bare Pi 的默认编码面主要是 `read`、`bash`、`edit`、`write`;OpenPI 的能力组和 Workflow/Subagent/Background 生命周期沿 Pi extension seam 按需加载。 | ||
| - OMP 同时改变 system prompt、工具面、执行反馈、上下文治理和恢复机制,因此它不是“多几个工具”的单变量对照。 | ||
| - OpenPI 既有 ARM64 source-build-derived 诊断已观察到首请求工具面税、动态轨迹差异和显著样本方差;这些材料不是官方 Terminal-Bench 分数。 | ||
| - 没有一次独立运行可以证明某个工具、提示或循环机制造成总体质量差异。 | ||
|
|
||
| 来源固定为: | ||
|
|
||
| - OpenPI:执行时记录完整 commit SHA、package version、Pi dependency version 和 active extension source。 | ||
| - Pi:记录使用的 release/tag 或完整 commit SHA。 | ||
| - OMP:记录完整 commit SHA 和公开版本。 | ||
| - 任务与 verifier:记录任务集版本、每题输入 hash、verifier 版本及其 SHA-256。 | ||
|
|
||
| ## 事实、推断与未知 | ||
|
|
||
| ### 事实 | ||
|
|
||
| 每个 cell 必须从 trace 中直接计数:verifier outcome、protocol outcome、请求数、input/output/cache/reasoning tokens、wall time、tool calls、实际加载的 capability、compaction/recovery/loop-guard 触发次数和基础设施失败。 | ||
|
|
||
| ### 推断 | ||
|
|
||
| 只有在固定模型、任务、verifier、权限、隔离和调度后,且 paired trace 显示机制触发与结果变化同时出现时,才可提出“静态脚手架成本”或“动态恢复收益”的相关性解释。相关性不得写成单机制因果结论。 | ||
|
|
||
| ### 未知 | ||
|
|
||
| - provider 的 implicit prompt cache 失效不一定等于 OpenPI 行为导致的 miss; | ||
| - 独立采样下的早期分叉可能放大后续 token 和 wall time; | ||
| - 不同 provider 的 usage 字段不能直接视为同一账本; | ||
| - 没有触发的机制不能用于解释该 cell 的结果。 | ||
|
|
||
| ## 对照设计 | ||
|
|
||
| 至少冻结四个 arm: | ||
|
|
||
| 1. Bare Pi:Pi 原生最小编码面。 | ||
| 2. OpenPI Explicit:普通 turn 保持 Pi-native,明确请求时才加载能力。 | ||
| 3. OpenPI capability profile:只加载实验指定的一个能力组。 | ||
| 4. OMP:固定版本的完整公开 harness。 | ||
|
|
||
| 每个模型 × 任务 cell 使用 AB/BA 顺序反转;每个 arm 使用独立 workspace、HOME、TMP、Session 和 provider credential 投影。任务完成后才运行 hidden verifier,禁止模型读取 verifier 或其他 arm 的目录。 | ||
|
|
||
| 模型至少覆盖一个较弱执行模型、一个中等模型和一个强 coding model。模型、provider、thinking/reasoning、timeout、最大请求数和权限配置必须在运行前冻结,不能看到结果后调整。 | ||
|
|
||
| ## 主要指标 | ||
|
|
||
| - 质量:verifier pass、paired outcome、错误类别和证据完整度。 | ||
| - 成本:首请求 input、total input/output/cache/reasoning tokens、估算成本。 | ||
| - 轨迹:turns、tool calls、重复失败、无新证据回合、wall time 和长尾。 | ||
| - 机制:能力加载成功率、实际采用率、每种机制触发次数及触发前后结果。 | ||
| - 安全:workspace 外访问、网络污染、verifier 泄漏、进程清理和 artifact 完整性。 | ||
|
|
||
| ## 解释门槛 | ||
|
|
||
| - 普通编码赛道中,OpenPI 首请求 input、总 tokens、turns、tool calls 和 wall time 只有在预注册阈值内,才可称为“没有明显静态/动态回退”。 | ||
| - 编排价值赛道必须预先声明任务确实需要委派、工作流或后台执行;只有能力被实际采用且主要质量、时间或人工步骤指标出现净改善,才可支持迁移建议。 | ||
| - 质量非劣、成本可接受和机制证据必须同时满足,才能提出默认行为变更。 | ||
| - 单次运行、基础设施失败、样本过小或未触发机制只能作为限制或待验证假设。 | ||
|
|
||
| ## 非目标 | ||
|
|
||
| - 不把 OMP 的完整脚手架复制到 OpenPI。 | ||
| - 不按模型名称硬编码“弱模型模式”。 | ||
| - 不把 Issue、研究协议或静态源码比较当作产品 Decision。 | ||
| - 不发布缺少 frozen identities、usage accounting、failure classification 或 raw-evidence reference 的正式 Benchmark 结果。 | ||
|
|
||
| ## 当前结论 | ||
|
|
||
| Issue #45 的现有材料支持建立上述配对协议,但不足以证明 Harness 强度与模型能力之间已经存在可迁移曲线。下一步必须先完成隔离有效、身份冻结、可审计 trace 的小规模重复;在此之前 OpenPI 默认 runtime 保持不变。 | ||
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
| Original file line number | Diff line number | Diff line change |
|---|---|---|
|
|
@@ -8,5 +8,6 @@ The following records predate [`Decision 0001`](../decisions/0001-documentation- | |
|
|
||
| - [`CLAUDE_CODE_WORKFLOW_FANOUT_POLICY_2026-08-23.md`](CLAUDE_CODE_WORKFLOW_FANOUT_POLICY_2026-08-23.md) — official-source research on dynamic fan-out and bounded execution. | ||
| - [`CLAUDE_CODE_WORKFLOW_RUNTIME_CONTRACT_2026-08-23.md`](CLAUDE_CODE_WORKFLOW_RUNTIME_CONTRACT_2026-08-23.md) — version-scoped Workflow contract interview and evidence boundary. | ||
| - [`OPENPI_HARNESS_STRENGTH_PROTOCOL_2026-08-30.md`](OPENPI_HARNESS_STRENGTH_PROTOCOL_2026-08-30.md) — Issue #45 的模型能力 × Harness 强度配对研究协议;不包含新的 Benchmark 结果。 | ||
|
Collaborator
There was a problem hiding this comment. Choose a reason for hiding this commentThe reason will be displayed to describe this comment to others. Learn more. 本列表声明文档早于 Decision 0001、尚未迁移到新 metadata 契约。但新增记录创建日期为 2026-08-30,Decision 0001 已于 2026-08-29 生效。这个位置会错误表达其历史豁免身份。 请移到当前研究小节,按现有 docs/README.md 约定补适用来源边界、相关 PR 和替代关系;无需新增格式框架或批量迁移旧文档。 |
||
|
|
||
| When research changes a project constraint, preserve the adopted choice in a Decision. Amend or supersede a historical record rather than silently rewriting its original conclusion. | ||
Oops, something went wrong.
Add this suggestion to a batch that can be applied as a single commit.
This suggestion is invalid because no changes were made to the code.
Suggestions cannot be applied while the pull request is closed.
Suggestions cannot be applied while viewing a subset of changes.
Only one suggestion per line can be applied in a batch.
Add this suggestion to a batch that can be applied as a single commit.
Applying suggestions on deleted lines is not supported.
You must change the existing code in this line in order to create a valid suggestion.
Outdated suggestions cannot be applied.
This suggestion has been applied or marked resolved.
Suggestions cannot be applied from pending reviews.
Suggestions cannot be applied on multi-line comments.
Suggestions cannot be applied while the pull request is queued to merge.
Suggestion cannot be applied right now. Please check back later.
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
Issue #45 §7 明确列出 Bare Pi、OpenPI Explicit、OMP、Maka 四个 Harness。这里将 Maka 换成 OpenPI capability profile,意味着完全按这份协议执行也无法完成原定的四 Harness 比较,而记录没有说明这一调整。
请恢复原四臂;单能力 profile 可另标为可选消融。若确实要改变主实验,明确给出新建议和理由,不要把它表现为已核验的原协议。四臂顺序也请沿用来源的 Latin-square/反转安排,或说明 AB/BA 实际配对对象,无需新增调度机制。