[loopx-agent]

Summary

Research synthesis can recommend a plausible ablation without establishing whether its changed limit actually engaged, or whether equal seeds and time budgets produced comparable work. This adds versioned synthesis guidance to reconcile completed comparisons, paired regressions and follow-up decisions, then make predictions conditional on the operational facts needed to interpret the test.

New standard producer requests persist revision 2. Frozen configurations without that revision retain their exact v1 prompt, key, schema, language caches and reports. The six-metric rubric, citation qualification and result schema stay unchanged; the observer remains an uncalibrated interpretation and does not execute proposed experiments.

Goal / Problem Context

• Requested outcome: useful, source-grounded explanations and discriminating next experiments in the existing research workbench.
• Current gap: configured ceilings, sampled proposals, completed work, accepted moves and final improvements can be conflated. A repeat can also be useful when contemporary controls replace historical measurements.
• Observable result: the existing controls/prediction fields distinguish exposure, actual stopping and work/randomness coupling, while allowing intended state divergence. New requests have a separate identity; saved analyses remain readable.
• Issue/task and intended base: accepted researcher-analysis refinement; main. No new solver, scorer, authorization or configuration owner.

Author Declaration

• Written by: model_agent — OpenAI GPT-6 family through Codex.

Implemented against

• Specification and revision: docs/reference/trajectory-assessment.md at base 8a718ca5f17b8036a57c6d4e88a02185eaf4581d, researcher synthesis and evidence-selection boundaries; this PR defines the bounded guidance refinement.

Criterion	Disposition	Symbol / path	Validation
Source-bound synthesis and narrow metric citations	implemented	prepareTrajectoryAssessment / qualifyTrajectoryAssessment	Both revisions exercise the same out-of-scope/missing-citation negatives
Saved results are not relabelled or automatically rerun	implemented	Persisted research_synthesis_revision; existing store/process	Frozen public identity oracles, separate real fixture jobs, existing report/citation readback
Useful explanations with counterevidence and interpretable next tests	implemented as guidance	Existing synthesis instruction and report fields	Real selected-case calibration disclosed below; scientific sufficiency is not machine-enforced
Preserve execution, failure, usage and language boundaries	implemented	Existing producer, store and HTTP paths	Process/store/HTTP regression suite; no new execution endpoint or extra model call

• Self-check: inspected the full changed production behavior, default change, protocol identities, citation scope and public boundary. Numeric revision extends the existing assessment method/config contract; it is not a new scoring or authority vocabulary. The semantic advisory has no supported carrier findings and does not analyze all dynamic forms.

Scope And Continuation

• Complete author slice: revised guidance is carried by the existing CLI preparation and pack-registration producers, qualified by the same typed owner, and read through the existing workbench.
• Existing v1 compatibility serves persisted reports and their actual language-specific jobs. It is not an unused external compatibility wrapper.
• New default: revision 2 applies to new standard requests only. Explicit custom configurations omitting the revision remain v1. No new automatic generation, retries or experiment execution.
• Broader evaluator calibration, independent reader benefit, full cost and installed adoption remain open. Related layout proposals remain separate.
• Bounded refactoring: unnecessary beyond the local version selector; the typed owner already owns interpretation and identity, while Python transports the selected producer revision. No new module, parser or generic framework.

Validation

• Tested revision: 4f10d882c07a56da582842451555a98d7d770774.
• Run state: finished for source checks and the bounded selected-case observer comparison.
• Input classes: synthetic; authorized_private_read_only.

Check kind	Result	Evidence / limitation
unit	passed	All five trajectory TypeScript suites: 65 tests, including old frozen identities, unsupported revisions, feature-off parity and unchanged citation/schema boundaries
integration	passed	Process, pack-store and HTTP research API: 62 tests; real files, disposable child and actual local HTTP; failure/cancel/retry, source tampering, owner isolation and old/new job separation
static	passed	Full control-plane typecheck; packaged dashboard build and source verification; staged bilingual check
regression_parity	passed	Four existing evidence packs across original/Chinese/English: all 12 request keys, prompt bytes and schemas unchanged; four saved Chinese reports available through the real HTTP reader; 32 old job files unchanged at verification
real_entrypoint	passed	Existing registered-pack workbench reads the prior report and original native measurements without generation; source snapshot unchanged and HTTP mutations zero
real_entrypoint	passed	One real observer invocation over unchanged authorized evidence and the same model configuration, apart from the new guidance revision; strict output/source qualification succeeded. This is selected-case author calibration, not an independent evaluation
manual	blocked	Exact-head standard premerge: 6 direct checks and 15 selected checks passed with zero failures; one benchmark-sensitive manual review hold remains. Overall manual_review_required, not a merge pass

• Initial focused test used the wrong citation response field; corrected to the existing text API and the complete relevant suite passed. A scan invocation placed the global format flag after its subcommand; the corrected bound invocation scanned all six candidate files successfully. Its credential-word warning points to an unchanged authentication-boundary comment, not credential data.
• The six-metric rubric and output schema are unchanged. Added instructions are model behavior, not a deterministic proof of better reasoning. No full repository suite, independent science/comprehension study, shared install or solver/evaluator run is claimed.

The selected-case report now identifies an unexposed deadline control, specifies actual phase entry/exit and completed work, and separates intended state divergence from failed random/work coupling. Residual gaps remain: its paired tradeoff narrative is incomplete, and a two-arm necessity test does not by itself identify the benefit of an omitted baseline. One metric label also changed despite the unchanged rubric and needs calibration; no count-based accuracy claim is made.

Observed input-cache coverage differed between the old and new calls, so elapsed time and token differences do not establish an efficiency gain. Usage and failures remain separate in their private evidence owner.

Frontend / Visual Evidence

• UI impact: none. No markup, layout, style, interaction or frontend bundle source changed.
• Before / after / first viewport: N/A for presentation changes. The existing packaged reader is used for report, old-version and literal-citation readback; private material and screenshots are not published.
• Attention/craft review: existing metrics → report → history journey and on-demand explanations are retained; no new controls, labels or configuration screen.

Type of Change

• New feature
• Documentation update
• Test update

LoopX Area

• Capability or extension (providers, adapters, skills)
• Benchmark boundary (evidence interpretation only)

Technical Direction

• S5/S11 research episode assessment: researcher value and evidence-qualified interpretation. This proposal does not qualify a treatment effect.

Shared-authority RFC fixture impact

N/A: no authority-store, shared state, permission, scheduler or scoring changes.

This is a normal author proposal for independent review. No self-merge or auto-merge is requested or enabled.

[songoow]

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

动机

研究者在已有工作台阅读研究汇报、选择下一项实验时，会遇到“配置上限”被当成“实际工作量”、重复比较缺少上下文的问题。

例如两次运行都在旧期限前停止，放宽期限本身不能说明增加了有效探索；新指引要求先核对干预是否实际生效、完成了什么工作，再提出有条件的预测。

已验证的新行为是标准准备与注册入口生成独立的第2版请求，并把修订指引交给原有分析调用；旧报告和关闭 synthesis 的输入保留原身份。 本次不修改求解器、原生评分、任务预算或实验执行权限，也不把引用校验通过当成科学有效性证明。

真实模型是否稳定采纳这些要求、是否改善研究判断及完整成本，仍需独立校准；本轮没有复用作者私有案例作为独立效果证据。

改动思路

版本选择留在已有 TypeScript assessment owner，Python 只声明新生产请求的版本；冻结配置是历史方法选择，不能根据当前默认值反推或覆盖。 本次边界是已有研究汇报的一次可回退指引升级；不新增调度器、结果数据库、执行端点或界面配置步骤。

我沿实际入口核对了 CLI trajectory-assess、trajectory-pack-register、持久化 store、TypeScript 准备／资格校验、HTTP research reader 和现有报告组件。新请求先由标准 producer 写入版本2，再由同一个 typed owner 生成 prompt、method 和 hash；只有原有的显式执行动作才启动 provider。读取旧 pack 时使用原有 record.config，而不是当前 DEFAULT_CONFIG，因此新版默认值不会悄悄升级历史报告。

输入	本次行为
新标准 synthesis 请求	保存 revision2，产生独立 v2 key
冻结／自定义配置缺少 revision	原 v1 prompt、schema、key 和语言缓存
synthesis 关闭且不带 revision	原 legacy 请求
revision2 但 synthesis 关闭，或非法版本	准备阶段拒绝

这是已有 owner 内的配置和指引变化；保留旧字符串服务的是实际历史任务，另开第二个分析器或存储没有必要。当前主干新增的 citation 内容呈现仍消费同一引用协议，本 PR 没有扩大它的数据来源。

具体改动

评审头：4f10d882c07a56da582842451555a98d7d770774。实际单提交父基准是 8a718ca5f17b8036a57c6d4e88a02185eaf4581d；GitHub 当前 base branch 前进并不应把后来 Lark 页面差异误算成本 PR。全量差异为6文件，+112/-11：生产代码+22/-6、测试+59/-2、文档+31/-3。

关键代码讲解

1. SYNTHESIS_DECISION_INSTRUCTION 复用完整 v1 文本后追加实际干预暴露、已完成对照、配对回退、固定工作量与墙钟期限的区别。原“证据不明则未知”“建议不是执行授权”仍在原文中；新增内容没有把模型判断升级为评分或任务命令。
2. prepareTrajectoryAssessment 只接受已启用 synthesis 的数字2，省略版本继续走1；method、config、instruction 一并进入 analysis key。同一套 output schema 与 source qualifier 继续使用。v2不能套用v1 key，continuation看到版本变化也不能直接复用旧判断。
3. Python prepare 通过现有 CLI 调用路径明确发送新版本，关闭 synthesis 仍发送原请求。我的真实 CLI 配对执行证实 base 输出 v1、head 输出 v2，两边都保持 preview 和显式 --execute 要求。
4. DEFAULT_CONFIG、register_pack 和 _prepared 把新默认限制在新注册入口；旧记录的重读和重试仍取冻结 config。实际文件／子进程／HTTP 测试覆盖旧任务保存、新任务分离、失败与取消恢复、引用原文、来源损坏和 owner 隔离。

规范基准为 docs/reference/trajectory-assessment.md，版本 8a718ca5f17b8036a57c6d4e88a02185eaf4581d，按原有约定逐项核对：source-bound synthesis 已实现，两版本继续拒绝越过 episode 的引用；frozen configurations 已实现，九组旧请求完整输出逐项相等；language-specific cache 已实现，未指定语言／中文／英文的旧身份分别保留；scientific qualification 保持原边界，机械校验和新指引都不证明科学有效性。

对主干的风险

最重要的回归场景是修改默认提示后旧分析 key 失效，或把旧判断当作新方法的结果。本次通过显式版本和冻结 config 避免了它。我独立执行了五组轨迹 TypeScript 测试，共 65 passed、0 skipped；三个 Python process/store/HTTP 模块共 62 passed、0 skipped。额外以相同合成输入在 immutable base/head 执行了九组省略／关闭／v1 × 未指定语言／中文／英文配对，完整 prompt、schema、config、key 均一致；v2新 key、伪造 key、非法版本及跨 episode 引用也做了反例验证。真实 CLI 的新增版本要求在旧基准失败、当前 head 通过，零付费模型调用。

语义与 CI 对齐

版本字段扩展现有 assessment 方法配置，没有新增泛化调度或授权词表。新默认明确披露在文档；模型指引和机器强制的配置／引用约束也有清楚区别。不能把配对测试通过说成模型已经更会研究：实际采纳、解释准确性和完整费用本轮未测。

远程 CI 仍有失败，应与本次代码结论分开。kernel-static 的5项 E701/E702 位于未修改的 tests/presentation/test_codex_attached_conversation.py；我用同一完整 Ruff 命令在 base/head 均复现，诊断逐字节一致。build 的 goal-activity 断言也已独立配对：使用同一锁定依赖、Node 24.22.1 和 Chromium 153，在 immutable base 与本次 exact head 执行原有 LOOPX_PERSONAL_WORKSPACE_SCENARIO=goal-activity node examples/personal-workspace-browser-smoke.mjs，两者均在 goal-activity.mjs:437 失败，actual 均为“正在执行\n0”，expected 均为“没有可确认的执行 · 2 个 Goal 执行方式未知”。断言、实际文本和期待值与远端一致；相关 frontend、fixture、workflow 和依赖未变，因此这条红项可归为 pre_existing_unrelated。没有更改断言，也没有把它称为已修复或整个 build 通过。stage2c-correctness 的日志表现为上游 skipped 后的汇总失败。相关全仓 CI／合并门仍然保留，以上65+62及独立探针只证明明确列出的变更边界。

我的整体评价

源码评审结论为 APPROVE，未发现可支持的阻塞性代码问题。这是一个 justified_increment：复用现有 owner，真实入口交付版本化指引，生产增量与问题匹配。长期持续工作能力判定为 preserved：历史结果、失败和显式恢复保留；用户旅程判定为 preserved：既有报告／来源入口和授权方式保持，不增加必要确认轮次。它没有因此完成独立研究质量或长期收益校准。

检查了 goal-research-analysis.tsx、goal-research-synthesis.tsx 及本次六文件差异，没有渲染结构、样式、交互或新增控件改动。 未重新验收真实模型生成长文的可读性；这不影响本次无布局变更的源级审查。

保留 v1 兼容有实际持久化读者，不建议为了少一个分支就删除。后续评估应关注模型是否真实识别未暴露干预与有效工作量，并保留阴性结果和完整成本。此结论仅针对 4f10d882c07a56da582842451555a98d7d770774，不是 merge-ready、CI 全绿或合并授权；若指引、版本选择、持久化／调用链改变，需要重新审查。

English verdict: APPROVE — 4f10d882c07a56da582842451555a98d7d770774. No supported implementation blocker was found. Independently verified 65 TypeScript tests, 62 real file/process/HTTP regression tests, nine immutable base/head frozen-request comparisons, and the real CLI v1→v2 default transition. Existing source/citation and explicit-execution boundaries are preserved. Model utility remains uncalibrated; remote CI failures and merge readiness remain separate and are not reported as passing.