Skip to content

Show verification evidence in the Goal research workbench - #9216

Open
loopx-agent wants to merge 2 commits into
mainfrom
codex/workbench-verifier-evidence
Open

loopx-agent wants to merge 2 commits into
mainfrom
codex/workbench-verifier-evidence

Conversation

@loopx-agent

@loopx-agent loopx-agent commented Nov 21, 2026 •

Copy link
Copy Markdown
Collaborator

Goal and delivered outcome

The Goal Research workbench currently exposes individual findings and execution observations. A reader auditing a candidate has to assemble the metric, exact baseline and artifact versions, negative evidence, independent check and measured cost across views. This change adds one expandable evidence packet to Research and a direct link from the existing Goal delivery evidence.

The packet keeps the local yardstick, independent acceptance and comparison coverage separate. It compares only exact, explicitly cited baseline/candidate pairs under matching measurement profiles. The per-run costs retain measured zero and display Unknown for omissions. It shows a host tier only when a current owner-pinned criterion names both the same candidate and the matching versioned proof outputs. A host identifier alone provides no tier.

Base branch: main.

Author declaration

Criterion Disposition Implementation and evidence
Candidate and evidence binding implemented Optional evidence within the existing Explore execution observation; same source and exact SHA-256 checked by the projection tests.
Independent verifier and lifecycle guards implemented Existing Goal acceptance and artifact-lifecycle reads; exact oracle outputs and criterion ID required for tier, exercised with the isolated SQLite authority.
Readable evidence on the existing Goal Research route implemented One collapsed entry plus a direct Goal link; packaged workbench visual and keyboard check below.
Validate deployed telemetry producers and production host behavior out_of_scope Existing emitters may omit this optional data. The fixtures exercise host-neutral consumption and do not certify installed host instrumentation.

Scope and continuation

The packet reads existing Explore, Goal acceptance and artifact lifecycle data. The local acceptance GET checks receipt freshness against source files. Browsing launches neither a verifier nor a model and leaves the acceptance receipt unchanged. The comparison reports ordering on matched samples; causal gain is unavailable unless supported by separate evidence. Requests for native telemetry production, public releases or deployment remain outside this read surface.

Validation

  • Tested revision: 330fd690132616e62046a05dc184a25522b8099f.
  • Input classes: synthetic, public fixture.
Check kind Result Evidence
static passed Dashboard npm run build includes TypeScript checks and both production bundles.
static passed Control-plane TypeScript and standard loopx canary premerge --from-git-diff --git-diff-base origin/main --tier standard, 19 selected checks.
unit passed Explore history, Yardstick and verification workbench suites, 31 tests. Covers conflicting or uncited measurement scopes, mismatched criterion and digest, legacy missing fields and measured zero.
real_backend passed Disposable SQLite authority with the actual acceptance and Explore HTTP routes. Four synthetic host observations, a current isolated oracle, and 16 persisted Explore records.
real_entrypoint passed Packaged desktop build in Chromium via Ego Browser, 1200 and 390 CSS pixels. Goal overview link, disclosure keyboard operation and host switch. Codex L2, DSH L1, Pi L0, and generic CLI Unknown reflected the fixture evidence. The backend authority revision and oracle receipt were unchanged after browsing.
manual passed Populated first viewport with history and the collapsed packet; expanded reading and zero/unknown states. At 390 px the disclosure is 44 px high and all four sections fit without horizontal overflow.

Coverage covers the read path with synthetic evidence and a real local backend. Hosted CI and independent exact-head review are separate. First-pass premerge needed the repository-root npm development dependencies; the final run used those dependencies without changing tracked manifests.

Digest consumer inventory follow-up

  • Current PR head fdcf9996743404a224b835c3b2730555a89de24d. The two genuine readers of the shared SHA-256 export are now in the existing consumer test.
  • The exact previous head reproduced the missing-inventory failure. The corrected test and verification workbench passed 20 checks on the candidate and 20 on its clean combination with a6ffb3be6e7bdb48e02311ec58d618200fe0501c.
  • Core TypeScript and standard premerge passed (19 selected checks). The full Core suite completed at concurrency two with 6,712 passed, 91 skipped and zero failures. The file-lock suite also passed 15 checks each on the candidate and the current base. The previous parallel Windows timing failure did not recur in these runs.
  • The source revision of the packaged UI captures below is 330fd690132616e62046a05dc184a25522b8099f. The follow-up diff consists of the two test inventory entries.

Frontend / visual evidence

The base capture shows the unchanged Research entry on the empty version of the synthetic Goal. The packaged candidate demonstrates the same Goal with 16 synthetic Explore records. The packet adds a single disclosure before the existing timeline and keeps the verifier status, coverage and refutations in the same reading flow. The synthetic backend has no connected execution host; the banner reports that fixture state.

Desktop and 390 px first viewports, with the 390 px expanded packet, are attached below. The baseline is 55bae0ac55c8858a09f05fdebd3a1b7b5dcc131f; the packaged candidate is the tested revision above.

before-desktop before-390 publish-desktop publish-390 publish-expanded-390

Boundary checklist

  • Public-safe fixtures and captures only.
  • Scope is one read surface and its existing typed inputs.
  • No paired RFC was edited; the bilingual gate passes.
  • DCO sign-off present.

@loopx-agent loopx-agent left a comment

Copy link
Copy Markdown
Collaborator Author

Choose a reason for hiding this comment

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

Reviewer: model_agent · gpt-6.2-sol · OpenAI · runtime_reported · xhigh
Request changes conclusion (author-owned PR; GitHub blocks formal self-review)

动机 / Motivation

Goal 操作者在 Research 中依据精确版本核查候选,判断证据是否仍有效。 既有 Goal Overview 依据当前检查显示通过,历史回执留在 Delivery & evidence。 打包 Chat 展开同一个已有合格存档的候选,把实时 Goal acceptance GET 切为 503。1200 和 390 CSS px 均同时显示「当前权威回执暂不可读」「产物检查通过」「权威终评已核对该产物」和「已验证档位 L2」。这时实时证据不可读,存档中的旧通过仍被当成当前通过。 提案在 Research 增设折叠的证据包,汇集候选、冻结基线、终评、反证与成本,同一 Goal 的入口可跳转。 独立检查覆盖准确头的本地打包 Chat 与隔离 SQLite Explore,核验 GET 的快照和 503 由合成测试夹具提供。已安装 App 与真实宿主发出证据的能力未复验。

操作者打开已有检查回执的研究候选。当当前接受服务不可读时,仍看到可信的候选与引用,同时需要清楚知道本次未确认当前状态。现行准确头把原存档的通过和实时故障并列。

改动思路 / Approach

原 Explore 与 Goal acceptance 分别提供事实,React 负责选择版本与显示,档位由领域中立的精确证明判断计算。 当前断言依赖当前读取。已保存的证据引用及日期可作为历史材料展示。 同一 GET 已经提供当前权威核验,可在现有 UI 分支收紧推导,无需新增 API、持久状态或核验动作。

具体改动 / Changes

逐项检查当前头 fdcf999 和 merge base 55bae0a,17 个文件,+555/-14。检查整个准确差量,并对 b4e1ded 得到无冲突整合预演。规范中的候选档位、当前终评和保存证据分别沿既有所有者核对。 评审依据为 55bae0a。 目标规范

  • versioned-proof · Require exact candidate, criterion and proof versions for the host tier.
  • current-authority · Display a current terminal claim only when the current owner evidence is readable.
  • read-only-recovery · Use the existing read path, surface unavailable state and provide retry without mutation.

关键代码讲解

对主干的风险 / Risks

[P2] 实时核验不可读时把已保存的通过显示成当前通过。 打包 Chat 展开同一个已有合格存档的候选,把实时 Goal acceptance GET 切为 503。1200 和 390 CSS px 均同时显示「当前权威回执暂不可读」「产物检查通过」「权威终评已核对该产物」和「已验证档位 L2」。这时实时证据不可读,存档中的旧通过仍被当成当前通过。 实时模式只用同 Goal 生命周期的最新成功响应导出当前通过、精确证明和档位;等待或失败时将当前结论标为未知,并保留重试。只读历史可继续展示在存档时有效的原始证据,明确其历史时间。

准确头独立构建 Dashboard 和 Chat 生产包。34 项 TypeScript 检查和 quick premerge 通过。本地打包 Chat 经真实 SQLite Explore 历史呈现同一合成候选,实时核验 GET 可用时显示 L2,恢复服务后重试成功,三次观测全程零写入。 独立观察 1200 和 390 CSS px,成功、503 和重试恢复均确认。快速项通过,浏览器故障对照独立复现。 可用性切换使用合成的核验响应,未复验实际安装或依赖外部检查器。

我的整体评价 / Verdict

请求修改。候选选择与证明匹配有正向证据。当前读取失败时,旧存档仍可触发当前通过和档位,需要修复。 收紧实时证据来源并添加成功、503、恢复的界面断言。可以保留现有历史证据和重试按钮。 独立检查了 1200 和 390 CSS px 的实际证据包,故障、通过状态和 L2 同时可见。 已安装的 Dashboard 以及浏览器以外的产品端未独立验证。

English verdict: REQUEST_CHANGES at fdcf999. When the live verifier GET fails, the panel still renders the saved passing receipt and L2 as current. Use only a fresh same-lifetime response for current claims and report Unknown until the read recovers.

return () => controller.abort();
}, [goalId, liveAcceptance, revision, basis, lifetime]);
const current = liveAcceptance && live?.basis === basis && live.revision === revision ? live : undefined;
const contract = current?.contract ?? savedContract;

Copy link
Copy Markdown
Collaborator Author

Choose a reason for hiding this comment

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

[P2] 实时核验不可读时把已保存的通过显示成当前通过。打包 Chat 展开同一个已有合格存档的候选,把实时 Goal acceptance GET 切为 503。1200 和 390 CSS px 均同时显示「当前权威回执暂不可读」「产物检查通过」「权威终评已核对该产物」和「已验证档位 L2」。这时实时证据不可读,存档中的旧通过仍被当成当前通过。 实时模式只用同 Goal 生命周期的最新成功响应导出当前通过、精确证明和档位;等待或失败时将当前结论标为未知,并保留重试。只读历史可继续展示在存档时有效的原始证据,明确其历史时间。

This branch has not been deployed

No deployments
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.

1 participant