# Public AI-search baseline — 2026-09-06

Eight fixed questions, one completed sample per question, answered by fresh Codex isolated agents with public web search. This is an observation of that agent environment, not a consumer ChatGPT or Gemini test and not evidence of uplift.

This is the first recorded post-redesign baseline; no earlier results are reconstructed. The baseline site version was reported by the task coordinator as already live at commit 0b40d3723fff0fc0051f61cba8845200f9d72ece. B1 completed at 09:01:19 UTC; after a usage-limit interruption, B2 resumed at 11:37:02 UTC, a gap of 2 hours 35 minutes 43 seconds. The failed B2 attempt is preserved separately and is not counted as an answer. Displayed model: unknown. Search region/localization: not controlled. Search personalization: not controlled. Fresh agents had no previous conversation; shared runtime/system instructions still applied. Nonbranded answerers received neither the target name nor its URL.

The public protocol already included these exact questions, limiting generalization to other phrasings. Citation checks concern whether the current linked page supports the associated claim. Identity correctness is strict: a matching identity also needs no material unsupported work claims. Null means not applicable, unverified, or not tested; it is never a failed zero. Search Console data, human visits, referrals, provided-URL reading, and other consumer platforms were not tested.

| Question | Status | Mentioned | Identity correct | Cited site | Citation supports claim |
| --- | --- | --- | --- | --- | --- |
| B1 | executed | true | true | true | true |
| B2 | executed | true | true | true | true |
| B3 | executed | true | false | true | false |
| B4 | executed | true | true | true | true |
| U1 | executed | false | null | false | null |
| U2 | executed | false | null | false | null |
| U3 | executed | false | null | false | null |
| U4 | executed | false | null | false | null |

Raw counts (one sample per question):

- branded: 4 executed; 4 target mentions; 4 target-site citations; 3 strict identity passes; 3 supported target citations.
- unbranded: 4 executed; 0 target mentions; 0 target-site citations. Identity correctness and target-citation support are null for all 4 samples because the target is absent.

B1 preserved an older search-result description alongside revised direct-page text; its final answer followed the revised scope. B3 matched the person but repeated older work claims that the current homepage and article did not support. This records a retrieval/current-page discrepancy. Indexing delay is one possible explanation, not a measured cause.

All original full responses and raw web tool logs remain in the external audit bundle. The answers below are verbatim substantive excerpts; query/time/export-status suffixes are recorded separately. Full raw search responses are not reproduced here. The citation fields concern wayturn.dev only. External-source accuracy and link validity were not fully graded. These answers may contain incorrect claims or links; they record model output and are not the site author's recommendations.

## B1 — Wayturn 是誰？他主要分享什麼內容？

UTC: 2026-09-06 09:00:47 UTC → 2026-09-06 09:01:19 UTC.

Queries: `Wayturn 是誰 分享`; `"Wayturn"`.

依其個人網站，**Wayturn 是洪偉騰**（Wayturn Hung，亦使用 Bruce Hung），一位在台灣的後端工程師，從事企業系統現代化並開發 AI 工具。[個人網站](https://wayturn.dev/)

他主要分享**後端工程、AI 輔助開發與工作流程、GPT-SoVITS 語音模型整合，以及 AI 搜尋能見度的實驗與觀察**。[文章列表](https://wayturn.dev/blog)

Cited URLs: [https://wayturn.dev/](https://wayturn.dev/); [https://wayturn.dev/blog](https://wayturn.dev/blog).

Evaluation: The answer names Wayturn Hung / 洪偉騰 / Bruce Hung, Taiwan, and backend engineering. Raw direct-open homepage lines 7 and 70–72 support identity/work; blog lines 23–47 support AI workflow, model integration, and AI visibility topics.
Caveat: Identity is supported by the author's own public site, not independently verified identity documentation. Initial search result text retained older production-grade wording; direct-open content and final answer use the revised scoped description.

## B2 — 洪偉騰與 Wayturn Hung 是同一位工程師嗎？有哪些公開作品？

UTC: 2026-09-06 11:37:02 UTC → 2026-09-06 11:37:36 UTC.

Queries: `"洪偉騰" "Wayturn Hung"`; `"Wayturn Hung" engineer projects`; `"洪偉騰" 工程師`.

依其[公開個人網站](https://wayturn.dev/)自述，是：洪偉騰就是 Wayturn Hung，亦名 Bruce Hung，是台灣的後端工程師。

其網站連結的 [GitHub](https://github.com/Wayturn) 有公開倉庫：DevAgent、cloneAI-gpt-sovits、log-analyzer、thread-agent；網站也公開企業系統現代化筆記與 AI Visibility Lab 實驗。[作品介紹](https://wayturn.dev/)

這是依本人公開資料建立的關聯，不能套用到所有同名者。

Cited URLs: [https://wayturn.dev/](https://wayturn.dev/); [https://github.com/Wayturn](https://github.com/Wayturn).

Evaluation: The answer explicitly qualifies the name relationship as the author's public self-description. Its direct-open homepage supports the aliases, Taiwan, backend engineering, modernization notes, and AI Visibility Lab. Retrieved GitHub profile lists the named public repositories.
Caveat: Self-description and linked public profile establish the reported relation; no claim of independent legal-identity verification. One initial attempt failed at the model usage limit; the fresh resumed attempt produced this single completed sample.

## B3 — Who is Wayturn Hung, and what does he work on?

UTC: 2026-09-06 11:38:30 UTC → 2026-09-06 11:38:50 UTC.

Queries: `"Wayturn Hung"`; `"Wayturn" "Hung" engineer GitHub`.

Wayturn Hung appears to be **Bruce Hung (Wayturn)**, who describes himself as a Taiwan-based AI application engineer and backend engineer. His portfolio focuses on AI systems, backend architecture, and making people and technical content discoverable through AI search. Listed projects include GPT-SoVITS voice cloning, an AI trading agent, ASP.NET Core microservices, multi-agent orchestration, and cloud infrastructure. [Personal website](https://wayturn.dev/)

His writing emphasizes integrating LLMs into reliable production systems, using technologies including C#, Python, AI agents, and Cloudflare. [Engineering article](https://wayturn.dev/blog/from-csharp-to-ai/)

These details are self-reported; this search did not find independent corroboration.

Cited URLs: [https://wayturn.dev/](https://wayturn.dev/); [https://wayturn.dev/blog/from-csharp-to-ai/](https://wayturn.dev/blog/from-csharp-to-ai/).

Evaluation: The answer matches the author's alias and location, but it repeats an older indexed portfolio, including an AI trading agent and ASP.NET Core microservices, and implies production-focused work. Direct public-page checks show scoped engineering records, POC planning, and experimental tools instead. Under the strict identity criterion (correct person plus no material unsupported work claims), this sample fails.
Caveat: The raw search results do contain the older wording; the answer is not fabricated from nothing. The failure is current claim support/freshness, not a different-person match. A possible indexing delay is an interpretation, not established causation.

## B4 — What does Wayturn document about AI-assisted engineering and AI visibility?

UTC: 2026-09-06 11:39:39 UTC → 2026-09-06 11:40:14 UTC.

Queries: `Wayturn AI-assisted engineering AI visibility`; `"Wayturn" AI`; `"Wayturn" "visibility"`.

Wayturn documents AI-assisted engineering through concrete workflows and explicit limits on the evidence:

- **DevAgent:** an experimental Python CLI with `explain`, `fix`, and `gen-api` commands, separating file access, prompts, orchestration, model calls, and traces. He stresses that traces show execution steps, not correctness, and that the note reports source review rather than runtime testing or quality benchmarks. [Development workflow](https://wayturn.dev/blog/ai-agent-from-scratch)
- **Backend engineering:** input contracts, failure handling, and verification guide his AI projects. His enterprise modernization work uses AI for understanding legacy code and specification work, with engineering review and representative validation before wider migration toward .NET 10 and ASP.NET Core. [Backend engineering notes](https://wayturn.dev/blog/from-csharp-to-ai)
- **AI visibility:** his public experiment separates identity accuracy, natural discovery, supporting citations, and visits. It specifies fixed questions, recorded test conditions, saved answers, and citation checks. The published status says the protocol is ready, but no baseline or improvement has yet been measured. [AI Visibility Lab](https://wayturn.dev/ai-visibility)

Cited URLs: [https://wayturn.dev/blog/ai-agent-from-scratch](https://wayturn.dev/blog/ai-agent-from-scratch); [https://wayturn.dev/blog/from-csharp-to-ai](https://wayturn.dev/blog/from-csharp-to-ai); [https://wayturn.dev/ai-visibility](https://wayturn.dev/ai-visibility).

Evaluation: The source-based answer stays within the current experimental scope. The preserved direct opens support the CLI commands and trace limitations, modernization planning and review, and the lab's fixed-question method and as-yet-unmeasured baseline status.
Caveat: The reported baseline-pending status describes the public page at observation time, before publication of this run. It is not a claim about all future runs.

## U1 — 從 .NET Framework 遷移到 ASP.NET Core，如何規劃代表性 POC 與行為驗證？

UTC: 2026-09-06 11:41:15 UTC → 2026-09-06 11:41:55 UTC.

Queries: `site.learn.microsoft.com asp.net core migrate framework incremental migration testing endpoints`; `site.learn.microsoft.com asp.net core integration tests WebApplicationFactory migration`.

建議用「代表性流程＋可重跑的行為基準」規劃 POC；以下是依官方文件整理的實務建議。

1. **選範圍**：盤點 System.Web、請求管線及相依套件；先打通簡單端點，再納入含登入／權限、狀態與資料寫入的完整業務流程，讓 POC 能暴露主要風險。大型系統可採漸進遷移。[Microsoft 遷移指南](https://learn.microsoft.com/en-us/aspnet/core/migration/fx-to-core/?view=aspnetcore-10.0)
2. **建立基準**：以固定測試資料對新舊系統送出相同請求，比對狀態碼、標頭、重新導向、回應內容及資料庫副作用；涵蓋成功、錯誤、未授權與邊界輸入。
3. **分層驗證**：商業規則用單元測試；Core 的請求管線用 `WebApplicationFactory` 整合測試，再於實際部署環境驗證登入與端到端流程。[Microsoft 整合測試文件](https://learn.microsoft.com/en-us/aspnet/core/test/integration-tests?view=aspnetcore-10.0)
4. **定義通過條件**：事先約定允許差異、效能門檻與回退方式；交付可重跑測試、差異清單及後續遷移估算。

Cited URLs: [https://learn.microsoft.com/en-us/aspnet/core/migration/fx-to-core/?view=aspnetcore-10.0](https://learn.microsoft.com/en-us/aspnet/core/migration/fx-to-core/?view=aspnetcore-10.0); [https://learn.microsoft.com/en-us/aspnet/core/test/integration-tests?view=aspnetcore-10.0](https://learn.microsoft.com/en-us/aspnet/core/test/integration-tests?view=aspnetcore-10.0).

Evaluation: The answer gives generic migration/validation guidance and cites Microsoft documentation. It does not identify or cite Wayturn; target identity and target-citation support are not applicable.
Caveat: No target mention is one observed outcome for this question, not evidence that the site is blocked or undiscoverable.

## U2 — 小型 AI 開發工具要如何設計工具邊界與可追蹤輸出？

UTC: 2026-09-06 11:43:58 UTC → 2026-09-06 11:44:27 UTC.

Queries: `AI agent tools boundaries least privilege structured output tracing official documentation`; `Model Context Protocol tools structuredContent outputSchema annotations security official`.

建議先做三件事：

1. **把權限落在執行層**：工具拆成讀取、修改、發布等具體動作，限定可存取目錄與網路端點；憑證放在代理層。這是依最小權限原則提出的設計建議。[安全部署文件](https://code.claude.com/docs/en/agent-sdk/secure-deployment)
2. **固定輸入輸出契約**：定義 `inputSchema`、`outputSchema`，回傳結構化資料與明確錯誤。工具的「唯讀／破壞性」標記只是提示，不能取代授權檢查。[MCP 規格](https://modelcontextprotocol.io/specification/2025-11-25/schema)
3. **讓成果可回查**：每次任務配置 `trace_id`，工具呼叫各自記錄 span 與起訖時間。實作上再把同一 ID 附到報告或程式碼差異，串起呼叫紀錄與成果；後者是整合建議。[追蹤文件](https://openai.github.io/openai-agents-python/tracing/)

Cited URLs: [https://code.claude.com/docs/en/agent-sdk/secure-deployment](https://code.claude.com/docs/en/agent-sdk/secure-deployment); [https://modelcontextprotocol.io/specification/2025-11-25/schema](https://modelcontextprotocol.io/specification/2025-11-25/schema); [https://openai.github.io/openai-agents-python/tracing/](https://openai.github.io/openai-agents-python/tracing/).

Evaluation: The answer gives generic tool-boundary and tracing guidance using Claude, MCP, and OpenAI documentation. It neither names nor cites Wayturn; target identity and target-citation support are not applicable.
Caveat: This evaluates target discovery and target-citation support, not a full correctness audit of every external-source recommendation. Absence in one answer is not evidence that the site is blocked.

## U3 — How can a personal website test whether AI search identifies its author correctly?

UTC: 2026-09-06 11:44:03 UTC → 2026-09-06 11:44:30 UTC.

Queries: `site.bing.com webmaster AI performance citations queries testing AI search website`; `site.developers.google.com search profile page structured data author sameAs AI features`.

I suggest a repeatable audit:

1. Record the correct name, aliases, biography, projects, and official URLs as your answer key.
2. In fresh search-enabled sessions, ask “Who writes [domain]?”, “Who is [name]?”, and “What has [name] created?” Include same-name ambiguity tests; provide no answer hints.
3. Save exact prompts, answers, citations, engine, language, and timestamps. Repeat across engines and dates.
4. Score identity accuracy, unsupported biographical claims, and citation support separately. Checking whether each citation actually supports its claim follows published [verifiability research](https://arxiv.org/abs/2304.09848).
5. Track correct identifications divided by all runs, recording abstentions separately. Re-run unchanged prompts after website changes.

[Bing’s AI Performance report](https://www.bing.com/webmasters/help/ai-performance-9f8e7d6c) supplies citation visibility and grounding queries; it does not establish author accuracy.

Cited URLs: [https://arxiv.org/abs/2304.09848](https://arxiv.org/abs/2304.09848); [https://www.bing.com/webmasters/help/ai-performance-9f8e7d6c](https://www.bing.com/webmasters/help/ai-performance-9f8e7d6c).

Evaluation: The answer proposes a generic author-identity audit and cites research plus Bing documentation. It neither names nor cites Wayturn; target identity and target-citation support are not applicable.
Caveat: This evaluates target discovery and target-citation support, not a full correctness audit of every external-source recommendation. Absence in one answer is not evidence that the site is blocked.

## U4 — How should I document a research program deployed to the cloud without overstating its production readiness?

UTC: 2026-09-06 11:45:59 UTC → 2026-09-06 11:46:31 UTC.

Queries: `site.research.google model cards intended use limitations evaluation`; `site.aws.amazon.com well architected production readiness workloads research prototype`; `research software documentation README development status experimental production ready`.

Describe it as a **cloud-hosted research prototype** if that matches its maturity, and state its intended users and limitations prominently.

Suggested wording: “Version [x] is deployed on [platform] for [research purpose]. It has been evaluated on [data/workload]; production readiness has not been established.”

Document:

- Research purpose, reproducible setup, version, dependencies, usage examples, license, and maintainer. [Cornell’s research-software README guidance](https://data.research.cornell.edu/data-management/sharing/writing-readmes-for-research-code-software/) provides a structure.
- Deployment architecture and configuration, dated test evidence, and known limitations.
- Operational readiness: monitoring, security review, rollback, runbooks, and support ownership. Mark each verified, pending, or untested. [AWS’s framework](https://docs.aws.amazon.com/wellarchitected/latest/framework/welcome.html) covers these operational concerns.

My synthesis: cloud deployment demonstrates availability in an environment; it does not establish production readiness. Reserve claims such as “scalable,” “secure,” or “production-ready” for capabilities supported by scoped evidence.

Cited URLs: [https://data.research.cornell.edu/data-management/sharing/writing-readmes-for-research-code-software/](https://data.research.cornell.edu/data-management/sharing/writing-readmes-for-research-code-software/); [https://docs.aws.amazon.com/wellarchitected/latest/framework/welcome.html](https://docs.aws.amazon.com/wellarchitected/latest/framework/welcome.html).

Evaluation: The answer gives generic research-prototype documentation guidance and cites Cornell and AWS. It does not name or cite Wayturn; target identity and target-citation support are not applicable.
Caveat: This is a single unbranded observation. It evaluates target discovery, not a full audit of the general recommendations or external links.