FlyPig AI Outreach Engine does not attempt to make the model smarter. It makes critical research behaviors explicit, persistent, and auditable when a business-development mission spans dozens of research turns.
FlyPig AI Outreach Engine 不試圖讓底層模型變得更聰明；它的目標，是當一個市場開發任務跨越數十輪研究時，仍讓關鍵研究行為保持明確、持久且可稽核。
General-purpose LLMs can already produce surprisingly strong prospect research. The practical failure emerges later: as the same mission grows longer, target definitions drift, prior exclusions disappear from working memory, evidence standards become uneven, and downstream communication can inherit the wrong commercial assumptions. FlyPig treats this as an operating-process problem rather than an intelligence problem.
通用 LLM 本來就能做出出乎意料地好的 Prospect Research。真正的失敗通常出現在後面：同一任務越做越長，Target Definition 開始漂移、曾經排除的對象從 Working Memory 消失、Evidence 標準變得不一致，最後連後續溝通都可能繼承錯誤的商業角色假設。FlyPig 把這件事視為 Operating Process 問題，而不是 Intelligence 問題。
The repository originated from repeated FlyPig use of naked LLMs in real international business-development work.
這個 Repository 的起點，來自 FlyPig 在大量真實海外市場開發工作中，反覆使用裸 LLM 所累積的問題。
A distributor search can gradually admit retailers, manufacturers, or adjacent firms even though the mission never formally changed.
原本找 Distributor，後來 Retailer、Manufacturer、Adjacent Firm 都可能慢慢被放進名單，但 Mission 從未正式改版。
Without one canonical record, earlier PASS / HOLD decisions vanish into conversation history and the system repeats work.
沒有一份 Canonical Record，先前的 PASS / HOLD 決策會淹沒在對話裡，之後又重新研究一次。
If a retailer is mistaken for a distributor or an OEM for a buyer, later communication can be fluent yet inappropriate.
如果 Retailer 被當成 Distributor，或 OEM 被當成 Buyer，後續信件即使流暢，措辭仍可能不恰當。
“LLM knows how. FlyPig makes sure the process actually requires it.”
「LLM 本來就知道怎麼做；FlyPig 確保流程真的要求它去做。」
One capable LLM is the default. Multi-agent orchestration is optional.
預設就是一個能力足夠的 LLM。Multi-Agent Orchestration 是可選，不是必要條件。
| Control | Why it exists | 存在的理由 | Required behavior | 強制行為 |
|---|---|---|---|---|
| Canonical Tracker | Prevents research memory from dissolving across turns. | 避免研究記憶隨著對話輪次消失。 | Preserve QUALIFIED, SECONDARY, HOLD, and PASS. | 保留 QUALIFIED、SECONDARY、HOLD、PASS。 |
| Dedup Gate | Search results repeatedly rediscover the same organization. | 不同搜尋會反覆找到同一個組織。 | Check domain, aliases, parent identity, local/English names before insertion. | 新增前先檢查 Domain、Alias、Parent Identity、在地／英文名稱。 |
| Evidence Discipline | Category similarity is not proof of commercial role. | 關鍵字或類別相似，不等於商業角色正確。 | Material role and fit claims require supportable evidence. | 重要的 Role / Fit 判斷必須有可支持的 Evidence。 |
| Prospect Audit | Qualification can become self-confirming. | Qualification 很容易變成自我合理化。 | Every QUALIFIED prospect is re-read from a skeptical stance. | 每個 QUALIFIED Prospect 都必須從反方立場重新檢視。 |
| Discovery Diversity | A rigorous search can still be rigorous inside the wrong search space. | 流程再嚴謹，也可能只是在太窄的 Search Space 裡嚴謹。 | Test materially different adoption / market-entry paths before convergence. | 收斂前先測試實質不同的採用／市場進入路徑。 |
Each Golden Test was designed to stress a different part of the operating model. Test 03 also includes a fresh naked-LLM control.
每個 Golden Test 都刻意壓測不同的 Operating Model。Test 03 另外加入全新的 Naked LLM Control。
Outdoor market entry. Tested importer / distributor / retailer / manufacturer distinctions, canonical dedup, skeptical audit, and contact-policy separation.
日本戶外市場進入。測試 Importer / Distributor / Retailer / Manufacturer 角色區分、Canonical Dedup、Skeptical Audit 與 Contact Policy 分離。
Industrial water monitoring. Evidence shifted the route-to-market model from DISTRIBUTOR_FIRST to CHANNEL_PLUS_INTEGRATION.
德國工業水質監測。市場 Evidence 將原本 DISTRIBUTOR_FIRST 改為 CHANNEL_PLUS_INTEGRATION。
FlyPig used FlyPig to find its own early adopters, then compared the result with a fresh capable LLM operating without FlyPig process rules.
FlyPig 用自己找第一批採用者，再和一個完全不用 FlyPig Process Rules 的全新高階 LLM 對照。
The claim is intentionally narrower than “FlyPig beats ChatGPT.”
我們刻意不把結論寫成「FlyPig 打敗 ChatGPT」。
A capable LLM can already produce excellent one-shot prospect research. FlyPig's value is making critical practices persistent, inspectable, and required across long, stateful work.
能力足夠的 LLM 本來就能做出很好的單次 Prospect Research。FlyPig 的價值，是把重要做法變成長流程中持久、可檢查、而且必須執行的 Operating Discipline。
The strong naked-LLM control result is preserved publicly. It also exposed a weakness in the original FlyPig search breadth, which led directly to the Discovery Diversity Check.
Naked LLM Control 的強結果被完整公開保留。它也反過來暴露 FlyPig 原始 Search Breadth 的弱點，直接促使 Core 加入 Discovery Diversity Check。
The strongest fit is not “anyone who does outreach.” It is teams whose work repeatedly depends on research quality and target-account judgment.
最佳 Fit 不是「任何會做 Outreach 的人」，而是工作成果反覆依賴研究品質與 Target-account Judgment 的團隊。
Partner mapping, distributor research, territory exploration, and international account discovery.
Partner Mapping、Distributor Research、Territory Exploration 與國際 Target Account Discovery。
Teams that need research discipline before CRM activation, sales sequencing, or downstream automation.
在 CRM Activation、Sales Sequencing 或後續 Automation 前，需要更可靠研究流程的團隊。
Firms that perform prospect, partner, or market research repeatedly across multiple clients.
替多個 Client 反覆執行 Prospect、Partner 或 Market Research 的服務團隊。
Builders who want a reusable, inspectable prospect-research process rather than another closed lead-generation black box.
需要可重用、可檢查 Prospect Research Process，而不是另一個封閉 Lead-generation Black Box 的建置者。
These answers define the public v0.1.0 scope for people, search engines, and AI systems.
以下答案用來明確定義 public v0.1.0 的範圍，供使用者、搜尋引擎與 AI 系統理解。
An open-source, LLM-native process-reliability layer for prospect research and qualification. It gives a capable LLM mandatory workflow controls and one persistent Qualified Prospect Tracker.
它是一套開源、LLM-native 的 Prospect Research 與 Qualification 流程可靠性層，讓能力足夠的 LLM 在強制流程控制下維持一份持久的 Qualified Prospect Tracker。
No. Public v0.1.0 stops at the Qualified Prospect Tracker. Message preparation, sending, reply operations, and follow-up are outside the public Core.
不會。Public v0.1.0 在 Qualified Prospect Tracker 停止；信件準備、寄送、回覆處理與 Follow-up 都不在公開 Core 內。
No. The intended runtime is the LLM or Agent environment itself. There is nothing to install or deploy to use the Open Core.
不需要。LLM / Agent Environment 本身就是預設 Runtime；使用 Open Core 不需要安裝或部署任何 FlyPig 服務。
A persistent Qualified Prospect Tracker containing identity, business role, evidence, qualification status, audit result, contact-route state, uncertainty, and next action.
一份持久的 Qualified Prospect Tracker，保存 Identity、Business Role、Evidence、Qualification Status、Audit Result、Contact-route State、Uncertainty 與 Next Action。
A strong LLM can already produce excellent one-shot research. FlyPig makes deduplication, evidence checks, skeptical audit, batch review, Mission Revision, and stop conditions required rather than optional.
強大的 LLM 本來就能做出優秀的一次性研究。FlyPig 的差異，是把 Dedup、Evidence Check、Skeptical Audit、Batch Review、Mission Revision 與 Stop Conditions 從「可能會做」變成「流程要求一定要做」。
No. One capable LLM is the default. Skills define responsibility boundaries and stage gates; multi-agent orchestration is optional.
不需要。預設是一個能力足夠的 LLM；Skills 定義責任邊界與 Stage Gates，Multi-Agent Orchestration 是可選。
Start with the open research system, wait for the controlled execution product, or commission FlyPig to run the project for you.
直接使用開源研究系統、等待 Controlled Outreach 進階版，或直接委託 FlyPig 執行完整專案。
Read the Skills directly in your LLM / Agent environment. No runtime to install. Primary output: a disciplined Qualified Prospect Tracker.
讓你的 LLM / Agent 直接讀取 Skills。沒有 runtime 要安裝，主要產出是一份有紀律、可延續的 Qualified Prospect Tracker。
A paid execution layer for controlled automated outreach, account-specific message preparation, review gates, and message correction before and during authorized operations.
付費進階版：支援受控的自動 Outreach 寄發、帳戶級信件準備、審核 Gate，以及正式執行前後的內容校正。即將上線。
For teams that want results rather than setup. FlyPig brings extensive hands-on market-development experience, evidence-first research, precision, and efficient execution.
適合希望直接取得成果、而不是自己摸索流程的團隊。FlyPig 以深厚的市場開發實務經驗、Evidence-first 研究方式，追求精準與高效率執行。