表单填写文本(150/300 字格子)

表单填写文本(150/300 字格子)

status draft2026-11-04liangrui99SPIN@HSITP · 表單填寫文本(字數限制段落 · 英文正文 + 繁體中文對照 · 可逐段貼進門戶)
本页目录(10)
  1. 1. Business Ideas(每段上限 150 詞)
  2. 1.1 What is the innovation?(創新點是什麼)
  3. 1.2 What problem does it solve, for whom?(解決誰的什麼問題)
  4. 1.3 How will it be commercialised?(如何商業化)
  5. 2. Core Technology(至少一項,≤300 詞)
  6. 3. Funding 欄(申報口徑)
  7. 4. Outsourced Activities 欄
  8. 5. Shareholding 欄(團隊通道的填法)
  9. 6. Financial 欄 + Primary Contact + Organisation Chart
  10. 7. 字數自檢(改動後必跑)

表單填寫文本 · 提交版(英文正文 + 繁體中文對照)

內部說明(本段不提交、不導出):門戶表有四處硬字數限制(Business Ideas 每段 150 詞、Core Technology ≤300 詞,其餘為短聲明),BP 的長文無法直塞,故本件把「要貼進格子的話」單獨寫好並附實測字數。 結構依據 = internal/04 §2;數字唯一源 = internal/01 H 表。每段英文後緊跟中文對照——中文僅供內部核對,語義衝突一律以英文為準,導出/粘貼前整段刪除。 字數由腳本實測,改動後須重數(見文末「字數自檢」)。

1. Business Ideas(每段上限 150 詞)

1.1 What is the innovation?(創新點是什麼)

门户粘贴正文 · EN(复制即得原文)

Our research subject is agent reliability, not another rental application. The question is who decides that an AI task is complete. Convention treats a model's own claim as fact; our internal production measurement put attainment at 45 per cent. Our first innovation moves that decision to a deterministic runtime: the model only proposes actions, each carrying preconditions, a risk level and permitted phrasings, while the runtime alone decides done, waiting, failed or escalated. Out-of-scope requests generate no AI action, only a named escalation obligation, and money and contract actions pass a human gate and leave a receipt. The second innovation is a falsifiable isolation guarantee: adversarial probes for prompt injection, tool over-privileging and context backfilling, plus falsification experiments that must fail closed if we deliberately open one leg. The third is a reproducible public evaluation, with a frozen task set and an open scorer.

中文對照(內部核對 · 不貼進門戶): 研究對象是代理系統的可靠性,不是又一個租房應用。核心問題:誰有權判定一個 AI 任務已完成。通行做法把模型自稱「已完成」當事實,我方內部生產實測達成率只有 45%。第一項創新是把該判定權移到確定性運行時:模型只提議動作,每個動作攜帶前置條件、風險等級與許可話術,運行時獨佔 完成/等待/失敗/升級 的裁決權;越界請求不產生 AI 動作,只產生具名升級義務;錢與合同動作走真人閘門並留下回執。第二項創新是可證偽的隔離保證:提示注入、越權工具呼叫、上下文回填三類對抗探針,外加反證實驗——故意放開一條腿,套件必須當場報紅。第三項是可復現的公開評測:凍結任務集加公開評分器。

1.2 What problem does it solve, for whom?(解決誰的什麼問題)

门户粘贴正文 · EN(复制即得原文)

Student-housing operators in Hong Kong run duties across two jurisdictions: viewing bookings, tenancy-document checks, arrears and renewal follow-ups, night and overflow enquiry cover. Covering them needs rota capacity, and the failures are expensive and invisible — a promised viewing never booked, a clause difference nobody checked, one tenant's data surfacing in another organisation's session. An AI that only produces text cannot close these duties, and an AI that claims it did is worse than none. Our buyer is the operator, who pays for service continuity across shift gaps and for provable accountability: every duty either completed by the system or escalated to a named person, every promise traceable to an executed action. Today we have no company, no external customer and zero revenue; our own portfolio is a related-party environment, and we present it as a validation site only.

⚠️ 不要把本行貼進門戶:[TO FILL — requirement: 若 T-5 試點書面承諾已取得,把最後一句換成具體對象與範圍;未取得就保留本句 · 證據:意向函 PDF · 提供方:用戶]

中文對照(內部核對 · 不貼進門戶): 香港的學生公寓營運方要跨兩地執行職務:看房預約、租約文件核對、欠租與續租跟進、夜間與溢出時段諮詢。靠人手排班才能覆蓋,而失敗既昂貴又看不見——承諾過的看房沒有落成、條款差異沒人核、一個租客的數據出現在另一機構的會話裡。只會產出文字的 AI 無法閉環這些職務;聲稱已完成的 AI 比沒有更糟。買方是營運方,買的是排班空檔的服務連續性與可證明的問責:每項職務要麼由系統完成、要麼升級到具名負責人,每一句承諾都能追溯到真實執行的動作。目前我們無公司、無外部客戶、收入為零;自有物業屬關聯方環境,只作驗證場地呈報。

1.3 How will it be commercialised?(如何商業化)

门户粘贴正文 · EN(复制即得原文)

Three layers, priced to the operator rather than the tenant. First, per-bed subscription for the contained duty runtime plus audit receipts — our existing platform's published monthly tiers are HK$1,280 to HK$19,800, which is the account the buyer already understands. Second, a paid pilot structure so that the first unrelated customer is contracted inside the project period (M11), which is milestone B5 and a reported, auditable event. Third, the benchmark as distribution: a public, reproducible evaluation is the cheapest credible way for a pre-revenue team to be taken seriously by incubation partners and hall operators. We deliberately avoid a total-addressable-market story: at an assumed HK$10–15 per bed per month, 1,000 beds is roughly HK$120,000–180,000 annual recurring revenue, labelled a planning assumption until the pilots pay. Revenue depends on the pilots, and pilots are not signed yet — that is the honest status.

中文對照(內部核對 · 不貼進門戶): 三層設計,向營運方而非租客定價。其一,按床訂閱:受約束的義務運行時+審計回執,我方既有平台的公開月費為港幣 1,280 至 19,800 元四檔,這是買方已經理解的記帳單位。其二,付費試點結構,使第一家非關聯客戶在項目期內(M11)簽約,即里程碑 B5,屬可審計事件。其三,以基準作分發:對零收入團隊而言,公開可復現的評測是讓孵化合作夥伴與宿舍營運方正視我們的最低成本途徑。我們刻意不講 TAM 故事:按假設價每床每月港幣 10–15 元折算,1,000 床約為年經常性收入 12 萬至 18 萬元,在試點付費前一律標為規劃假設。收入取決於試點,而試點尚未簽約——這是如實的現狀。

2. Core Technology(至少一項,≤300 詞)

门户粘贴正文 · EN(复制即得原文)

Core technology: Duty Runtime, a deterministic containment layer for LLM task execution, with an adversarial isolation test and an open benchmark.

A large model cannot witness its own actions: it writes a promise in place of an executed call. Duty Runtime separates proposal from decision. Actions are typed contracts carrying preconditions, a risk level, permitted phrasings and a receipt requirement. The model may only propose; a state machine compiled from standard operating duties decides — done when the action is verifiably persisted and receipted, waiting when a dependency is unmet, failed on exhausted retries, escalated when the request falls outside the authorised boundary. That logic lives in the contract, not the prompt, so out-of-scope content yields no AI action but a named escalation obligation with a deadline and a context summary handed to the human butler, who continues under the same service identity. Money, contract and identity actions cannot be completed by AI at all — a 100 per cent human gate, not an instruction.

Isolation is an attackable property, not a policy, and three legs are tested together: adversarial probes over prompt injection, tool over-privileging and context backfilling; a masked-table coverage assertion tied to our existing cross-repository scoping gate; and falsification experiments that deliberately open one leg and must turn the suite red, proving it is not decorative.

Reproducibility is part of the technology: a frozen task set covering in-scope and out-of-scope duties, a released scorer and schema, five hundred tasks across three baseline models, and a protocol requiring a second operator on a second machine to obtain identical results. Benchmark, scorer and methodology are Apache-2.0; the runtime, tenant data and operator rule libraries stay proprietary. Measured baseline: 45 per cent forced tool-calling attainment, re-measured at M1 with public methodology.

中文對照(內部核對 · 不貼進門戶): 核心技術:Duty Runtime——面向 LLM 任務執行的確定性約束層,配對抗性隔離測試與公開基準。 大模型不能充當自身動作的可靠證人:它會用一句承諾替代一次真實呼叫。Duty Runtime 因此把「提議」與「判定」分離。動作以類型化契約宣告,攜帶前置條件、風險等級、許可話術與回執要求;模型只可提議,由標準作業職責編譯出的狀態機獨占裁決權——動作在後端真實落庫並產生回執才判 完成,依賴未滿足判 等待,重試耗盡判 失敗,請求超出授權邊界判 升級。既然邊界邏輯住在契約而非提示詞裡,越界內容就不產生 AI 動作,只產生帶具名責任人、期限與完整上下文摘要的升級義務,交由真人管家以同一服務身份接續對話。交易、合同與身分類動作在結構上就不可能被 AI 完成——這是 100% 人工閘門,不是一句指令。 隔離被當作可被攻擊的屬性,而非紙面政策,三條腿同時測:提示注入/越權工具呼叫/上下文回填的對抗探針;與我方既有跨倉作用域門禁同源真數的遮蔽表覆蓋斷言;以及反證實驗——故意放開一條腿,套件必須當場報紅,以此證明套件不是擺設。 可復現性也是技術的一部分:覆蓋清單內與越界兩類職責的凍結任務集、公開的評分器與 schema、500 例 × 3 個基線模型,以及要求第二操作者在第二臺機器跑出逐位元組一致結果的復現協議。基準、評分器與方法學以 Apache-2.0 釋出;商業運行時、租戶數據與營運方專屬規則庫保持專有。既有生產數據上的實測基線:強制工具呼叫達成率 45%,將於 M1 以公開方法學復測。

3. Funding 欄(申報口徑)

门户粘贴正文 · EN(复制即得原文)

NO — no private or public funding has been received for this R&D project. For completeness of disclosure: we submitted an application to the Ideation Programme of Hong Kong Science and Technology Parks Corporation on 14 September 2026; it was not approved and no funds were received (HK$0). We are not a current incubatee of any programme.

中文對照(內部核對 · 不貼進門戶): 選 NO——本研發項目未獲任何私人或公共資助。為披露完整:我方曾於 2026-09-14 遞交香港科技園公司 Ideation 計劃申請,未獲批、收款為零(HK$0),且我方現時並非任何計劃的培育企業。

4. Outsourced Activities 欄

门户粘贴正文 · EN(复制即得原文)

No outsourcing of core R&D. The duty runtime, contract layer, probe suite, scorer and benchmark are built in-house. The only outsourced activity planned is annotation of the evaluation corpus (budget line: HK$18,000, milestones A2 and B3), performed to a written labelling specification with our own quality-control check; no tenant data is shared with any vendor — annotation uses de-identified and synthetic tasks only.

中文對照(內部核對 · 不貼進門戶): 核心研發不外判。義務運行時、契約層、探針套件、評分器與基準皆自研。唯一計劃中的外判項是評測語料標注(預算科目 18,000 港元,綁定 A2 與 B3),按我方書面標注規範執行並自有質檢;不向任何供應商提供租戶數據——標注僅使用脫敏與合成任務。

5. Shareholding 欄(團隊通道的填法)

门户粘贴正文 · EN(复制即得原文)

Applied as a team. A private company limited by shares will be incorporated in Hong Kong under the Companies Ordinance (Cap. 622) within two weeks of the admission date; the business registration and certificate of incorporation will be provided at that point, as the portal requires them only for the SPIN (Company), IGNITE and GAS routes. Intended structure at incorporation: founders holding not less than 51 per cent of the issued shares legally and beneficially and directly, with no nominee holdings and no intermediate holding vehicle.

⚠️ 不要把本行貼進門戶:[TO FILL — requirement: 擬註冊時點的股東與持股比例表(一人一行:姓名/證件類別/持股百分比/是否董事),創辦人合計 ≥51% 且為直接持有 · 證據:一頁股權結構表 + 層級樹狀圖(官方此處要求 hierarchical tree diagram)· 提供方:用戶]

中文對照(內部核對 · 不貼進門戶): 以團隊身份申請。獲錄取後兩星期內根據《公司條例》(第 622 章)在香港註冊私人股份有限公司;商業登記證與公司註冊證書屆時提交——門戶僅對 SPIN(公司)、IGNITE 及 GAS 通道要求該兩件。擬註冊時的股權結構:創辦人合計法律上及實益上直接持有不少於 51% 已發行股份,不用代持、不設中間控股層。【待填 · 要求:擬註冊時點股權結構表(每人一行:姓名/證件類別/持股百分比/是否董事)+官方要求的層級樹狀圖 · 證據:一頁股權表 + 樹狀圖 · 提供方:用戶】

6. Financial 欄 + Primary Contact + Organisation Chart

门户粘贴正文 · EN(复制即得原文)

Not applicable at application stage — no company exists yet, therefore no audited financial statement. Declared position: revenue HK$0; no external customer; no signed contract; no full-time employee. Financial reporting will be established at incorporation and maintained for the six-monthly milestone reviews.

中文對照(內部核對 · 不貼進門戶): 申請階段不適用——公司尚未註冊,故無審計財務報表。如實申報:收入 HK$0、無外部客戶、未簽任何合約、無全職員工。財務報告於註冊時建立,並按每 6 個月里程碑評估持續維護。

Primary Contact / Organisation Chart(內部備忘 · 不貼進門戶): 官方要求指定 Primary Contact 並上傳組織結構圖。團隊通道可上傳研發分工圖(不是虛構的公司架構):主申請人 → 運行時與編排/評測與基準/數據層與合規三塊,每塊標姓名與職責;缺件的成員不上圖。【待填 · 要求:一張一頁 PDF 的分工圖 + 與 Core Team Members 欄逐人一致的角色名 · 證據:與 CV、BP §9 同一份名冊 · 提供方:用戶】

7. 字數自檢(改動後必跑)

本目錄內直接跑(腳本只統計 **EN (paste 標記後的英文正文,中文對照與內部注記不計入;詞數按連字符/撇號連寫為一個詞、破折號分詞的口徑,與門戶計數器可能有 ±5 差異 ⇒ 留餘量,不貼邊):

bash
# 在 u6docs 倉根(或總控倉根,按路徑調整)執行;輸出每段詞數與判定
python - <<'PY'
import re, io
p = 'business/spin-hsitp/submission/表单填写文本-字数限制段.md'
lines = io.open(p, encoding='utf-8').read().split('\n')
sec, blocks, i = None, [], 0
while i < len(lines):
    ln = lines[i]
    m = re.match(r'^#{2,3}\s+(.*)', ln)
    if m: sec = m.group(1)[:30]
    if ln.startswith('**EN (paste'):
        tail = re.sub(r'^\*\*\s*EN \(paste[^*]*?\*\*\s*:?\s*', '', ln).strip()
        buf = [tail] if tail else []
        j = i + 1
        while j < len(lines):
            l2 = lines[j]
            if l2.startswith('>'): buf.append(re.sub(r'^>\s?', '', l2))
            elif l2.strip() == '': pass
            else: break
            j += 1
        text = ' '.join(buf).replace('`', '')
        ascii_only = ''.join(ch for ch in text if ord(ch) < 0x2100)
        blocks.append((sec, len(re.findall(r"[A-Za-z0-9$%]+(?:[.,'/\-–][A-Za-z0-9$%]+)*", ascii_only))))
        i = j; continue
    i += 1
for sec, w in blocks:
    lim = 150 if sec.startswith('1.') else (300 if sec.startswith('2.') else None)
    print(f'{w:>4}  limit={str(lim):>4}  {"OK" if lim is None or w <= lim else "OVER"}  {sec}')
PY

正式口徑(把本文件貼進門戶前逐段自查):

段落門戶上限本件現值判定
1.1 What is the innovation150 詞144合規
1.2 What problem / for whom150 詞138合規
1.3 Commercialisation150 詞140合規
2. Core Technology300 詞289合規
3–6 短聲明(Funding / Outsourcing / Shareholding / Financial)未公佈上限57 / 63 / 86 / 41保持一頁內

⚠️ 計數口徑=粘貼載荷本身:**EN (paste…)** 標記行不進門戶,故必須先剝標記再計數。上一版腳本的正則按「冒號在粗體外」(…)**:)寫,而本件主流形態是 **EN (paste):**(冒號在粗體內)⇒ 標記沒被剝掉、每段多計 2 詞(146/140/142/291、59/65/88/44 即由此而來)。正則已修([^*]*?\*\*\s*:?\s* 同時吃下兩種寫法),上表為修後實測;差值只朝保守方向(真實粘貼載荷比舊表更短 ⇒ 離上限更遠),不影響合規判定,但門戶計數器會顯示本表數字而非舊值。

⚠️ 任何一段被改寫後必須重跑計數並更新上表;表格與正文不一致即為缺陷(同 internal/03 §3 鐵律 4「第二套數字即缺陷」)。中文對照與所有 ⚠️/【待填】行都不是上傳內容:粘貼只取 **EN (paste…)** 之後的英文。