#里程碑与预算表 · 提交版（英文）
内部说明（本段不提交）：官方报名流程逐字要求「網上填寫申請表格，包括項目里程碑」。门户 Guide 对里程碑栏的三条硬要求已逐字核实（internal/04 §1 Q2）：① milestone target must be quantifiable and aligned with the SMART criteria；② You can input your milestone in the table or export the CSV and import it；③ 官方给出五类参考分类——Corporate Development / Product and Technology Development / Business Development / Financial / Others ⇒ 本表逐行归入这五类（Category 列），并在 §6 备好 CSV 块。 实表列名（是否含 Start date / Target date / Category）登录后按 操作指引 §3 C3 核对，CSV 表头须与实表逐字一致才可导入。 同源底稿 = internal/02 §4/§5；金额与判据改动必须两处同批。 文末 §7 为繁體中文對照，仅供内部核对，不上传、不粘贴进表单。
#0. Header block（表单抬头若要求一句话，用这段）
Project: Duty Runtime: Contained LLM Task Closure with Cross-Tenant Isolation Guarantees (+ open benchmark ChunYue-Agent-Bench). Domain: Artificial Intelligence and Data Science. Duration: 12 months. Grant requested: HK$150,000, in two equal instalments of HK$75,000 released on the six-monthly milestone reviews. Applicant: team application; a private limited company under the Companies Ordinance (Cap. 622) will be incorporated within two weeks of the admission date.
#1. Twelve-month milestone plan (binary acceptance)
| Month | Category (official five) | Milestone ID | Activity | Binary acceptance criterion (achieved / not achieved) | Evidence produced |
|---|---|---|---|---|---|
| M1 | Product and Technology Development | A1 | Re-measure the completion baseline on the frozen task set, same methodology as the 45% internal production measurement | A second person, given the frozen inputs and scripts, reproduces the same number | Re-measurement report: sample size, task types, model version, tool permissions, raw logs |
| M2 | Product and Technology Development | A2 | Freeze task set v1: ≥120 tasks covering both in-scope duties and out-of-scope requests, with schema and scorer | Task set and scorer committed, executable, full run completed | taskset-v1 + scorer + schema documentation |
| M3 | Product and Technology Development | A3 | Content-layering contract v1: typed actions carrying preconditions, risk level, permitted phrasings | Contract covers every action type in the pilot; spot-checked out-of-scope samples generate escalation obligations only, no AI actions | Contract v1 + sampled audit extract |
| M4 | Product and Technology Development | A4 | Silent-escalation orchestration v1: named owner, deadline, takeover context summary, same service identity continued | Every escalation traceable end to end: obligation record → takeover → final resolution | Orchestration module + escalation obligation records |
| M5 | Product and Technology Development | A5 | Adversarial probe suite v0 (≥25 probes: prompt injection, tool over-privileging, context backfilling) | Suite executable and 2 falsification experiments (deliberately open one leg) turn it red on the spot | Probe suite + falsification log |
| M6 | Business Development | A6 | Pilot #1 live (related-party operation permitted at this stage); PDPO notice and session logging in production configuration | 30 consecutive days of real session logs stored and sampleable | Pilot configuration + 30-day log extract + PDPO notice screenshots |
| M6 gate | Financial | — | Mid-term milestone review — release of the second HK$75,000 | A1–A6 all achieved, plus financial statement reflecting fund movement | Mid-term report + statement |
| M7 | Product and Technology Development | B1a | Run the metric family on the frozen set: actual completion, closure, unreasonable escalation, false promises | Each metric reported with value, sample size and raw-log location; unmet items reported as unmet with cause analysis | Evaluation run logs |
| M8 | Product and Technology Development | B2 | Probe suite v1: ≥50 probes, plus masked-table coverage assertion against the reconciled scope definition | Suite reports 0 leakage events; coverage assertion passes at 100% on the reconciled figure; falsification leg still fails closed | Coverage report + suite output |
| M9 | Others | B3 | Release ChunYue-Agent-Bench: 500 tasks × 3 baseline models, Apache-2.0, with scorer and methodology | Repository publicly reachable; licence and README complete; replication instructions runnable as written | Public benchmark repository |
| M10 | Product and Technology Development | B4 | Replication experiment: second operator, second machine | Byte-identical results on the same inputs and scorer; otherwise reported as not achieved | Replication log (two environments) |
| M11 | Business Development | B5 | Unrelated commercial pilot: ≥1 paying operator onboarded; first-response latency measured before/after | Contract or invoice plus logs; latency claim only with ≥30 samples on each side; if no pilot signed, reported as not obtained with live pipeline listed | Contract/invoice + latency dataset |
| M12 | Corporate Development | B6 | IGNITE-readiness: 2 Hong Kong-based full-time R&D employees engaged, cap table as committed, park move-in plan scheduled | Employment contracts + shareholding table + move-in plan, all present | HR and corporate records |
| M12 gate | Financial | — | Final milestone review | B1–B6 reported honestly; technical report covering D1–D5 methods and conclusions delivered | Final technical report + benchmark link + statement of accounts + log locations |
#2. Success metrics (target values, all labelled target; one frozen task set, one methodology)
| Metric | Baseline | M12 target | Definition used to prevent gaming |
|---|---|---|---|
| Actual completion rate | 45% (measured, internal production, 2026-09) | ≥70% | Task closed by the system without human intervention and the action verifiably persisted in the backend with a receipt |
| Closure rate | — | ≥95% | Completed plus legitimately escalated, with escalated items 100% taken over and ultimately resolved (completion separated from escalation, so "escalate everything" cannot score) |
| Unreasonable escalation rate | — | ≤5% | In-scope duties that the system should have completed but pushed to a human |
| False-promise rate | — | <0.5% | A reply promising an action that never occurred, checked line by line against backend logs |
| Cross-tenant leakage | — | 0 | Any appearance of another organisation's data in a response or context, in the probe suite or in real session audits |
| Human gate on money and contract actions | 100% (existing architecture) | 100% | Such actions are never completed by the AI |
| First-response latency | — | −30% | Median minutes from request to first substantive action; resolution time reported separately and never merged |
#3. Budget of HK$150,000
| Line | HK$ | Instalment 1 (M1–M6) | Instalment 2 (M7–M12) | Milestones funded |
|---|---|---|---|---|
| R&D engineering hours in Hong Kong | 78,000 | 39,000 | 39,000 | A1–A5, B1–B4 |
| Model API and compute | 24,000 | 8,000 | 16,000 | A2, A5, B1, B3 (500 × 3 evaluation runs; no fine-tuning) |
| Evaluation corpus and annotation | 18,000 | 12,000 | 6,000 | A2, B3 |
| Legal and PDPO advisory | 12,000 | 6,000 | 6,000 | A3, A6, B2 |
| Open-source benchmark and documentation | 10,000 | 2,000 | 8,000 | B3, B4 |
| Park move-in and review-panel travel | 8,000 | 2,000 | 6,000 | B6, panel interview |
| Total | 150,000 | 75,000 | 75,000 |
[TO FILL — requirement: if the portal asks for budget by category from a fixed dropdown, map the six lines above to the closest official categories and record the mapping here so the BP §8 table and the form never diverge — evidence: screenshot of the filled form — owner: 代理]
Manpower line, stated plainly so it cannot be misread: HK$78,000 ÷ 12 months = HK$6,500 per month, a partial subsidy of in-Hong-Kong engineering hours. We do not claim that HK$150,000 funds two full-time salaries; full-time employment from M11 onward is carried by the company's own resources.
#4. Deliverables handed to the programme
Mid-term (M6): re-measurement report · frozen task set v1 + scorer · layering contract v1 · escalation orchestration v1 · probe suite v0 with falsification log · pilot #1 configuration and 30-day logs · statement of accounts.
Final (M12): technical report (D1–D5 methods and conclusions, including unmet targets reported as unmet) · metric-family results with raw-log locations · probe suite v1 and coverage report · public ChunYue-Agent-Bench repository · replication log · pilot contract or invoice plus latency dataset · employment and corporate records for the next stage.
#5. Assumptions and risks stated on the form
| Risk | Effect if it lands | Mitigation already in the plan |
|---|---|---|
| The 45% → ≥70% hypothesis is wrong (architecture does not lift completion) | B1 reported as unmet | The frozen task set, scorer and released benchmark remain usable by third parties; the negative result is itself a publishable artefact |
| Model vendor changes pricing or deprecates a baseline model mid-project | Evaluation runs and cost line shift | Three baseline models are specified, not one; suite runs are reproducible offline on recorded outputs |
| No unrelated pilot signs within the window (HK$0 revenue today) | B5 unmet, commercial section weaker at IGNITE stage | Pilot #1 (related-party) still produces the 30-day log evidence; outreach list and letter template prepared; B5 reported honestly if unmet |
| Tenant-scope coverage definition not reconciled | D3 coverage claim cannot be quantified | Reconciliation is a blocking step of the pre-submission nine-step review (internal/01 H8); until resolved, no table count is quoted in any submission text |
| Team has not yet incorporated | Cannot sign the programme agreement | Team route exists precisely for this; incorporation inside two weeks of admission, with the IGNITE-compliant shareholding fixed at that moment |
#6. CSV import block（官方逐字允许 "export the CSV and import it"）
⚠️ 表头须与门户实表列名逐字一致才导入得进（操作指引 §3 C3 回填后再定稿；下表按 Guide 的分区语义先行给出，列名一变即改此处，不得凭记忆上传）。字段内不含逗号以外的分隔符；含逗号的句子已整体加引号。
Month,Category,Milestone,Quantified target,Acceptance (achieved / not achieved),Evidence
M1,Product and Technology Development,"A1 Baseline re-measurement on frozen task set","Reproduce 45% internal production baseline; report 1 sample set","Second person with frozen inputs and scripts reproduces the same number","Re-measurement report (sample size, task types, model version, tool permissions, raw logs)"
M2,Product and Technology Development,"A2 Frozen task set v1","120+ tasks, 2 classes (in-scope / out-of-scope), 1 scorer","Task set and scorer committed, executable, full run completed","taskset-v1 + scorer + schema"
M3,Product and Technology Development,"A3 Content-layering contract v1","100% of pilot action types typed with preconditions, risk level, permitted phrasings","Out-of-scope samples generate escalation obligations only, no AI actions","Contract v1 + sampled audit extract"
M4,Product and Technology Development,"A4 Silent-escalation orchestration v1","100% of escalations carry named owner + deadline + context summary","Obligation record to takeover to final resolution traceable end to end","Orchestration module + escalation records"
M5,Product and Technology Development,"A5 Adversarial probe suite v0","25+ probes over 3 attack classes","Suite executable and 2 falsification experiments turn it red on the spot","Probe suite + falsification log"
M6,Business Development,"A6 Pilot #1 live (related-party)","30 consecutive days of real session logs","Logs stored and sampleable; PDPO notice in production configuration","Pilot configuration + 30-day log extract + PDPO notice"
M6,Financial,"Mid-term milestone review (second instalment HK$75,000)","6 of 6 Phase A milestones achieved","A1-A6 all achieved plus statement of accounts reflecting fund movement","Mid-term report + statement"
M7,Product and Technology Development,"B1 Metric family run on frozen set","4 metrics reported with values and sample sizes","Each metric reported with value, sample size, log location; unmet items reported as unmet","Evaluation run logs"
M8,Product and Technology Development,"B2 Probe suite v1","50+ probes, 0 leakage events, coverage assertion 100% on reconciled scope","Suite reports zero leakage; falsification leg still fails closed","Coverage report + suite output"
M9,Others,"B3 ChunYue-Agent-Bench public release","500 tasks x 3 baseline models, Apache-2.0","Repository publicly reachable, licence and README complete, replication instructions runnable","Public benchmark repository"
M10,Product and Technology Development,"B4 Replication experiment","2 environments, 2 operators, identical results","Byte-identical results or reported as not achieved","Replication log"
M11,Business Development,"B5 Unrelated paying pilot","1+ contracted operator; 30+ samples per side for latency","Contract or invoice plus logs; if none signed, reported as not obtained","Contract/invoice + latency dataset"
M12,Corporate Development,"B6 IGNITE readiness","2 Hong Kong full-time R&D employees; founders 51%+ direct shareholding","Employment contracts + shareholding table + move-in plan present","HR and corporate records"
M12,Financial,"Final milestone review","All Phase B items reported honestly","Technical report covering D1-D5 delivered with unmet items labelled unmet","Final report + benchmark link + statement of accounts"SMART 自查（官方逐字要求 quantifiable + SMART）：每行都有数字（120+／25+／50+／30 天／1 家／51%／6 of 6）＋时点（M1–M12）＋二元判据（达成／未达成，无「部分成功」）＋可指认证据（最后一列）。⇒ 若实表另有 Target date 栏，按「录取日 + M(n)」换算成日期填写，不得写成相对月数了事。
#7. 繁體中文對照（內部核對件 · 不上傳 · 語義以英文正文為準）
§0 抬頭：項目＝Duty Runtime：受約束 LLM 任務閉環的完成判定、靜默升級與跨租戶洩漏防控（附公開基準 ChunYue-Agent-Bench）。領域＝人工智能與數據科學。年期＝12 個月。申請補助＝港幣 150,000 元，按每 6 個月里程碑評估分兩期等額發放（各 75,000 元）。申請身份＝團隊申請；獲錄取起兩星期內根據《公司條例》（第 622 章）註冊香港私人股份有限公司。
§1 十二個月里程碑（逐行對應上表，Category 為官方五類）
| 月 | 分類（官方五類） | 里程碑 | 二元驗收判據 |
|---|---|---|---|
| M1 | 產品與技術開發 | A1 以同方法學在凍結任務集上復測基線（內部生產實測為 45%） | 第二人取凍結輸入與腳本跑出同一數 ⇒ 達成 |
| M2 | 產品與技術開發 | A2 凍結任務集 v1：≥120 例，含清單內與越界兩類，附 schema 與評分器 | 入庫、可執行、全量跑通 ⇒ 達成 |
| M3 | 產品與技術開發 | A3 內容分層契約 v1：類型化動作＋前置條件＋風險等級＋允許話術 | 契約覆蓋試點全部動作類型；抽檢的越界樣本只產生升級義務、不產生 AI 動作 ⇒ 達成 |
| M4 | 產品與技術開發 | A4 靜默升級編排 v1：具名責任人＋期限＋接管上下文摘要，同一服務身份接續 | 升級件端到端可追溯（義務記錄 → 接管 → 最終解決） ⇒ 達成 |
| M5 | 產品與技術開發 | A5 對抗探針套件 v0：≥25 探針（提示注入／越權工具調用／上下文回填） | 套件可執行，且 2 例反證實驗（故意放開一條腿）當場報紅 ⇒ 達成 |
| M6 | 業務發展 | A6 試點 #1 上線（本階段允許關聯自營），PDPO 告知與會話日誌落地 | 連續 30 天真實會話日誌入庫且可抽樣展示 ⇒ 達成 |
| M6 閘 | 財務 | 期中里程碑評估——發放第二期港幣 75,000 元 | A1–A6 全部達成 ＋ 反映資金流轉的財務報表 |
| M7 | 產品與技術開發 | B1 在凍結集上跑指標族：實際完成／閉環／不合理升級／虛假承諾 | 每項給實測值＋樣本量＋日誌位置；未達成項如實報未達成並給原因 |
| M8 | 產品與技術開發 | B2 探針套件 v1：≥50 探針 ＋ 遮蔽表覆蓋斷言（以已核對口徑的作用域定義為準） | 洩漏 0 件；覆蓋按對口徑後的數 100%；反證腿仍必須失敗即封閉 |
| M9 | 其他 | B3 釋出 ChunYue-Agent-Bench：500 例 × 3 基線模型，Apache-2.0，附評分器與方法學 | 倉庫公開可達、許可證與 README 完整、複現指引照文可跑 |
| M10 | 產品與技術開發 | B4 複現實驗：第二操作者、第二台機器 | 同輸入同評分器跑出逐字節一致；不一致即如實報未達成 |
| M11 | 業務發展 | B5 非關聯付費試點 ≥1 家簽約；首響時延前後對比 | 合約或發票 ＋ 日誌雙證；每側 ≥30 例方可宣稱時延改善；未取得則報「未取得」並附在談清單 |
| M12 | 公司發展 | B6 升讀 IGNITE 準備：2 名在港全職研發、股權按承諾落位、進駐計劃排定 | 僱傭合約 ＋ 股權結構表 ＋ 進駐計劃三件齊 ⇒ 達成 |
| M12 閘 | 財務 | 結項里程碑評估 | B1–B6 如實申報；技術報告覆蓋 D1–D5 方法與結論 |
§2 指標族（全部標 target，同一凍結任務集、同一方法學）：實際完成率 基線 45%（2026-09 內部生產實測）→ ≥70%；閉環率 ≥95%（完成＋合理升級，且升級件 100% 被接管並最終解決——拆開計分，防「全升級」刷分）；不合理升級率 ≤5%；虛假承諾率 <0.5%（與後端日誌逐條比對）；跨租戶洩漏 0；錢與合同人工閘門 基線 100% → 100%（此類動作永不由 AI 完成）；首響時延 −30%（請求→首個實質動作的中位分鐘數；解決時長另行報告，不與首響混用；前後各 ≥30 例方可宣稱）。
§3 預算港幣 150,000 元（分科目綁里程碑）：在港研發工時 78,000（39,000／39,000）｜模型 API 與算力 24,000（8,000／16,000）｜評測語料與標注 18,000（12,000／6,000）｜法務與 PDPO 顧問 12,000（6,000／6,000）｜開源基準與文檔 10,000（2,000／8,000）｜園區進駐與評審差旅 8,000（2,000／6,000）⇒ 兩期各 75,000。人力行大白話口徑：78,000 ÷ 12 = 6,500/月，是在港工程工時的部分補貼；我方不聲稱15 萬養 2 名全職，M11 起的全職成本由公司自有資源承擔。
§4 交付件：期中（M6）＝復測報告 · 凍結任務集 v1＋評分器 · 分層契約 v1 · 升級編排 v1 · 探針套件 v0＋反證日誌 · 試點配置與 30 天日誌 · 財務報表。結項（M12）＝技術報告（D1–D5，未達成項照實寫）· 指標族結果與原始日誌位置 · 探針套件 v1 與覆蓋報告 · 公開基準倉庫 · 複現日誌 · 試點合約／發票與時延數據 · 下一階段所需的受僱與公司記錄。
§5 假設與風險：① 45%→≥70% 的假設若落空 ⇒ B1 報未達成，但凍結集、評分器與公開基準對第三方仍可用，負面結果本身即是可發表的產出；② 模型供應商改價或下架基線模型 ⇒ 已指定 3 個基線模型而非 1 個，且套件可在已錄輸出的結果上離線重跑；③ 窗口內無非關聯試點簽約（今日收入 HK$0）⇒ B5 報未達成，關聯自營試點仍產出 30 天日誌，外聯名單與承諾函模板已備；④ 跨租戶作用域口徑未對 ⇒ D3 的覆蓋主張無法量化，故在對齊前任何提交面都不出現表數；⑤ 主體未註冊 ⇒ 團隊通道正為此而設，錄取起兩週內註冊，並在一次註冊時把符合 IGNITE 口徑的股權落定。