Conrad Challenge Archive 2024 — 2026

Home首页· Guide指南·The topic engine选题引擎

The Conrad Topic Engine

A systematic, repeatable method for discovering competitive Conrad Challenge topics.

Grounded in: the 25 rubric sub-criteria (knowledge-base/01-competition-facts.md §7), the seven winning archetypes (knowledge-base/03-winner-patterns.md §3), and the failure modes observed in five real judged briefs (knowledge-base/02-judging-playbook.md).

Designed to be run by a student team with a coach, in one 3-hour workshop plus one week of verification. It is deliberately mechanical: the goal is that two different teams running it on the same interests produce different good topics, not the same obvious one.

Design principle

Most topic advice tells students to "find a problem you care about." That produces generic sustainability projects — the single most common failure mode.

This method inverts the order:

Don't start from a problem. Start from an asymmetry you can exploit, then find the problem that asymmetry unlocks.

The rubric rewards defensible differentiation (Innovation 30%, whose sub-criteria are Originality, Impact, Ambition, Verification, IP Defensibility). Differentiation cannot be added to a topic afterwards. It has to be the seed.

The pipeline

STAGE 0   Asset inventory          → what asymmetry do we actually have?         [45 min]
STAGE 1   Domain scan              → 12 candidate problem territories            [45 min]
STAGE 2   Archetype crossing       → 7 archetypes × territories = 30-80 seeds    [45 min]
STAGE 3   Kill screen              → cut to 8 on 4 hard gates                    [30 min]
STAGE 4   Originality verification → the judge's own search, run on yourself     [1 week]
STAGE 5   Rubric prescore          → score survivors on the real 25 sub-criteria [60 min]
STAGE 6   Category placement       → arbitrage decision                          [15 min]
STAGE 7   Falsification            → find the one reason it can't work           [2 hours]
                                     ↓
                                  1 topic

Expected yield: 30–80 seeds → 8 candidates → 3 survivors → 1 topic. If you end Stage 3 with fewer than 5, your Stage 1 territories were too narrow. Go back.

STAGE 0 — Asset inventory (45 min)

Before any ideas. Write down what this specific team actually has that a random team does not. Be concrete and unflattering.

Asset class Prompt Why it matters
Access What place, machine, dataset, organisation, or population can you get to that others can't? A family business, a farm, a clinic, a lab, a specific city, a language community. Enables Archetype G (hyper-specificity) and real Practicality evidence
Domain What does a team member know unusually well? A sport, an illness, a craft, a job, a hobby with equipment. Reveals real pain points others can't see
Skill Who can actually build? CAD, ML, electronics, video, sales, writing. Be honest. Sets what evidence rungs are reachable
Adult network Which experts can you email and get a reply from? Parents' colleagues, teachers, a doctor, a local engineer. Evidence ladder rung 2 costs one email
Lived constraint What has your family or community had to do without? The most under-used asset. Produces the specificity that defeats the originality search.
Budget & time Real numbers. Determines which archetypes are viable

Output: a list of 5–15 concrete assets.

The rule this stage enforces: every topic that survives to Stage 5 must trace to at least one named asset here. A topic with no asset behind it is a topic any of the other 1,500 teams could also submit — and probably did.

Worked example. A 2-student team: one has a beekeeping uncle (access + domain); one can train image models (skill); a family member had a stroke (lived constraint); a school 3D printer (skill); an aunt who is a nurse (network). Note that this exact asset profile is roughly what produced BeeGuard and BeCure — two independent 2026 finalists on the same underlying access.

STAGE 1 — Domain scan (45 min)

For each of the five categories, generate 2–3 problem territories — not solutions. A territory is a bounded population with a recurring, expensive, unglamorous problem.

Good territory tests:

  • Can you name who has it, how often, and what it costs them?
  • Is it boring enough that big companies ignore it? (StomaLock: ostomy care. Kipos: aging subway cars.)
  • Does someone currently pay money or spend labour on a bad workaround?

Prompts that work:

  1. The workaround hunt — where do people use tape, buckets, spreadsheets, or a person standing there watching?
  2. The price cliff — what capability exists only at institutional cost? (Feeds Archetype A.)
  3. The infrastructure assumption — what breaks when there's no power / no internet / no cold chain / it freezes? (Feeds B and G.)
  4. The installed base — what expensive old equipment is everywhere and can't be replaced? (Feeds C.)
  5. The sample-and-send — what still requires shipping a physical sample to a lab? (Feeds D.)
  6. The dignity gap — where is the existing solution functional but humiliating? (StomaLock's "discreet" alerts.)

Output: ~12 territories, each one line: who + what problem + current bad workaround.

STAGE 2 — Archetype crossing (45 min)

The generative step. Take a matrix: territories (rows) × the 7 archetypes (columns). Fill cells. Most will be nonsense — that's expected and fine.

A Price collapse B Zero-resource C Retrofit D Workflow E Data moat F Razor/blade G Hyper-specific
Territory 1
Territory 2

Cell prompt: "What would it mean to solve [territory] via [archetype]?"

Examples of the move:

  • Territory "hive health" × A → hive diagnosis at consumer price → what's the expensive incumbent? A bee inspector.
  • Territory "hive health" × B → hive diagnosis with zero added hardware → BeeGuard.
  • Territory "hive health" × D → replace physical inspection with behavioural analysis → BeCure.
  • Territory "hospital sterilisation" × C → clip onto existing autoclaves, change nothing → EcoClave.
  • Territory "drinking water safety" × G → freeze damage in high-altitude pastoral pipelines → Soaring Seeds.

Output: 30–80 one-line seeds. Do not filter yet.

Why this beats brainstorming: brainstorming converges on the most available idea (solar panel, water filter, recycling app). A forced matrix produces the cells nobody's mind travels to on its own, which is exactly where the originality search finds nothing.

STAGE 3 — Kill screen (30 min)

Four hard gates. Any failure kills the seed. Be ruthless; you want to reach 8.

Gate 1 — Buildable-to-evidence. Can you reach at least 3 rungs of the evidence ladder by the January deadline with your Stage 0 assets and budget? (Ladder: component precedent / expert testimony / literature / graphic model / prototype / described future experiment.) You do not need a prototype — you need three rungs.

Gate 2 — Nameable incumbent. Can you name, right now, a real product or method people currently use? If you cannot name what you're replacing, you cannot write Q7, and Innovation collapses. This is the Puppy WC gate.

Gate 3 — Conservation sanity. Does it violate energy or mass balance? Is any loop closed on itself? Is the "free" energy actually free? This is the P-Bump / Ocean Energy gate — and it is worth ~15 points.

Gate 4 — Moat statement. Can you finish this sentence with something other than a shrug? "A well-funded competitor can't just copy this because ______." Valid endings: a patentable mechanism, a trade secret process, an accumulating dataset, an exclusive relationship, a locked-in ecosystem, a cost structure they can't match without cannibalising their own product. Invalid: "we'll copyright it", "we'll work harder", "we'll be first."

Output: ~8 surviving candidates.

STAGE 4 — Originality verification (1 week, mostly waiting)

Run the judge's own required search against yourself, before they run it against you. Judges are instructed to "perform an online search to verify originality." The Verification sub-criterion scores exactly this. Do it first.

For each of the 8 candidates:

  1. Search as the industry, not as a student. Find the trade term. Not "sun-following solar panel" but "single-axis solar tracker"; not "pet toilet" but "automatic pet waste disposal system." Ask: what would a salesperson in this industry call it?
  2. Search commercial products, not just papers. Amazon, Alibaba, industry trade press, Crunchbase, Product Hunt. Puppy WC failed because the competitor was a consumer product, not a paper.
  3. Search patents. Google Patents / Espacenet for the mechanism. A near-identical granted patent is a kill; an adjacent one is a gift — it proves the field is real and shows you the claim language.
  4. Search prior Conrad entries. Check knowledge-base/data/winners.jsonl and the published finalist pages. A close match from last year is a strong negative — judges have seen it.
  5. Record everything. Every search feeds the references PDF and Q7. Nothing here is wasted.

Scoring the result:

Finding Action
Nothing similar exists at all ⚠️ Suspicious. Usually means the search terms are wrong, or there's a reason nobody does it. Search harder before celebrating.
Similar exists, yours differs on a named axis ✅ Ideal. This is the finalist position.
Similar exists, yours differs only in degree 🟡 Salvageable if the degree is large and quantified (Clarity: 40× cheaper)
Essentially identical product on the market ❌ Kill, or pivot to a specific unserved segment of it

Output: ~3–5 survivors with a documented competitive landscape. You have now written most of Q7.

STAGE 5 — Rubric prescore (60 min)

Score each survivor 1–5 against the real 25 sub-criteria, weighted as the judges weight them. Use the actual sub-criteria — this is a prediction of your score, not a vibe.

Theme Weight Sub-criteria to score
Innovation ×6 Originality · Impact Potential · Ambition · Verification · IP Defensibility
Storytelling ×4 Investor Appeal · Clarity of Opportunity · Credibility Boost · Narrative · Polish
Practicality ×4 Feasibility · Realism · Proof of Concept · Evidence Base · Next Steps
Marketing ×4 Market Insight · Entry & Adoption · Ecosystem · Differentiation · Channels
Finances ×2 Cost Estimation · Revenue · Funding Strategy · Budget · Viability

Theme score = mean of its 5 sub-criteria, rounded to a whole number (judges cannot award 3.5). Total = Innov×6 + Story×4 + Pract×4 + Mkt×4 + Fin×2.

At topic stage, score honestly for what you can reach by January, not what exists today.

Interpretation:

  • < 55 — kill it
  • 55–70 — viable but needs a specific fix; identify which sub-criterion is dragging
  • 70–85 — strong candidate
  • > 85 — either excellent or you are scoring yourself generously. Have the coach re-score blind.

Tiebreaker when two are close: pick the one with the higher Innovation score. Innovation is ×6 — double any other theme — and it is the hardest to improve later. Marketing and Finances can be fixed with a week of work in December. Originality cannot.

STAGE 6 — Category placement (15 min)

Deliberate, not automatic. For your chosen topic, ask:

  1. What is the obvious category? (Where will the most similar projects be?)
  2. What adjacent category has a defensible claim on it?
  3. In which category is your project surprising?

Then apply the category arbitrage rule:

Choose the category where your project is unusual but honestly placed. You compete against the ~300 teams in your category, not the 1,500 in the field.

Precedent from 2026 (all finalists): BeCure — bee detection → Health & Nutrition; Robo Therapy — medical exoskeleton → Cyber-Technology; Echo/Pulse — speech restoration → Cyber-Technology. And BeeGuard entered hive health under Energy & Environment and won Pete Conrad Scholar.

Also weigh the special category. The rotating special category (Water Challenge in 2024-25 and 2025-26) has the same ~6 finalist slots but no accumulated body of obvious projects. Best expected value per entrant. Check what the current cycle's special category is.

⚠️ Honesty limit: the fit must be genuine. A forced placement reads as confusion and hurts Clarity of Opportunity. If you can't write one sentence explaining the fit, take the obvious category.

STAGE 7 — Falsification (2 hours, before committing)

The last gate, and the one teams skip. You are looking for the reason your project fails.

  1. Find one domain adult — not your coach, not a business person. An engineer, clinician, or scientist in the actual field. Ask exactly one question:

    "Is there a reason this can't work?" Ocean Energy Dynamics and P-Bump would both have gained ~15 points from this single email.

  2. Run the energy/mass balance yourself. Write down where every joule and gram comes from and goes. If anything is "free," find whose it is.

  3. Steelman the incumbent. Why hasn't the obvious big company done this? There is always a reason. Options: (a) they can't — that's your moat, name it; (b) they won't — the market is too small for them, that's your moat, name it; (c) they have — go back to Stage 4.

  4. Write the one-sentence version and read it to someone outside the team. Use the finalist template from 03-winner-patterns.md §2. If they don't understand it in one pass, the problem is the topic, not the sentence.

If falsification kills the topic — that is a success. It costs 2 hours in November instead of $499 and ten weeks.

Output artifact

The engine produces one filled card. This becomes the Lean Canvas and the spine of the brief.

TOPIC CARD
──────────────────────────────────────────────────────
Innovation name:
One-sentence pitch:   [finalist template — artifact / user / mechanism /
                       quantified delta / named incumbent / moat]

Archetype(s):         A B C D E F G
Team asset it uses:   [must trace to Stage 0]
Category chosen:      ____________  (obvious was: ____________)
  Arbitrage rationale:

Named incumbents:     1.        2.        3.
Our defensible delta:
Moat sentence:        "A well-funded competitor can't copy this because..."

Evidence ladder plan (target ≥3 rungs by Jan):
  [ ] component precedent   [ ] expert testimony (who: ______)
  [ ] literature            [ ] graphic/CAD model
  [ ] prototype             [ ] described future experiment

Falsification result:  Expert consulted: ______  Verdict: ______
                       Energy/mass balance: OK / issue: ______
                       Why hasn't a big company done it: ______

Prescore:  Innov __×6  Story __×4  Pract __×4  Mkt __×4  Fin __×2  = ___/100
Kill-gate status:  Buildable ☐  Incumbent ☐  Conservation ☐  Moat ☐
──────────────────────────────────────────────────────

Anti-gaming note for coaches

This method can be misused: a coach could run Stages 1–7 and hand a student a finished topic card. That would violate the rule that coaches "may not create any of the submitted elements or provide excessive influence on the concept design."

Keep the method legitimate:

  • Stage 0 must be filled by the students — it is about their assets.
  • The coach runs the process (timekeeping, forcing the matrix to be filled, refusing weak moat sentences).
  • The coach must not supply cells in Stage 2 or verdicts in Stage 5.
  • The Stage 4 searches must be run by the students; they will be asked about them at the Summit.
  • The Stage 7 expert email must be written and sent by a student.

A useful test: at the Power Pitch, can the student explain why they rejected the other seven candidates? If not, the topic isn't theirs. See knowledge-base/04-mentoring-market.md.

Conrad 选题引擎

一套系统化、可重复的方法，用于发现具备竞争力的康莱德挑战赛选题。

依据：25 条评分子标准（knowledge-base/01-competition-facts.zh.md §7）、 从真实决赛队伍中挖掘出的七种获奖原型（knowledge-base/03-winner-patterns.zh.md §3）， 以及在五份真实已评审简述中观察到的失败模式（knowledge-base/02-judging-playbook.zh.md）。

设计目标：一支学生队伍在指导老师协助下，用一次 3 小时的工作坊加一周的核实即可跑完。 它刻意做得很机械：目标是让两支不同的队伍用同样的兴趣跑完之后，得到不同的好选题， 而不是同一个显而易见的选题。

设计原则

大多数选题建议告诉学生"找一个你关心的问题"。这产生的是 笼统的可持续性项目 —— 也就是最常见的单一失败模式。

本方法把顺序反了过来：

不要从问题出发。从你能利用的不对称性出发，再去找那个不对称性能打开的问题。

评分表奖励的是可防御的差异化（创新占 30%，其子标准为原创性、影响、野心、 验证、IP 可防御性）。差异化没法事后添加，它必须是种子。

流水线

阶段 0   资产盘点            → 我们究竟有什么不对称性？          [45 分钟]
阶段 1   领域扫描            → 12 个候选问题领地                 [45 分钟]
阶段 2   原型交叉            → 7 种原型 × 领地 = 30-80 个种子    [45 分钟]
阶段 3   淘汰筛选            → 4 道硬关卡，砍到 8 个             [30 分钟]
阶段 4   原创性核实          → 把评委的搜索先对自己跑一遍        [1 周]
阶段 5   评分表预打分        → 按真实的 25 条子标准给幸存者打分  [60 分钟]
阶段 6   赛道落位            → 套利决策                          [15 分钟]
阶段 7   证伪                → 找出它行不通的那一个理由          [2 小时]
                              ↓
                           1 个选题

预期产出：30–80 个种子 → 8 个候选 → 3 个幸存者 → 1 个选题。 如果阶段 3 结束时不足 5 个，说明阶段 1 的领地太窄了。回去重做。

阶段 0 —— 资产盘点（45 分钟）

在任何想法之前。写下这支队伍实际拥有、而随机一支队伍没有的东西。 要具体，而且不必好听。

资产类别 提问 为什么重要
通道 Access 你能接触到什么别人接触不到的场所、机器、数据集、组织或人群？家族企业、农场、诊所、实验室、某个特定城市、某个语言社群。 使原型 G（极端具体）和真实的可行性证据成为可能
领域 Domain 队里有谁对某件事懂得异常多？一项运动、一种疾病、一门手艺、一份工作、一个有器材的爱好。 揭示别人看不见的真实痛点
技能 Skill 谁真的能做出东西？CAD、机器学习、电子、视频、销售、写作。诚实一点。 决定哪些证据层级够得着
成人网络 Network 你发邮件后，哪些专家真的会回你？父母的同事、老师、一位医生、本地工程师。 证据阶梯第 2 级只需一封邮件
亲历的约束 Lived constraint 你的家庭或社区曾经不得不将就什么？ 最被低估的资产。 它产生的具体性能击败原创性搜索。
预算与时间 真实数字。 决定哪些原型可行

输出： 5–15 项具体资产。

本阶段执行的规则： 每一个能活到阶段 5 的选题，都必须能追溯到这里的某一项具名资产。 一个背后没有资产的选题，是另外 1,499 支队伍也能提交的选题 —— 而且很可能已经提交了。

实例。 一支两人队伍：一人有个养蜂的叔叔（通道+领域）；一人会训练图像模型（技能）； 一位家人中过风（亲历的约束）；学校有 3D 打印机（技能）；有个当护士的姨妈（网络）。 请注意，正是这套资产组合大致产生了 BeeGuard 和 BeCure —— 两支独立的 2026 决赛队伍，来自同一种通道。

阶段 1 —— 领域扫描（45 分钟）

对五个赛道各生成 2–3 个问题领地 —— 不是解决方案。 领地是一个有边界的人群，带着一个反复出现的、昂贵的、不体面的问题。

好领地的检验：

  • 你能说出谁有这个问题、多久一次、代价多少吗？
  • 它是否无聊到大公司懒得管？（StomaLock：造口护理。Kipos：老旧地铁车厢。）
  • 现在是否有人在为一个糟糕的变通方案花钱或花人力？

有效的提示：

  1. 变通方案搜寻 —— 哪里的人在用胶带、水桶、Excel 表格，或者派一个人站在那儿盯着？
  2. 价格悬崖 —— 哪些能力只在机构级成本下才存在？（喂给原型 A。）
  3. 基础设施假设 —— 没有电 / 没有网络 / 没有冷链 / 结冰时，什么会坏？（喂给 B 和 G。）
  4. 存量设备 —— 什么昂贵的老设备到处都是、又换不掉？（喂给 C。）
  5. 取样送检 —— 什么至今仍然需要把实物样本寄去实验室？（喂给 D。）
  6. 尊严缺口 —— 哪里的现有方案功能上可用、但让人难堪？（StomaLock 的"隐蔽提醒"。）

输出： 约 12 个领地，每个一行：谁 + 什么问题 + 现在的糟糕变通方案。

阶段 2 —— 原型交叉（45 分钟）

生成步骤。做一个矩阵：领地（行）× 七种原型（列）。 把格子填满。大多数会是废话 —— 这是预期之内，也没关系。

A 价格坍塌 B 零资源 C 存量改造 D 工作流 E 数据护城河 F 剃刀刀片 G 极端具体
领地 1
领地 2

单格提示："用【原型】来解决【领地】会意味着什么？"

这个动作的示例：

  • 领地"蜂群健康" × A → 用消费级价格做蜂群诊断 → 昂贵的在位者是谁？一位养蜂检查员。
  • 领地"蜂群健康" × B → 零额外硬件的蜂群诊断 → BeeGuard。
  • 领地"蜂群健康" × D → 用行为分析替代开箱检查 → BeCure。
  • 领地"医院灭菌" × C → 夹在现有高压灭菌器上，什么都不改 → EcoClave。
  • 领地"饮用水安全" × G → 高原牧区管道的冻损 → Soaring Seeds。

输出： 30–80 个一行种子。此时不要筛选。

为什么这比头脑风暴强： 头脑风暴会收敛到最容易想到的点子 （太阳能板、净水器、回收 App）。强制矩阵会产生那些你的大脑自己走不到的格子 —— 而那正是原创性搜索什么也搜不到的地方。

阶段 3 —— 淘汰筛选（30 分钟）

四道硬关卡。任一不过即淘汰。 要狠，目标是砍到 8 个。

关卡 1 —— 可达成证据。 在 1 月截止前，用阶段 0 的资产和预算，你能爬到证据阶梯的至少 3 级吗？ （阶梯：元器件先例 / 专家证言 / 文献 / 图形模型 / 原型 / 描述未来实验。） 你不需要原型 —— 你需要三级。

关卡 2 —— 可点名的在位者。 你现在，立刻，能说出一个人们今天在用的真实产品或方法吗？ 如果说不出来，你就写不了第 7 题，创新分随之崩塌。这是 Puppy WC 关卡。

关卡 3 —— 守恒常识。 它违反能量或质量守恒吗？有没有哪个回路是自我闭合的？那个"免费"的能量真的免费吗？ 这是 P-Bump / Ocean Energy 关卡 —— 价值约 15 分。

关卡 4 —— 护城河陈述。 你能不能用"耸肩"以外的东西把这句话补完？ "一个资金充足的竞争者没法直接抄走这个，因为 ______。" 有效的补法：一个可申请专利的机制、一道商业秘密工艺、一个持续积累的数据集、 一段独家关系、一个被锁定的生态、一种他们不自我蚕食就无法匹配的成本结构。 无效的："我们会申请著作权"、"我们会更努力"、"我们会是第一个"。

输出： 约 8 个幸存候选。

阶段 4 —— 原创性核实（1 周，大部分时间在等）

把评委被要求做的那个搜索，在他们对你做之前，先对自己做一遍。 评委被要求"执行在线搜索以核实原创性"。Verification 子标准打的就是这个。先做。

对 8 个候选逐一：

  1. 用行业的语言搜，不要用学生的语言。 找到那个行业术语。 不是"会跟着太阳转的太阳能板"，而是"单轴太阳跟踪器"； 不是"宠物马桶"，而是"自动宠物排泄物处理系统"。 问自己：这个行业里的销售会怎么称呼它？
  2. 搜商业产品，不只是论文。 Amazon、阿里巴巴、行业媒体、Crunchbase、Product Hunt。 Puppy WC 就是因为竞品是消费品而不是论文才失败的。
  3. 搜专利。 Google Patents / Espacenet，搜机制。 一份近乎相同的已授权专利是致命的；一份相邻专利是礼物 —— 它证明这个领域是真实的，还把权利要求的措辞白送给你。
  4. 搜往届 Conrad 参赛项目。 查 knowledge-base/data/winners.jsonl 和公开的决赛名单页。 与去年高度相似是强负面信号 —— 评委见过。
  5. 全部记录。 每一次搜索都会喂给参考文献 PDF 和第 7 题。这里没有一点功夫是白费的。

结果判读：

发现 行动
完全找不到类似的东西 ⚠️ 可疑。 通常意味着搜索词错了，或者存在一个"没人做"的理由。庆祝前先加倍搜索。
存在类似的，但你在一条具名的轴上不同 ✅ 理想。 这就是决赛位置。
只有程度差异 🟡 若程度很大且已量化，尚可挽救（Clarity：便宜 40 倍）
市面上有基本相同的产品 ❌ 淘汰，或转向该市场中一个未被服务的细分

输出： 3–5 个幸存者，附带一份有据可查的竞争格局。你已经把第 7 题写掉了大半。

阶段 5 —— 评分表预打分（60 分钟）

对每个幸存者，按真实的 25 条子标准打 1–5 分，并按评委的权重加权。 用真正的子标准 —— 这是对你分数的预测，不是感觉。

维度 权重 要打分的子标准
创新 ×6 原创性 · 影响潜力 · 野心 · 验证 · IP 可防御性
叙事 ×4 投资人吸引力 · 机会清晰度 · 可信度加成 · 叙事结构 · 打磨
可行性 ×4 可行性 · 现实性 · 概念验证 · 证据基础 · 下一步
市场 ×4 市场洞察 · 进入与采纳 · 生态 · 差异化 · 渠道
财务 ×2 成本估算 · 收入 · 融资策略 · 预算 · 可持续性

维度分 = 其 5 条子标准的平均值，四舍五入到整数（评委不能打 3.5）。 总分 = 创新×6 + 叙事×4 + 可行性×4 + 市场×4 + 财务×2。

在选题阶段，按你到 1 月能够得着的水平诚实打分，而不是按今天已有的水平。

判读：

  • < 55 —— 淘汰
  • 55–70 —— 可行，但需要一个具体的修补；找出是哪条子标准在拖后腿
  • 70–85 —— 强候选
  • > 85 —— 要么真的优秀，要么你在给自己放水。让指导老师盲打一遍。

当两个候选接近时的决胜规则：选创新分更高的那个。 创新是 ×6 —— 是任何其他维度的两倍 —— 而且它是最难后期补救的。 市场和财务可以在 12 月用一周补上。原创性不行。

阶段 6 —— 赛道落位（15 分钟）

要刻意，不要自动。对选定的选题问：

  1. 显而易见的赛道是哪个？（最相似的项目会在哪里？）
  2. 哪个相邻赛道对它有站得住的主张？
  3. 在哪个赛道里，你的项目是出人意料的？

然后应用赛道套利规则：

选那个让你的项目不寻常但落位诚实的赛道。 你竞争的是你赛道里的约 300 支队伍，不是全场 1,500 支。

2026 年的先例（全部为决赛队伍）：BeCure —— 蜂螨检测 → Health & Nutrition； Robo Therapy —— 医疗外骨骼 → Cyber-Technology；Echo/Pulse —— 语音恢复 → Cyber-Technology。而 BeeGuard 把蜂群健康报进 Energy & Environment， 拿下了 Pete Conrad Scholar。

也要权衡特别赛道。 轮换的特别赛道（近两届为水挑战）拥有同样的约 6 个决赛席位， 但没有累积下来的一堆显而易见的项目。每参赛者的期望值最好。 查清本届的特别赛道是什么。

⚠️ 诚实边界： 契合必须是真实的。牵强的落位读起来像是没想清楚， 会伤害"机会清晰度"。如果你写不出一句话解释这个契合，就选显而易见的那个赛道。

阶段 7 —— 证伪（2 小时，在投入之前）

最后一关，也是队伍最常跳过的一关。你要找的是你的项目失败的理由。

  1. 找一位领域内的成年人 —— 不是你的指导老师，不是商科的人。是这个真实领域里的 工程师、临床医生或科学家。只问一个问题：

    "有没有什么原因让这个方案行不通？" Ocean Energy Dynamics 和 P-Bump 各自都能从这一封邮件里拿回约 15 分。

  2. 自己算一遍能量/质量平衡。 写下每一焦耳、每一克从哪来、到哪去。 如果有什么是"免费的"，找出它是从谁那儿来的。

  3. 为在位者做最强辩护。 那个显而易见的大公司为什么还没做？总是有理由的。 可能是：(a) 他们做不到 —— 这就是你的护城河，点名它； (b) 他们不愿做 —— 市场对他们太小，这也是你的护城河，点名它； (c) 他们已经做了 —— 回到阶段 4。

  4. 写出一句话版本，读给队伍之外的人听。 用 03-winner-patterns.zh.md §2 的决赛模板。 如果对方一遍听不懂，问题出在选题上，不在句子上。

如果证伪杀死了这个选题 —— 那是成功。 它在 11 月花了你 2 小时， 而不是 499 美元加十周。

输出产物

引擎产出一张填好的卡片。它会成为 Lean Canvas，以及整份简述的骨架。

选题卡
──────────────────────────────────────────────────────
创新名称：
一句话陈述：  [决赛模板 —— 物件 / 用户 / 机制 /
               量化差值 / 具名在位者 / 护城河]

使用的原型：  A B C D E F G
依托的队伍资产：[必须追溯到阶段 0]
选定赛道：    ____________ （显而易见的是：____________）
  套利理由：

点名的在位者：1.        2.        3.
我们可防御的差值：
护城河句式：  "一个资金充足的竞争者没法抄走这个，因为……"

证据阶梯计划（1 月前目标 ≥3 级）：
  [ ] 元器件先例      [ ] 专家证言（谁：______）
  [ ] 文献            [ ] 图形/CAD 模型
  [ ] 原型            [ ] 描述未来实验

证伪结果：    咨询的专家：______  结论：______
              能量/质量平衡：通过 / 问题：______
              大公司为什么还没做：______

预打分：创新 __×6  叙事 __×4  可行 __×4  市场 __×4  财务 __×2 = ___/100
淘汰关卡状态：可达成证据 ☐  在位者 ☐  守恒 ☐  护城河 ☐
──────────────────────────────────────────────────────

给指导老师的防滥用提示

这套方法可能被误用：一位指导老师可以自己跑完阶段 1–7，然后把一张做好的选题卡交给学生。 那会违反"指导老师不得创作任何提交内容，也不得对概念设计施加过度影响"的规则。

保持方法的合法性：

  • 阶段 0 必须由学生自己填 —— 它讲的是他们的资产。
  • 指导老师负责跑流程（计时、逼着把矩阵填满、拒绝无力的护城河陈述）。
  • 指导老师不得在阶段 2 提供格子内容，也不得在阶段 5 给出结论。
  • 阶段 4 的搜索必须由学生执行；他们在峰会上会被问到。
  • 阶段 7 的专家邮件必须由学生撰写并发送。

一个有用的检验：在 Power Pitch 上，学生能不能解释他们为什么否掉了另外七个候选？ 答不出来，这个选题就不是他们的。见 knowledge-base/04-mentoring-market.zh.md。