「今年有 22% 的概率发生核爆」：这个价格作为证据值多少？"A 22% chance of a nuclear detonation this year": what is that price worth as evidence?
2026 年 3 月 4 日，Polymarket 下架「Nuclear weapon detonation by…?」，下架前全年合约报价约 22%。本页只用 Polymarket 公开接口的原始数据和一手来源，对这个数打五道折扣。不评价平台，也不评价下注的人。On 4 March 2026 Polymarket pulled "Nuclear weapon detonation by…?"; the full-year contract was quoting about 22%. This page applies five discounts to that number, using only raw data from Polymarket's public APIs and primary sources. No judgement of the platform or its traders.
一 · 五道折扣1 · Five discounts
| # | 问的问题QUESTION | 查到的事实WHAT THE DATA SHOWS | 对 22% 的影响EFFECT ON 22% | 判决VERDICT |
|---|---|---|---|---|
| 1 | 合约问的，是不是你以为的那个问题？Is the contract asking the question you think it is? | 规则原文：任何核爆都算 Yes，「includes offensive usages, nuclear tests, and accidental detonations」。1998 年以后全球每一次核爆都是朝鲜核试验（2006、2009、2013、2016×2、2017），2018 年起为零。另：接口显示三个 2026 合约已于 3 月 11 日按 No 结算，其中两个尚未到期。Rules: any detonation counts, "includes offensive usages, nuclear tests, and accidental detonations". Every detonation since 1998 has been a North Korean test (2006, 2009, 2013, 2016×2, 2017); none since 2018. The API also shows all three 2026 contracts settled No on 11 March, two of them before their deadline. | 这是「今年有没有核试验」的价格，不是核战概率。长期基准约 19%，近 8 年为 0A price on "any test this year", not on nuclear war. Long-run base rate ~19%, last 8 years 0 | not_comparable for "nuclear war" |
| 2 | 价格能直接读成概率吗？Can the price be read as a probability? | 买 No 要付 0.785 并锁到年底（约 0.83 年）。按 4% 资金成本，报价折回约 18.9%。同一平台的研究：10 美分以下的买入平均每美元亏 19.3 美分（低价合约偏贵）；政治类价格长期被压向 50%。Buying No costs 0.785 locked to year-end (~0.83y). At a 4% cost of capital the quote carries back to ~18.9%. Polymarket studies: purchases under 10¢ lose 19.3¢ per dollar (longshots overpriced); political prices compress toward 50%. | 方向向下；资金成本只解释约 2.6 个百分点，偏差大小无法精确给出Downward; carry explains only ~2.6pp, the bias size cannot be pinned down | band_only |
| 3 | 这类盘有没有足够的结算记录证明它准？Is there enough settlement history to show it is calibrated? | 此前 9 个同类合约（2022-10 → 2025-12）全部 No，日均报价 3%–11%。期限互相重叠，独立观测约 3 年。若真实年概率为 15%，连续 3 年全 No 的机会是 61%；要靠全 No 记录在 5% 水平上推翻 15%，需要 19 年。All 9 earlier contracts (2022-10 → 2025-12) resolved No, average daily quotes 3%–11%. Windows overlap: about 3 independent years. If the true annual rate were 15%, three straight No years happen 61% of the time; rejecting 15% at the 5% level on an all-No record takes 19 years. | 「过去都是 No」既不能说明它高估，也不能说明它准"Always No so far" shows neither overpricing nor accuracy | underpowered |
| 4 | 这个价格是多少人定的？How many people set this price? | 全年合约公开成交 1,610 笔、约 $80.6k、364 个钱包；前 5 个占 50.4%，前 10 个占 63%。按集中度（HHI = 0.063）折算，相当于约 16 个同等体量的独立钱包。报出 22% 的 3 月 4 日，当天成交约 $4.3k。The full-year contract: 1,610 public trades, ~$80.6k, 364 wallets; top 5 hold 50.4% of notional, top 10 hold 63%. Concentration (HHI = 0.063) is equivalent to about 16 equal-sized independent wallets. On 4 March, the day of the 22% quote, about $4.3k traded. | 「市场共识」实际是十几个钱包的共识The "crowd" is a dozen or so wallets | thin |
| 5 | 下注的人能不能左右结果？Can traders influence the outcome? | 普通交易者无法引发一次核爆，这一条基本不成立。可能存在的是知情者（例如提前知道试验时间），那会让价格更有信息量，而不是更像激励。No ordinary trader can cause a detonation, so this largely does not apply. Informed insiders (e.g. advance knowledge of a test) are possible; that makes the price more informative, not an incentive. | 不打折No discount | passes |
结论Verdict
22% 是一个约 16 个有效钱包、几千美元日成交给出的「今年会不会有核试验或核爆」的价格。读成概率要向下修，修多少不确定；它和过去 27 年的核试验频率（19%）大致相当。拿它当「核战概率」，题目就错了；拿它当「这个盘准不准」的证据，样本也远远不够。22% is a price on "any test or detonation this year", set by about 16 effective wallets trading a few thousand dollars a day. Read as a probability it should be revised down, by an uncertain amount; it sits close to the 27-year test frequency (19%). As a "probability of nuclear war" it answers the wrong question; as evidence the market is calibrated, the record is far too short.
二 · 此前 9 个已到期的同类合约2 · Nine earlier contracts of the same kind
| 合约CONTRACT | 交易区间TRADED | 日均报价MEAN DAILY QUOTE | 成交额（平台口径）VOLUME (PLATFORM) | 结算RESOLVED |
|---|---|---|---|---|
| before 2023 | 2022-10-31 → 2022-12-31 | — | $33,205 | No |
| by June 30, 2023 | 2023-01-31 → 2023-06-30 | 5.0% | $427,620 | No |
| by Dec 31, 2023 | 2023-05-23 → 2023-12-31 | 7.9% | $179,951 | No |
| by June 30, 2024 | 2023-12-29 → 2024-06-30 | 11.4% | $90,601 | No |
| in 2024 (Jul–Dec) | 2024-07-02 → 2024-12-31 | 9.7% | $4,742,982 | No |
| by Mar 31, 2025 | 2024-12-03 → 2025-03-31 | 5.5% | $827,456 | No |
| by June 30, 2025 | 2024-12-03 → 2025-06-30 | 7.0% | $883,595 | No |
| before Oct 2025 | 2025-07-03 → 2025-09-30 | 3.1% | $224,981 | No |
| in 2025 | 2024-12-30 → 2025-12-31 | 11.0% | $2,903,760 | No |
日均报价为 Yes 合约日线收盘的简单平均，期限各不相同，不能横向直接比较。成交额为平台接口的 volume 字段，与第一节用公开成交记录（数量 × 价格）算出的口径不同。Mean daily quote is the simple average of daily Yes closes; horizons differ, so rows are not directly comparable. Volume is the platform API's volume field, a different measure from the size × price sum used in section 1.
三 · 换成 Evidence Deflator 的输入3 · As Evidence Deflator inputs
这五道折扣对应一组由调用方提供、工具不抓取的输入（草案，尚未实现）：The five discounts map to inputs supplied by the caller, never fetched by the tool (draft, not implemented):
resolution_text+your_question— 合约规则原文和你真正要问的问题，逐条对照contract rules and the question you actually mean, compared clause by clauseprice,expiry,cost_of_capital— 折回带资金成本的区间carry-adjusted bandresolved_comparables— 同类已结算合约数和独立年数count of resolved comparable contracts and independent periodswallet_notional— 按钱包的成交额分布，用于算有效钱包数notional by wallet, for the effective-wallet countact_sensitive+source— 能否被交易者左右，并写明依据whether traders can move the outcome, with the basis stated
前人工作Prior work
Decomposing Crowd Wisdom（3.53 亿笔交易）：「a price's meaning depends on what, when and how much is traded」。The Favorite–Longshot Bias … Polymarket（5.88 亿笔）发现：按合约等权时低价合约每美元亏 6.3 美分，先按母事件归组则变成赚 4.1 美分——相关合约是否算成独立样本，会翻转结论，这正是第三、四道折扣的问题。Chaos Contracts（Jonker & Victoria, 2026）给出第五道折扣的判据：交易能否改变结果。本页的增量更窄：把这些判据逐条套到一个具体价格上，每个数都可以从公开接口复现。Decomposing Crowd Wisdom (353M trades): "a price's meaning depends on what, when and how much is traded". The Favorite–Longshot Bias … Polymarket (588M trades): longshots lose 6.3¢ per dollar when contracts are weighted equally but gain 4.1¢ once related contracts are grouped by parent event — whether related contracts count as independent flips the answer, the issue behind discounts 3 and 4. Chaos Contracts (Jonker & Victoria, 2026) supplies the test for discount 5. What this page adds is narrower: the criteria applied one by one to a single price, every number reproducible from public endpoints.
做法与限定Method and limits
- 数据于 2026-10-06 取自 Polymarket 公开接口：
gamma-api（规则、结算、平台成交额）、clob /prices-history（日线）、data-api /trades（逐笔成交）。核试验年份取自 Arms Control Association 核试验统计表。Data pulled 2026-10-06 from Polymarket's public endpoints:gamma-api(rules, settlement, platform volume),clob /prices-history(daily),data-api /trades(fills). Test years from the Arms Control Association tally. - 媒体报道的 $65 万–$83.8 万是整个事件（三个期限合计）按平台口径的成交额，本页第一节只统计全年合约的公开成交记录，两者不可直接比较。中文报道中的「24%」在日线数据里没有找到，日线最高为 0.215（不含上市首日）。The $650k–$838k reported by media is event-level platform volume across three deadlines; section 1 counts only the full-year contract's public fills, so the two are not comparable. The "24%" in some Chinese reports does not appear in the daily data; the daily high was 0.215 (excluding listing day).
- 「3 月 11 日按 No 结算」是接口字段（
outcomePrices = [0, 1]，closedTime）显示的状态；平台是否另有退款安排，本页无法核实。"Settled No on 11 March" is what the API fields show (outcomePrices = [0, 1],closedTime); whether the platform made other arrangements for holders cannot be checked here. - 4% 资金成本为假设值。低价合约偏差的数字来自全平台研究，这个盘属于哪一类、偏差多大，没有单独估计。有效钱包数按成交额而非持仓计算。The 4% cost of capital is an assumption. Longshot-bias figures are platform-wide; this market's category and bias were not estimated separately. Effective wallets are computed on traded notional, not positions.
- 一人审计，欢迎指正。One auditor; corrections welcome.
Evidence Deflator · GitHub · LLM 交易论文公开审计LLM trading paper audit