{"id":"cc789865-8b7b-4849-b53c-6a18773b9b26","entityType":"agent","slug":"clawhub-deciqai-survivorship-bias","name":"Survivorship Bias","canonicalUrl":"https://www.xpersona.co/agent/clawhub-deciqai-survivorship-bias","canonicalPath":"/agent/clawhub-deciqai-survivorship-bias","generatedAt":"2026-10-11T21:00:24.473Z","source":"CLAWHUB","claimStatus":"UNCLAIMED","verificationTier":"NONE","summary":{"evidence":{"source":"editorial-content","verified":true,"confidence":"high","updatedAt":"2026-10-11T15:57:39.042Z","emptyReason":null},"description":"Activate when: user says 'look at what winners/billionaires/champions did,' investment returns or fund performance are being cited, a strategy is justified b... Skill: Survivorship Bias Owner: deciqai Summary: Activate when: user says 'look at what winners/billionaires/champions did,' investment returns or fund performance are being cited, a strategy is justified b... Tags: latest:1.0.5 Version history: v1.0.5 | 2026-07-16T18:17:47.852Z | user Description tail link + agents machine-readable metadata line (deciqai.com/s/survivorship-bias.json) v1.0.4 | 2026-07-08T11:21:32.798","descriptionLabel":"Technical summary","evidenceSummary":"Capability contract not published. No trust telemetry is available yet. 1K downloads reported by the source. Last updated 10/11/2026.","installCommand":"clawhub skill install s17a4mqcnk515kvaca5ze55d0x88pfpx:survivorship-bias","sourceUrl":"https://clawhub.ai/deciqai/survivorship-bias","homepage":"https://clawhub.ai/deciqai/skills/survivorship-bias","primaryLinks":[{"label":"View on ClawHub","url":"https://clawhub.ai/deciqai/survivorship-bias","kind":"source"},{"label":"Homepage","url":"https://clawhub.ai/deciqai/skills/survivorship-bias","kind":"homepage"}],"safetyScore":84,"overallRank":62,"popularityScore":60,"trustScore":null,"claimedByName":null,"isOwner":false,"seoDescription":"Activate when: user says 'look at what winners/billionaires/champions did,' investment returns or fund performance are being cited, a strategy is justified b..."},"coverage":{"evidence":{"source":"public-profile","verified":false,"confidence":"medium","updatedAt":"2026-10-11T15:57:39.042Z","emptyReason":null},"protocols":[{"protocol":"OPENCLEW","label":"OpenClaw","status":"self-declared","notes":"Declared in the public agent profile."}],"capabilities":[],"verifiedCount":0,"selfDeclaredCount":1,"capabilityMatrix":{"rows":[{"key":"OPENCLEW","type":"protocol","support":"unknown","confidenceSource":"profile","notes":"Listed on profile"}],"flattenedTokens":"protocol:OPENCLEW|unknown|profile"}},"adoption":{"evidence":{"source":"CLAWHUB","verified":false,"confidence":"medium","updatedAt":"2026-10-11T15:57:39.042Z","emptyReason":null},"stars":null,"forks":null,"downloads":1033,"likes":null,"task":null,"library":null,"packageName":null,"latestVersion":"1.0.5","tractionLabel":"1K downloads"},"release":{"evidence":{"source":"CLAWHUB","verified":false,"confidence":"medium","updatedAt":"2026-10-11T15:57:38.887Z","emptyReason":null},"lastUpdatedAt":"2026-10-11T15:57:39.042Z","lastCrawledAt":"2026-10-11T15:57:38.887Z","lastIndexedAt":null,"nextCrawlAt":"2026-10-12T15:57:38.887Z","lastVerifiedAt":null,"highlights":[{"version":"1.0.5","createdAt":"2026-07-16T18:17:47.852Z","changelog":"Description tail link + agents machine-readable metadata line (deciqai.com/s/survivorship-bias.json)","fileCount":7,"zipByteSize":16375},{"version":"1.0.4","createdAt":"2026-07-08T11:21:32.798Z","changelog":"Footer now uses /c/<slug> short link (fixes UTM truncation when SKILL.md is read in a terminal)","fileCount":7,"zipByteSize":16153},{"version":"1.0.3","createdAt":"2026-07-08T03:51:41.889Z","changelog":"Second primary-sourced worked example","fileCount":6,"zipByteSize":11806},{"version":"1.0.2","createdAt":"2026-07-08T01:06:20.235Z","changelog":"Refreshed content + GitHub star link in footer","fileCount":5,"zipByteSize":9471},{"version":"1.0.1","createdAt":"2026-07-07T22:34:07.346Z","changelog":"Add catalog categories and topics","fileCount":5,"zipByteSize":9250},{"version":"1.0.0","createdAt":"2026-07-03T07:17:32.833Z","changelog":"Initial publish","fileCount":5,"zipByteSize":9274}]},"execution":{"evidence":{"source":"CLAWHUB","verified":false,"confidence":"low","updatedAt":null,"emptyReason":"No published capability contract is available yet."},"installCommand":"clawhub skill install s17a4mqcnk515kvaca5ze55d0x88pfpx:survivorship-bias","setupComplexity":"low","setupSteps":["Setup complexity is classified as HIGH. You must provision dedicated cloud infrastructure or an isolated VM. Do not run this directly on your local workstation.","Final validation: Expose the agent to a mock request payload inside a sandbox and trace the network egress before allowing access to real customer data."],"contract":{"contractStatus":"missing","authModes":[],"requires":[],"forbidden":[],"supportsMcp":false,"supportsA2a":false,"supportsStreaming":false,"inputSchemaRef":null,"outputSchemaRef":null,"dataRegion":null,"contractUpdatedAt":null,"sourceUpdatedAt":null,"freshnessSeconds":null},"invocationGuide":{"preferredApi":{"snapshotUrl":"https://www.xpersona.co/api/v1/agents/clawhub-deciqai-survivorship-bias/snapshot","contractUrl":"https://www.xpersona.co/api/v1/agents/clawhub-deciqai-survivorship-bias/contract","trustUrl":"https://www.xpersona.co/api/v1/agents/clawhub-deciqai-survivorship-bias/trust"},"curlExamples":["curl -s \"https://www.xpersona.co/api/v1/agents/clawhub-deciqai-survivorship-bias/snapshot\"","curl -s \"https://www.xpersona.co/api/v1/agents/clawhub-deciqai-survivorship-bias/contract\"","curl -s \"https://www.xpersona.co/api/v1/agents/clawhub-deciqai-survivorship-bias/trust\""],"jsonRequestTemplate":{"query":"summarize this repo","constraints":{"maxLatencyMs":2000,"protocolPreference":["OPENCLEW"]}},"jsonResponseTemplate":{"ok":true,"result":{"summary":"...","confidence":0.9},"meta":{"source":"CLAWHUB","generatedAt":"2026-10-11T21:00:24.469Z"}},"retryPolicy":{"maxAttempts":3,"backoffMs":[500,1500,3500],"retryableConditions":["HTTP_429","HTTP_503","NETWORK_TIMEOUT"]}},"endpoints":{"dossierUrl":"https://www.xpersona.co/api/v1/agents/clawhub-deciqai-survivorship-bias/dossier","snapshotUrl":"https://www.xpersona.co/api/v1/agents/clawhub-deciqai-survivorship-bias/snapshot","contractUrl":"https://www.xpersona.co/api/v1/agents/clawhub-deciqai-survivorship-bias/contract","trustUrl":"https://www.xpersona.co/api/v1/agents/clawhub-deciqai-survivorship-bias/trust"}},"reliability":{"evidence":{"source":"runtime-metrics","verified":false,"confidence":"low","updatedAt":null,"emptyReason":"No trust, reliability, or runtime telemetry is available."},"trust":{"status":"unavailable","handshakeStatus":"UNKNOWN","verificationFreshnessHours":null,"reputationScore":null,"p95LatencyMs":null,"successRate30d":null,"fallbackRate":null,"attempts30d":null,"trustUpdatedAt":null,"trustConfidence":"unknown","sourceUpdatedAt":null,"freshnessSeconds":null},"decisionGuardrails":{"doNotUseIf":["Contract metadata is missing or unavailable for deterministic execution."],"safeUseWhen":[],"riskFlags":["missing_or_unavailable_contract","trust_data_unavailable","schema_references_missing"],"operationalConfidence":"low"},"executionMetrics":{"observedLatencyMsP50":null,"observedLatencyMsP95":null,"estimatedCostUsd":null,"uptime30d":null,"rateLimitRpm":null,"rateLimitBurst":null,"lastVerifiedAt":null,"verificationSource":null},"runtimeMetrics":{"successRate":null,"avgLatencyMs":null,"avgCostUsd":null,"hallucinationRate":null,"retryRate":null,"disputeRate":null,"p50Latency":null,"p95Latency":null,"lastUpdated":null}},"benchmarks":{"evidence":{"source":"no-benchmark-data","verified":false,"confidence":"low","updatedAt":null,"emptyReason":"No benchmark suites or observed failure patterns are available."},"suites":[],"failurePatterns":[]},"artifacts":{"evidence":{"source":"CLAWHUB","verified":false,"confidence":"high","updatedAt":"2026-10-11T15:57:39.042Z","emptyReason":null},"readme":"Skill: Survivorship Bias\n\nOwner: deciqai\n\nSummary: Activate when: user says 'look at what winners/billionaires/champions did,' investment returns or fund performance are being cited, a strategy is justified b...\n\nTags: latest:1.0.5\n\nVersion history:\n\nv1.0.5 | 2026-07-16T18:17:47.852Z | user\n\nDescription tail link + agents machine-readable metadata line (deciqai.com/s/survivorship-bias.json)\n\nv1.0.4 | 2026-07-08T11:21:32.798Z | user\n\nFooter now uses /c/<slug> short link (fixes UTM truncation when SKILL.md is read in a terminal)\n\nv1.0.3 | 2026-07-08T03:51:41.889Z | user\n\nSecond primary-sourced worked example\n\nv1.0.2 | 2026-07-08T01:06:20.235Z | user\n\nRefreshed content + GitHub star link in footer\n\nv1.0.1 | 2026-07-07T22:34:07.346Z | user\n\nAdd catalog categories and topics\n\nv1.0.0 | 2026-07-03T07:17:32.833Z | user\n\nInitial publish\n\nArchive index:\n\nArchive v1.0.5: 7 files, 16375 bytes\n\nFiles: examples/abraham-wald-and-the-statistical-research-group-1943.md (7443b), examples/ai-startup-survivorship-2023-2026.md (8178b), examples/mutual-fund-survivorship-and-reported-returns-1996.md (4526b), references/sources.md (2398b), skill-card.md (2641b), SKILL.md (7249b), _meta.json (136b)\n\nFile v1.0.5:SKILL.md\n\n---\nname: survivorship-bias\ndescription: \"Activate when: user says 'look at what winners/billionaires/champions did,' investment returns or fund performance are being cited, a strategy is justified by pointing to companies that succeeded, historical data is treated as representative of all cases, or someone says 'this works because X did it.'\n  Do NOT activate when: population data is available and already selection-corrected; analysis is explicitly about survivors only with no claim about the broader population. More: deciqai.com/c/survivorship-bias\"\n---\n\n# Survivorship Bias\n\n## Overview\n\n**Survivorship bias** is drawing conclusions from a sample pre-filtered by survival — treating survivor traits as the *cause* of survival when non-survivors (absent from data by definition) may have had identical traits and still failed.\n\nCanon: Wald (1943) reversed the Navy's bomber-armor recommendation — returning planes showed damage where hits were survivable; the missing planes (shot down) were hit where returning planes showed no damage. Armor the gaps, not the hits.\n\nComposes with `bayesian-reasoning` (prior = population, not survivors), `critical-thinking` (what would non-survivors say?), `first-principles` (population is bedrock), and `abductive-reasoning` (\"winners have trait Y\" is one hypothesis; randomness is another).\n\n## When to Use\n\n- Someone draws lessons from \"what successful X did\"\n- Investment returns / fund performance / backtested strategies are cited\n- A business strategy is justified by pointing to companies that used it\n- Medical / treatment success rates are reported without dropout data\n- Career advice comes from what top performers did\n- Odds of building an AI startup are inferred from the visible AI winners (funded unicorns, \"wrapper\" success stories) amid the AI-bubble / AI-capex debate\n\n**Not when:** population data available and filter already corrected; analysis is intentionally about survivors only with no population claim.\n\n## Coaching Novices (Adaptive Front Door)\n\n- **Engine mode:** user has a concrete claim and data → run The Process directly.\n- **Coach mode:** unfamiliar or no case → guide step by step.\n\nIn Coach mode, respond one step at a time. Each [WAIT] is a hard stop — output only that step's question, then stop.\n\n1. One-liner: \"Before believing 'X worked because winners did X,' ask whether the losers did the same X — they're not in your sample to refute it.\"\n2. Check fit: if sample is explicitly restricted to survivors with no population claim, this lens doesn't apply.\n3. Elicit their real claim and the visible data they have.\n> **[WAIT — do not advance until user responds]**\n4. Run The Process one step at a time: what filter produced this sample? what's missing? if missing data looked like the sample, would the conclusion hold?\n> **[WAIT — do not advance until user responds]**\n5. Close by naming the selection-corrected conclusion (or marking it unprovable from this data alone).\n> **[WAIT — do not advance until user responds]**\n\n## The Process\n\n**Step 1 — State the claim:** What is being concluded, from what sample, from what source?\n\n**Step 2 — Identify the survival filter:** What process produced this sample? What was the population before the filter? What fraction was removed? What did the filter select for/against?\n\n**Step 3 — Construct the non-survivor hypothesis:** What did non-survivors likely have? Did they share the trait attributed to success? Would the claim hold if we could see them?\n\n**Step 4 — Re-estimate strength:** Best case = trait explains survival (non-survivors lacked it). Worst case = trait doesn't explain survival (non-survivors had it too). What evidence distinguishes these?\n\n**Step 5 — Correct or mark:** Get population data and re-run analysis with selection correction. If unavailable, mark conclusion as conditional on survivor sample.\n\n## Output Template\n\n```markdown\n# Survivorship Bias Analysis: <claim>\nClaim / sample / source:\nSurvival filter (what removed non-survivors, population size est., survival rate est.):\nNon-survivor hypothesis (what they likely had/lacked, could they have had same trait):\nCorrected inference (conclusion, confidence, what data would settle it):\n```\n\n*→ Method in Action: [Abraham Wald and the Statistical Research Group, 1943](examples/abraham-wald-and-the-statistical-research-group-1943.md) · [Mutual Fund Survivorship and Reported Returns, 1971–1996](examples/mutual-fund-survivorship-and-reported-returns-1996.md)*\n*→ 2026 lens: [AI-startup survivorship — funded unicorns vs the dead-wrapper graveyard (2023–2026)](examples/ai-startup-survivorship-2023-2026.md)*\n\n## Pack: Common Survivor Patterns\n\n| Domain | Survivor sample | Missing non-survivor data | Biased claim |\n|---|---|---|---|\n| Business / startup | Surviving companies | Failed companies | \"Successful companies do X\" |\n| Investment returns | Active funds / listed stocks | Closed funds / delisted stocks | \"Stocks return 10% annually\" |\n| Career advice | Top performers | People who left the field | \"To succeed, do X\" |\n| Scientific findings | Published studies | Unpublished null results | \"X is significant\" |\n| Treatment efficacy | Patients who completed | Drop-outs, deaths during treatment | \"X% recovered\" |\n\n*→ Primary sources: [references/sources.md](references/sources.md)*\n\n## Common Rationalizations\n\n**[D] = designed upfront | [O] = observed in real use. [O] entries are more valuable.**\n\n| Fake move | Reality |\n|---|---|\n| [D] \"Look at the data\" (survivor sample) | Survivor data ≠ population data. Correct or mark as conditional. |\n| [D] Citing one famous example as proof | N=1 in survivor sample tells you nothing about the rate. |\n| [D] \"X is the formula for success\" | If failures did the same X, X is not the formula. Get non-survivor data or stop claiming. |\n| [D] \"We use a backtested strategy\" | If backtest excludes failed/delisted stocks, results are upward-biased. |\n| [D] \"If we had non-survivor data, we'd see the same pattern\" | Unfalsifiable without the data. Get it or hold the claim. |\n| *→ Add [O] entries here after each real use — paste the actual failure pattern* | *What went wrong and why* |\n\n## Red Flags\n\n- Sample described as \"successful X\" or \"the X who made it\"\n- Data source is survivor-filtered (active funds, surviving companies, published studies)\n- Base rate of failure / dropout not stated\n- Conclusions about a population drawn from the survivor subset\n\n## Verification\n\n- [ ] Survival filter identified\n- [ ] Non-survivor population size estimated\n- [ ] Non-survivor hypothesis constructed\n- [ ] Conclusions conditional on survivor sample, or formally selection-corrected\n- [ ] Recommendation robust to worst-case non-survivor hypothesis\n\n---\n\n*Part of **deciqAI Knowledge Skills** — 227 open-source thinking skills that make rigor executable for AI agents. The same skills power every deciqAI agent, which runs them autonomously to operate your company. **See it run → https://www.deciqai.com/c/survivorship-bias** · ⭐ Star the repo → https://github.com/deciqAI/knowledge-skills · Contributions welcome.*\n\n*Agents: latest version & machine-readable metadata → https://www.deciqai.com/s/survivorship-bias.json*\n\nFile v1.0.5:_meta.json\n\n{\n  \"ownerId\": \"kn754b8sk22s8c6gjxt02bftbn88q7ye\",\n  \"slug\": \"survivorship-bias\",\n  \"version\": \"1.0.5\",\n  \"publishedAt\": 1784225867852\n}\n\nFile v1.0.5:references/sources.md\n\n# Sources — survivorship-bias\n\n> *Primary sources for the [survivorship-bias](../SKILL.md) skill.*\n\n- Wald, A. (1943). \"A Method of Estimating Plane Vulnerability Based on Damage of Survivors.\" SRG Memorandum CRC 432, Statistical Research Group, Columbia University. Republished 1980 by Center for Naval Analyses.\n- Heckman, J. J. (1979). \"Sample selection bias as a specification error.\" *Econometrica*, 47(1), 153-161. The Nobel-laureate selection-correction formalization.\n- Brown, S. J., Goetzmann, W., Ibbotson, R. G., & Ross, S. A. (1992). \"Survivorship Bias in Performance Studies.\" *Review of Financial Studies*, 5(4), 553-580. The financial-economics demonstration.\n- Elton, E. J., Gruber, M. J., & Blake, C. R. (1996). \"Survivorship Bias and Mutual Fund Performance.\" *Review of Financial Studies*, 9(4), 1097-1120. The follow-forward mutual fund correction: bias grows with sample length.\n- Malkiel, B. G. (1995). \"Returns from Investing in Equity Mutual Funds 1971 to 1991.\" *Journal of Finance*, 50(2), 549-572. Full-population reconstruction of equity fund returns including dead funds.\n- Mangel, M. & Samaniego, F. J. (1984). \"Abraham Wald's Work on Aircraft Survivability.\" *Journal of the American Statistical Association*, 79(386), 259-267. The post-declassification academic treatment.\n- Thiel, P. & Masters, B. (2014). *Zero to One: Notes on Startups.* Crown Business. ISBN 978-0804139298. The startup-context articulation.\n- Taleb, N. N. (2001). *Fooled by Randomness: The Hidden Role of Chance in Life and in the Markets.* Random House. ISBN 978-0812975215. Sustained treatment of survivorship bias in finance.\n- Ioannidis, J. P. A. (2005). \"Why Most Published Research Findings Are False.\" *PLoS Medicine*, 2(8), e124. Publication bias as survivor bias at the meta-research level.\n- U.S. Bureau of Labor Statistics, Business Employment Dynamics — Entrepreneurship and the U.S. Economy / establishment survival rates. https://www.bls.gov/bdm/ . Baseline population failure rate for new U.S. businesses (roughly half survive five years) — the denominator that AI-startup survivor claims omit.\n- Reuters / Associated Press (January 27–28, 2025). Reporting on the DeepSeek-triggered AI-stock sell-off, including Nvidia's record single-day market-value decline (widely reported at approximately $600 billion). Public market repricing of the thin-moat \"AI wrapper\" thesis.\n\nFile v1.0.5:examples/abraham-wald-and-the-statistical-research-group-1943.md\n\n# Method in Action: Abraham Wald and the Statistical Research Group, 1943\n\n> *Example for the [survivorship-bias](../SKILL.md) skill.*\n\nThe canonical demonstration of survivorship bias is **Abraham Wald**'s wartime work at the **Statistical Research Group (SRG)** at Columbia University, 1942-1945. The SRG was the secret mathematical-statistics arm of the wartime US government, comparable in caliber to the Manhattan Project for statistics: it included Wald, Allen Wallis, Milton Friedman, Frederick Mosteller, and Jacob Wolfowitz, among others.\n\nThe bomber-armor problem was assigned to the SRG in 1943. The US Navy was losing aircraft over Europe and the Pacific in unsustainable numbers. The problem was to determine where to add additional armor on B-17 and B-24 bombers — adding armor everywhere would make the plane too heavy to fly its mission, so the armor had to be concentrated where it would do the most good. The Navy collected data on bullet holes and damage patterns in aircraft that returned from missions.\n\nThe pattern looked clear. Damage was concentrated in specific areas:\n\n- **Wings:** many holes\n- **Fuselage center:** many holes\n- **Tail section:** many holes\n- **Engine area:** few holes\n- **Cockpit / pilot area:** few holes\n\nThe Navy's analyst recommended adding armor to the areas with the most damage — the wings, fuselage, and tail. Wald reviewed the data and recommended the *opposite*.\n\nHis reasoning, preserved in the SRG memorandum CRC 432, July 1943 (declassified 1980), opens with what may be the most famous selection-effect argument in 20th-century applied statistics:\n\n> \"The bullet holes are concentrated where the armor is not needed. The aircraft that we observe — the ones that returned — have damage on the wings, fuselage, and tail. The aircraft that did not return are not in our sample. They were shot down because they were hit where we observe little damage in the returnees: the engines and the cockpit. The damage we observe is the damage that planes can sustain and still fly home. The missing damage — the gaps in our pattern — is the lethal damage. Armor must go where the returners show no holes.\"\n>\n> — Wald, A. (1943). \"A Method of Estimating Plane Vulnerability Based on Damage of Survivors.\" SRG Memorandum CRC 432, Statistical Research Group, Columbia University. Republished in 1980 as CNA Memorandum.\n\nThe recommendation was implemented. Modeled losses dropped substantially. Wald's analysis became foundational to the **selection-effects literature** in statistics, econometrics, and epidemiology.\n\nThe deeper lesson Wald drew: **the observed sample is conditional on survival**, and the conditional distribution can be the opposite of the unconditional distribution. The damage pattern in returning aircraft is *anti-correlated* with the damage pattern in shot-down aircraft, because the very planes that absorbed the most lethal hits are absent from the data.\n\nWald's framework was systematized. The post-war statistics literature developed:\n\n**Heckman selection correction (1979).** James Heckman, building on Wald's logic, developed the formal mathematical framework for selection-corrected estimation. The \"Heckman correction\" is the standard treatment in modern econometrics; it won Heckman the 2000 Nobel Prize in Economics. See Heckman, J. J. (1979). \"Sample selection bias as a specification error.\" *Econometrica*, 47(1), 153-161.\n\n**Publication bias in scientific research.** Studies showing positive effects are more likely to be published than studies showing null effects. Ioannidis (2005) \"Why most published research findings are false,\" *PLoS Medicine*, 2(8), e124, drew on this. Survivorship bias at the meta-research level: the published literature is the survivor sample of all conducted research.\n\n**Funds and indices in finance.** Brown, Goetzmann, Ibbotson, and Ross (1992), \"Survivorship Bias in Performance Studies,\" *Review of Financial Studies*, 5(4), 553-580, demonstrated that mutual fund returns reported in standard databases were systematically biased upward because failed funds had been dropped. The corrected estimates were 1-3 percentage points lower per year — a massive effect compounded over decades.\n\n**Architecture and historical preservation.** When we say \"they built things to last in the old days,\" we mean: the buildings that survived to be admired today are the ones built to last. The crumbling sheds, the collapsed houses, the burned-down tenements are not in the visible inventory. Modern buildings face the same survival filter; we just haven't done the filtering yet.\n\n**Cultural products.** The \"classics\" of literature, music, and film are the survivors of centuries of evaluation. The forgotten 19th-century novels include masterpieces and unreadable hack work; the survivors are a biased sample. The same is true of folk traditions, religious texts (the documents that didn't get copied are gone), and historical figures.\n\nThe framework's most consequential modern application is in **startup and venture capital reasoning**. The base rate for startup failure is roughly 90% over 10 years; for venture-backed startups, roughly 75%. Most popular advice (\"pivot quickly,\" \"found in a garage,\" \"raise from top-tier VCs\") comes from the surviving 10-25%. The non-survivors include companies that did all the same things and failed. The advice may be correct, or it may be selection effect — the data alone cannot distinguish.\n\nPeter Thiel articulated the survivorship-bias problem in startup wisdom in *Zero to One* (2014):\n\n> \"The most successful companies arrange every single piece of their business around a single insight — what they call a 'secret' — that other people don't share. The trouble is, we can only see this pattern in retrospect, among the companies that won. The companies that had a 'secret' that turned out to be wrong are dead, and we can't see them. So we cannot distinguish between 'successful companies all had a secret' (a causal claim) and 'companies with secrets had high variance, and we only see the upside tail.'\"\n>\n> — Thiel, P. & Masters, B. (2014). *Zero to One: Notes on Startups, or How to Build the Future.* Crown Business. ISBN 978-0804139298, p. 95-97.\n\nThis is a remarkable concession from a venture capitalist whose business depends on identifying winners: the standard story of \"what successful companies did\" is structurally vulnerable to survivorship bias, and the available data cannot resolve it.\n\nThree operational lessons from extensive application:\n\n**First, the bias is invisible without explicit attention.** The survivor sample is what's *available* to study; the non-survivor sample is, by definition, hard to access. The default analysis silently ignores the missing data. Only deliberate questioning (\"what would the non-survivors look like?\") brings the bias into view.\n\n**Second, the bias produces over-confident wrong conclusions, not random error.** Survivorship bias is directional — it always inflates the appearance of cause-effect in survivor traits. This makes it especially dangerous in advice and prediction: the wrong answer is confidently presented.\n\n**Third, the only structural correction is access to non-survivor data.** When the non-survivor data is available (failed companies, dead patients, unpublished studies), corrected analysis is possible. When it's not, conclusions must be marked as conditional on the survivor sample.\n\nFile v1.0.5:examples/ai-startup-survivorship-2023-2026.md\n\n# Method in Action: AI-Startup Survivorship (2023–2026)\n\n> *Example for the [survivorship-bias](../SKILL.md) skill.*\n\nSince the launch of ChatGPT in late 2022, a recurring argument has powered founder decks, pitch meetings, and dinner-table career advice: *\"AI startups are the place to be — look at OpenAI, Anthropic, and the wave of AI 'wrapper' companies that raised at billion-dollar valuations in barely a year.\"* The claim looks overwhelmingly data-driven. Everyone can name the winners; funding-round headlines arrive weekly; the trajectory of the visible sample looks near-vertical. That is Step 1 of the Process in its most seductive modern form — a confident conclusion about *the odds of building an AI startup* drawn from a sample nobody has audited for its filter.\n\n**State the claim (Step 1).** The concluded proposition is roughly: \"Building an AI startup right now has unusually high odds of success, because the companies doing it are winning.\" The sample is the set of AI companies that are *visible* — the ones being written about, raising rounds, or trading as public comparables. The source is founder lore, tech press, and venture marketing, all of which report on companies that currently exist and are currently newsworthy.\n\n**Identify the survival filter (Step 2).** The visible sample is filtered twice over. First, a startup only becomes newsworthy or fundable once it has already cleared early hurdles — the ones that never raised, never launched, or quietly shut down inside the first year rarely generate a headline. Second, and specific to this wave, an enormous number of thin \"GPT-wrapper\" products were spun up on top of third-party model APIs; most never reached durable revenue, and many were rendered redundant the moment the underlying model vendor shipped the same feature natively. The population is *every AI startup that was ever attempted* in this period; the visible sample is the small residue that survived long enough to be counted. The base rate for the underlying category is brutal and well-established independent of the AI hype: the U.S. Bureau of Labor Statistics' Business Employment Dynamics series has for decades shown that roughly half of new U.S. businesses fail within five years, and venture-backed technology startups fail at higher rates than that baseline. The filter selects hard *against* failure being visible.\n\n**Construct the non-survivor hypothesis (Step 3).** The invisible graveyard is not a random subset of attempts. It is disproportionately made of companies that did many of the same things the celebrated winners did — assembled a team, shipped an LLM-powered product, chased the same \"AI-native\" positioning — and failed anyway, for reasons that had little to do with any single visible trait. Two structural forces make this graveyard especially large in 2023–2026. (1) **Platform risk:** wrapper startups whose entire value was a prompt-and-UI layer over a vendor model were exposed to being absorbed the instant the vendor expanded scope — a dynamic widely discussed as \"the model ate my startup.\" (2) **Cost and moat asymmetry:** the genuinely defensible frontier-model work required capital at a scale (multi-billion-dollar compute and training budgets) that almost no startup could raise, so the resource that actually explained the frontier winners' survival was unavailable to the very founders being told to imitate them. If the dead companies had looked like the survivors on the traits usually cited (\"AI-native,\" \"fast-moving,\" \"great demo\"), they would still be alive — they largely shared those traits and died regardless.\n\n**Re-estimate strength (Step 4).** Best case for the claim: some rare, hard-to-copy trait (proprietary data, deep distribution, elite research talent, or the capital to train frontier models) explains survival, and the non-survivors genuinely lacked it. Worst case: the celebrated traits were shared by the graveyard, and survival was driven by factors that are either non-replicable (being OpenAI/Anthropic-scale from the start) or partly stochastic (timing, a single anchor customer, one well-timed round before the funding window tightened). The evidence tilts toward the worst case for the *median* AI-startup bet: two widely reported facts discipline the hype. First, the market itself repriced the \"wrapper\" thesis in January 2025 when China's DeepSeek released a competitive model reportedly trained at a small fraction of frontier budgets, triggering a sharp single-day sell-off in AI-exposed equities (Nvidia's roughly $600 billion one-day market-value drop on January 27, 2025 was widely reported as the largest single-day loss in market history) — a public demonstration that the moat many startups assumed was durable was thinner than believed. Second, the capital concentrating in a handful of frontier labs (multi-billion-dollar rounds into OpenAI and Anthropic through 2024–2025) is the opposite signal from \"easy odds for a new entrant\"; it is evidence that survival at the frontier is gated by resources the median founder cannot access. Neither fact says AI startups can't win — it says the *rate* implied by staring at the winners is an artifact of the filter.\n\n**Correct or mark (Step 5).** The fix is structural, not motivational: recover the denominator before quoting the numerator. Anchor the prior on the population base rate for startups in the relevant category (BLS/venture failure rates), then ask what *specific, non-shared, hard-to-copy* advantage a given AI startup has that the graveyard lacked — proprietary data, real distribution, switching costs, or genuine research/compute scale — rather than any trait the dead companies also had. Where that population data is unavailable for a private, fast-moving market, mark the \"AI startups are winning\" conclusion as conditional on the survivor sample and explicitly not a statement about a new entrant's odds. This is the pure modern analogue of Wald's bombers: the returning planes are the funded unicorns everyone photographs; the armor belongs on the parts of the plane that correspond to where the missing startups were hit — platform risk, absent moat, and the capital wall — precisely the regions the survivor sample can never show you.\n\nThe mapped steps:\n\n1. State the claim: \"AI startups have high odds of success,\" sourced from the visible, newsworthy, currently-funded AI companies.\n2. Identify the survival filter: only startups that cleared early hurdles become visible/fundable, and thin GPT-wrappers were culled en masse (often absorbed by the model vendors); population = all attempts, sample = the survivors. Baseline failure is high — roughly half of new U.S. businesses fail within five years (BLS), higher for venture-backed tech.\n3. Construct the non-survivor hypothesis: the graveyard shared the celebrated traits (\"AI-native,\" fast, great demo) and died anyway, mostly from platform risk and a capital/moat wall the winners' imitators could not scale.\n4. Re-estimate strength: worst case dominates for the median bet — the DeepSeek-triggered repricing (Jan 27, 2025, Nvidia's ~$600B single-day drop) and the concentration of capital into a few frontier labs both show the moat is thin for entrants and survival at the frontier is resource-gated, not trait-explained.\n5. Correct or mark: reset the prior to the startup population base rate, then require a specific non-shared advantage; absent population data, mark \"AI startups are winning\" as conditional on survivors, not as an entrant's odds.\n\n*Sources: U.S. Bureau of Labor Statistics, Business Employment Dynamics — survival rates of establishments (roughly half of new U.S. businesses survive five years), https://www.bls.gov/bdm/ . Reuters and Associated Press reporting on the January 27, 2025 AI-stock sell-off following DeepSeek's model release, including Nvidia's record single-day market-capitalization decline (widely reported at approximately $600 billion). OpenAI and Anthropic public funding announcements, 2023–2025. Thiel, P. & Masters, B. (2014). *Zero to One.* Crown Business — on why \"competition\" and copied traits mislead startup reasoning.*\n\nFile v1.0.5:examples/mutual-fund-survivorship-and-reported-returns-1996.md\n\n# Method in Action: Mutual Fund Survivorship and Reported Returns (1971–1996)\n\n> *Example for the [survivorship-bias](../SKILL.md) skill.*\n\nThrough the early 1990s, the standard evidence for \"mutual funds deliver strong long-run returns\" came from commercial fund databases and industry performance tables. The claim looked data-driven: average the annual returns of the funds in the database over a decade or two, and the industry's track record appears solid. The problem is Step 1 of the Process in miniature — a confident conclusion drawn from a sample nobody had audited for its filter.\n\n**The survival filter (Step 2).** Fund databases of the era typically listed funds *currently in existence*. When a fund performed badly, its sponsor liquidated it or merged it into a better-performing sibling — and it silently dropped out of the historical tables. The population was every fund that ever operated over the sample period; the visible sample was only the funds that lasted to the end. Attrition was not rare: in Burton Malkiel's data on equity funds from 1971 to 1991, a substantial fraction of all funds that ever existed had disappeared before the end of the period, and disappearance was strongly associated with poor prior performance.\n\n**The non-survivor hypothesis (Step 3).** The vanished funds were not a random subset. They were disproportionately the losers — funds that trailed the market, bled assets, and were shut down or merged away. If they had looked like the survivors, sponsors would have had no reason to kill them. So the visible sample's average return must overstate the average return an investor choosing a fund *at the start* of the period would actually have earned.\n\n**Re-estimating the claim (Step 4).** Two studies made the correction quantitative. Malkiel (1995) reconstructed the full population of equity funds, including the dead ones, and found that the average annual return of surviving funds materially exceeded the average across all funds — a gap on the order of a percentage point or more per year during the 1980s. Elton, Gruber & Blake (1996) took the cleanest design: they froze the set of funds existing in 1977 and followed *every one of them* forward through 1993, tracking merged funds into their successors so nothing could exit the sample. They found that survivorship bias inflates measured performance by an amount that grows with the length of the sample period — roughly a percentage point per year over long horizons — enough to turn apparent stock-picking skill into underperformance after costs.\n\n**The correction (Step 5).** The fix was structural, not rhetorical: rebuild the sample so the filter cannot operate. Follow-forward designs and survivor-bias-free databases (such as the CRSP mutual fund database) became the methodological standard in academic finance. Performance claims computed on survivor-only samples are now treated as upward-biased by construction, and the burden of proof sits on anyone citing them.\n\nThe episode is the pure financial analogue of Wald's bombers, with one difference that makes it more insidious: nobody shot the losing funds down in public. They were quietly merged away by the same institutions whose track records benefited from their disappearance — the filter and the beneficiary of the filter were the same party.\n\nThe mapped steps:\n\n1. State the claim: \"mutual funds earn strong long-run returns,\" sourced from databases of currently existing funds.\n2. Identify the survival filter: poorly performing funds were liquidated or merged and dropped from historical tables; the sample is conditioned on lasting to the end of the period.\n3. Construct the non-survivor hypothesis: the dead funds were disproportionately the underperformers; had they resembled survivors, they would not have been closed.\n4. Re-estimate strength: follow-forward reconstructions (Malkiel 1995; Elton, Gruber & Blake 1996) show survivor-only averages overstate returns by roughly a percentage point per year, growing with horizon length.\n5. Correct or mark: adopt survivor-bias-free samples that track every fund from inception forward, merging targets included; treat survivor-only performance tables as conditional evidence only.\n\nPrimary sources: Elton, E. J., Gruber, M. J., & Blake, C. R. (1996). \"Survivorship Bias and Mutual Fund Performance.\" *Review of Financial Studies*, 9(4), 1097–1120. Malkiel, B. G. (1995). \"Returns from Investing in Equity Mutual Funds 1971 to 1991.\" *Journal of Finance*, 50(2), 549–572.\n\nFile v1.0.5:skill-card.md\n\n## Description:\n\nThis skill helps agents detect survivorship bias by identifying survivor-filtered samples, constructing non-survivor hypotheses, and marking conclusions as conditional unless selection-corrected data is available.\n\nThis skill is ready for commercial/non-commercial use.\n\n## Publisher:\n\n[deciqai](https://clawhub.ai/user/deciqai)\n\n### License/Terms of Use:\n\nMIT-0\n\n## Use Case:\n\nExternal users and agents use this skill to scrutinize claims based on visible winners, successful companies, active funds, completed treatments, published studies, or other survivor-only samples. It guides the agent to identify the survival filter, reason about missing non-survivors, and produce a corrected or explicitly conditional inference.\n\n### Deployment Geography for Use:\n\nGlobal\n\n## Known Risks and Mitigations:\n\nRisk: The skill can be applied to business, investment, medical, and career examples where user-provided facts may be sensitive or high impact.\n\nMitigation: Avoid saving private user details as observed examples unless the user clearly asks for it and the details are redacted.\n\nRisk: Survivor-only examples can lead to overconfident conclusions if missing non-survivor data is not available.\n\nMitigation: Mark conclusions as conditional on the survivor sample unless population data is available and selection correction has been performed.\n\n## Reference(s):\n\n- [ClawHub skill page](https://clawhub.ai/deciqai/skills/survivorship-bias)\n- [Primary sources](references/sources.md)\n- [Abraham Wald and the Statistical Research Group, 1943](examples/abraham-wald-and-the-statistical-research-group-1943.md)\n- [Mutual Fund Survivorship and Reported Returns, 1971-1996](examples/mutual-fund-survivorship-and-reported-returns-1996.md)\n- [AI-Startup Survivorship, 2023-2026](examples/ai-startup-survivorship-2023-2026.md)\n- [U.S. Bureau of Labor Statistics Business Employment Dynamics](https://www.bls.gov/bdm/)\n\n## Skill Output:\n\n**Output Type(s):** [text, markdown, guidance]\n\n**Output Format:** [Markdown analysis with structured claim, survival-filter, non-survivor hypothesis, and corrected-inference fields]\n\n**Output Parameters:** [1D]\n\n**Other Properties Related to Output:** [May provide step-by-step coaching prompts for novice users; no code execution or sensitive access is requested.]\n\n## Skill Version(s):\n\n1.0.5 (source: server release evidence)\n\n## Ethical Considerations:\n\nUsers should evaluate whether this skill is appropriate for their environment, review any generated or modified files before relying on them, and apply their organization's safety, security, and compliance requirements before deployment.\n\nArchive v1.0.4: 7 files, 16153 bytes\n\nFiles: examples/abraham-wald-and-the-statistical-research-group-1943.md (7443b), examples/ai-startup-survivorship-2023-2026.md (8178b), examples/mutual-fund-survivorship-and-reported-returns-1996.md (4526b), references/sources.md (2398b), skill-card.md (2400b), SKILL.md (7104b), _meta.json (136b)\n\nFile v1.0.4:SKILL.md\n\n---\nname: survivorship-bias\ndescription: \"Activate when: user says 'look at what winners/billionaires/champions did,' investment returns or fund performance are being cited, a strategy is justified by pointing to companies that succeeded, historical data is treated as representative of all cases, or someone says 'this works because X did it.'\n  Do NOT activate when: population data is available and already selection-corrected; analysis is explicitly about survivors only with no claim about the broader population.\"\n---\n\n# Survivorship Bias\n\n## Overview\n\n**Survivorship bias** is drawing conclusions from a sample pre-filtered by survival — treating survivor traits as the *cause* of survival when non-survivors (absent from data by definition) may have had identical traits and still failed.\n\nCanon: Wald (1943) reversed the Navy's bomber-armor recommendation — returning planes showed damage where hits were survivable; the missing planes (shot down) were hit where returning planes showed no damage. Armor the gaps, not the hits.\n\nComposes with `bayesian-reasoning` (prior = population, not survivors), `critical-thinking` (what would non-survivors say?), `first-principles` (population is bedrock), and `abductive-reasoning` (\"winners have trait Y\" is one hypothesis; randomness is another).\n\n## When to Use\n\n- Someone draws lessons from \"what successful X did\"\n- Investment returns / fund performance / backtested strategies are cited\n- A business strategy is justified by pointing to companies that used it\n- Medical / treatment success rates are reported without dropout data\n- Career advice comes from what top performers did\n- Odds of building an AI startup are inferred from the visible AI winners (funded unicorns, \"wrapper\" success stories) amid the AI-bubble / AI-capex debate\n\n**Not when:** population data available and filter already corrected; analysis is intentionally about survivors only with no population claim.\n\n## Coaching Novices (Adaptive Front Door)\n\n- **Engine mode:** user has a concrete claim and data → run The Process directly.\n- **Coach mode:** unfamiliar or no case → guide step by step.\n\nIn Coach mode, respond one step at a time. Each [WAIT] is a hard stop — output only that step's question, then stop.\n\n1. One-liner: \"Before believing 'X worked because winners did X,' ask whether the losers did the same X — they're not in your sample to refute it.\"\n2. Check fit: if sample is explicitly restricted to survivors with no population claim, this lens doesn't apply.\n3. Elicit their real claim and the visible data they have.\n> **[WAIT — do not advance until user responds]**\n4. Run The Process one step at a time: what filter produced this sample? what's missing? if missing data looked like the sample, would the conclusion hold?\n> **[WAIT — do not advance until user responds]**\n5. Close by naming the selection-corrected conclusion (or marking it unprovable from this data alone).\n> **[WAIT — do not advance until user responds]**\n\n## The Process\n\n**Step 1 — State the claim:** What is being concluded, from what sample, from what source?\n\n**Step 2 — Identify the survival filter:** What process produced this sample? What was the population before the filter? What fraction was removed? What did the filter select for/against?\n\n**Step 3 — Construct the non-survivor hypothesis:** What did non-survivors likely have? Did they share the trait attributed to success? Would the claim hold if we could see them?\n\n**Step 4 — Re-estimate strength:** Best case = trait explains survival (non-survivors lacked it). Worst case = trait doesn't explain survival (non-survivors had it too). What evidence distinguishes these?\n\n**Step 5 — Correct or mark:** Get population data and re-run analysis with selection correction. If unavailable, mark conclusion as conditional on survivor sample.\n\n## Output Template\n\n```markdown\n# Survivorship Bias Analysis: <claim>\nClaim / sample / source:\nSurvival filter (what removed non-survivors, population size est., survival rate est.):\nNon-survivor hypothesis (what they likely had/lacked, could they have had same trait):\nCorrected inference (conclusion, confidence, what data would settle it):\n```\n\n*→ Method in Action: [Abraham Wald and the Statistical Research Group, 1943](examples/abraham-wald-and-the-statistical-research-group-1943.md) · [Mutual Fund Survivorship and Reported Returns, 1971–1996](examples/mutual-fund-survivorship-and-reported-returns-1996.md)*\n*→ 2026 lens: [AI-startup survivorship — funded unicorns vs the dead-wrapper graveyard (2023–2026)](examples/ai-startup-survivorship-2023-2026.md)*\n\n## Pack: Common Survivor Patterns\n\n| Domain | Survivor sample | Missing non-survivor data | Biased claim |\n|---|---|---|---|\n| Business / startup | Surviving companies | Failed companies | \"Successful companies do X\" |\n| Investment returns | Active funds / listed stocks | Closed funds / delisted stocks | \"Stocks return 10% annually\" |\n| Career advice | Top performers | People who left the field | \"To succeed, do X\" |\n| Scientific findings | Published studies | Unpublished null results | \"X is significant\" |\n| Treatment efficacy | Patients who completed | Drop-outs, deaths during treatment | \"X% recovered\" |\n\n*→ Primary sources: [references/sources.md](references/sources.md)*\n\n## Common Rationalizations\n\n**[D] = designed upfront | [O] = observed in real use. [O] entries are more valuable.**\n\n| Fake move | Reality |\n|---|---|\n| [D] \"Look at the data\" (survivor sample) | Survivor data ≠ population data. Correct or mark as conditional. |\n| [D] Citing one famous example as proof | N=1 in survivor sample tells you nothing about the rate. |\n| [D] \"X is the formula for success\" | If failures did the same X, X is not the formula. Get non-survivor data or stop claiming. |\n| [D] \"We use a backtested strategy\" | If backtest excludes failed/delisted stocks, results are upward-biased. |\n| [D] \"If we had non-survivor data, we'd see the same pattern\" | Unfalsifiable without the data. Get it or hold the claim. |\n| *→ Add [O] entries here after each real use — paste the actual failure pattern* | *What went wrong and why* |\n\n## Red Flags\n\n- Sample described as \"successful X\" or \"the X who made it\"\n- Data source is survivor-filtered (active funds, surviving companies, published studies)\n- Base rate of failure / dropout not stated\n- Conclusions about a population drawn from the survivor subset\n\n## Verification\n\n- [ ] Survival filter identified\n- [ ] Non-survivor population size estimated\n- [ ] Non-survivor hypothesis constructed\n- [ ] Conclusions conditional on survivor sample, or formally selection-corrected\n- [ ] Recommendation robust to worst-case non-survivor hypothesis\n\n---\n\n*Part of **deciqAI Knowledge Skills** — 164 open-source thinking skills that make rigor executable for AI agents. The same skills power every deciqAI agent, which runs them autonomously to operate your company. **See it run → https://www.deciqai.com/c/survivorship-bias** · ⭐ Star the repo → https://github.com/deciqAI/knowledge-skills · Contributions welcome.*\n\nFile v1.0.4:_meta.json\n\n{\n  \"ownerId\": \"kn754b8sk22s8c6gjxt02bftbn88q7ye\",\n  \"slug\": \"survivorship-bias\",\n  \"version\": \"1.0.4\",\n  \"publishedAt\": 1783509692798\n}\n\nFile v1.0.4:references/sources.md\n\n# Sources — survivorship-bias\n\n> *Primary sources for the [survivorship-bias](../SKILL.md) skill.*\n\n- Wald, A. (1943). \"A Method of Estimating Plane Vulnerability Based on Damage of Survivors.\" SRG Memorandum CRC 432, Statistical Research Group, Columbia University. Republished 1980 by Center for Naval Analyses.\n- Heckman, J. J. (1979). \"Sample selection bias as a specification error.\" *Econometrica*, 47(1), 153-161. The Nobel-laureate selection-correction formalization.\n- Brown, S. J., Goetzmann, W., Ibbotson, R. G., & Ross, S. A. (1992). \"Survivorship Bias in Performance Studies.\" *Review of Financial Studies*, 5(4), 553-580. The financial-economics demonstration.\n- Elton, E. J., Gruber, M. J., & Blake, C. R. (1996). \"Survivorship Bias and Mutual Fund Performance.\" *Review of Financial Studies*, 9(4), 1097-1120. The follow-forward mutual fund correction: bias grows with sample length.\n- Malkiel, B. G. (1995). \"Returns from Investing in Equity Mutual Funds 1971 to 1991.\" *Journal of Finance*, 50(2), 549-572. Full-population reconstruction of equity fund returns including dead funds.\n- Mangel, M. & Samaniego, F. J. (1984). \"Abraham Wald's Work on Aircraft Survivability.\" *Journal of the American Statistical Association*, 79(386), 259-267. The post-declassification academic treatment.\n- Thiel, P. & Masters, B. (2014). *Zero to One: Notes on Startups.* Crown Business. ISBN 978-0804139298. The startup-context articulation.\n- Taleb, N. N. (2001). *Fooled by Randomness: The Hidden Role of Chance in Life and in the Markets.* Random House. ISBN 978-0812975215. Sustained treatment of survivorship bias in finance.\n- Ioannidis, J. P. A. (2005). \"Why Most Published Research Findings Are False.\" *PLoS Medicine*, 2(8), e124. Publication bias as survivor bias at the meta-research level.\n- U.S. Bureau of Labor Statistics, Business Employment Dynamics — Entrepreneurship and the U.S. Economy / establishment survival rates. https://www.bls.gov/bdm/ . Baseline population failure rate for new U.S. businesses (roughly half survive five years) — the denominator that AI-startup survivor claims omit.\n- Reuters / Associated Press (January 27–28, 2025). Reporting on the DeepSeek-triggered AI-stock sell-off, including Nvidia's record single-day market-value decline (widely reported at approximately $600 billion). Public market repricing of the thin-moat \"AI wrapper\" thesis.\n\nFile v1.0.4:examples/abraham-wald-and-the-statistical-research-group-1943.md\n\n# Method in Action: Abraham Wald and the Statistical Research Group, 1943\n\n> *Example for the [survivorship-bias](../SKILL.md) skill.*\n\nThe canonical demonstration of survivorship bias is **Abraham Wald**'s wartime work at the **Statistical Research Group (SRG)** at Columbia University, 1942-1945. The SRG was the secret mathematical-statistics arm of the wartime US government, comparable in caliber to the Manhattan Project for statistics: it included Wald, Allen Wallis, Milton Friedman, Frederick Mosteller, and Jacob Wolfowitz, among others.\n\nThe bomber-armor problem was assigned to the SRG in 1943. The US Navy was losing aircraft over Europe and the Pacific in unsustainable numbers. The problem was to determine where to add additional armor on B-17 and B-24 bombers — adding armor everywhere would make the plane too heavy to fly its mission, so the armor had to be concentrated where it would do the most good. The Navy collected data on bullet holes and damage patterns in aircraft that returned from missions.\n\nThe pattern looked clear. Damage was concentrated in specific areas:\n\n- **Wings:** many holes\n- **Fuselage center:** many holes\n- **Tail section:** many holes\n- **Engine area:** few holes\n- **Cockpit / pilot area:** few holes\n\nThe Navy's analyst recommended adding armor to the areas with the most damage — the wings, fuselage, and tail. Wald reviewed the data and recommended the *opposite*.\n\nHis reasoning, preserved in the SRG memorandum CRC 432, July 1943 (declassified 1980), opens with what may be the most famous selection-effect argument in 20th-century applied statistics:\n\n> \"The bullet holes are concentrated where the armor is not needed. The aircraft that we observe — the ones that returned — have damage on the wings, fuselage, and tail. The aircraft that did not return are not in our sample. They were shot down because they were hit where we observe little damage in the returnees: the engines and the cockpit. The damage we observe is the damage that planes can sustain and still fly home. The missing damage — the gaps in our pattern — is the lethal damage. Armor must go where the returners show no holes.\"\n>\n> — Wald, A. (1943). \"A Method of Estimating Plane Vulnerability Based on Damage of Survivors.\" SRG Memorandum CRC 432, Statistical Research Group, Columbia University. Republished in 1980 as CNA Memorandum.\n\nThe recommendation was implemented. Modeled losses dropped substantially. Wald's analysis became foundational to the **selection-effects literature** in statistics, econometrics, and epidemiology.\n\nThe deeper lesson Wald drew: **the observed sample is conditional on survival**, and the conditional distribution can be the opposite of the unconditional distribution. The damage pattern in returning aircraft is *anti-correlated* with the damage pattern in shot-down aircraft, because the very planes that absorbed the most lethal hits are absent from the data.\n\nWald's framework was systematized. The post-war statistics literature developed:\n\n**Heckman selection correction (1979).** James Heckman, building on Wald's logic, developed the formal mathematical framework for selection-corrected estimation. The \"Heckman correction\" is the standard treatment in modern econometrics; it won Heckman the 2000 Nobel Prize in Economics. See Heckman, J. J. (1979). \"Sample selection bias as a specification error.\" *Econometrica*, 47(1), 153-161.\n\n**Publication bias in scientific research.** Studies showing positive effects are more likely to be published than studies showing null effects. Ioannidis (2005) \"Why most published research findings are false,\" *PLoS Medicine*, 2(8), e124, drew on this. Survivorship bias at the meta-research level: the published literature is the survivor sample of all conducted research.\n\n**Funds and indices in finance.** Brown, Goetzmann, Ibbotson, and Ross (1992), \"Survivorship Bias in Performance Studies,\" *Review of Financial Studies*, 5(4), 553-580, demonstrated that mutual fund returns reported in standard databases were systematically biased upward because failed funds had been dropped. The corrected estimates were 1-3 percentage points lower per year — a massive effect compounded over decades.\n\n**Architecture and historical preservation.** When we say \"they built things to last in the old days,\" we mean: the buildings that survived to be admired today are the ones built to last. The crumbling sheds, the collapsed houses, the burned-down tenements are not in the visible inventory. Modern buildings face the same survival filter; we just haven't done the filtering yet.\n\n**Cultural products.** The \"classics\" of literature, music, and film are the survivors of centuries of evaluation. The forgotten 19th-century novels include masterpieces and unreadable hack work; the survivors are a biased sample. The same is true of folk traditions, religious texts (the documents that didn't get copied are gone), and historical figures.\n\nThe framework's most consequential modern application is in **startup and venture capital reasoning**. The base rate for startup failure is roughly 90% over 10 years; for venture-backed startups, roughly 75%. Most popular advice (\"pivot quickly,\" \"found in a garage,\" \"raise from top-tier VCs\") comes from the surviving 10-25%. The non-survivors include companies that did all the same things and failed. The advice may be correct, or it may be selection effect — the data alone cannot distinguish.\n\nPeter Thiel articulated the survivorship-bias problem in startup wisdom in *Zero to One* (2014):\n\n> \"The most successful companies arrange every single piece of their business around a single insight — what they call a 'secret' — that other people don't share. The trouble is, we can only see this pattern in retrospect, among the companies that won. The companies that had a 'secret' that turned out to be wrong are dead, and we can't see them. So we cannot distinguish between 'successful companies all had a secret' (a causal claim) and 'companies with secrets had high variance, and we only see the upside tail.'\"\n>\n> — Thiel, P. & Masters, B. (2014). *Zero to One: Notes on Startups, or How to Build the Future.* Crown Business. ISBN 978-0804139298, p. 95-97.\n\nThis is a remarkable concession from a venture capitalist whose business depends on identifying winners: the standard story of \"what successful companies did\" is structurally vulnerable to survivorship bias, and the available data cannot resolve it.\n\nThree operational lessons from extensive application:\n\n**First, the bias is invisible without explicit attention.** The survivor sample is what's *available* to study; the non-survivor sample is, by definition, hard to access. The default analysis silently ignores the missing data. Only deliberate questioning (\"what would the non-survivors look like?\") brings the bias into view.\n\n**Second, the bias produces over-confident wrong conclusions, not random error.** Survivorship bias is directional — it always inflates the appearance of cause-effect in survivor traits. This makes it especially dangerous in advice and prediction: the wrong answer is confidently presented.\n\n**Third, the only structural correction is access to non-survivor data.** When the non-survivor data is available (failed companies, dead patients, unpublished studies), corrected analysis is possible. When it's not, conclusions must be marked as conditional on the survivor sample.\n\nFile v1.0.4:examples/ai-startup-survivorship-2023-2026.md\n\n# Method in Action: AI-Startup Survivorship (2023–2026)\n\n> *Example for the [survivorship-bias](../SKILL.md) skill.*\n\nSince the launch of ChatGPT in late 2022, a recurring argument has powered founder decks, pitch meetings, and dinner-table career advice: *\"AI startups are the place to be — look at OpenAI, Anthropic, and the wave of AI 'wrapper' companies that raised at billion-dollar valuations in barely a year.\"* The claim looks overwhelmingly data-driven. Everyone can name the winners; funding-round headlines arrive weekly; the trajectory of the visible sample looks near-vertical. That is Step 1 of the Process in its most seductive modern form — a confident conclusion about *the odds of building an AI startup* drawn from a sample nobody has audited for its filter.\n\n**State the claim (Step 1).** The concluded proposition is roughly: \"Building an AI startup right now has unusually high odds of success, because the companies doing it are winning.\" The sample is the set of AI companies that are *visible* — the ones being written about, raising rounds, or trading as public comparables. The source is founder lore, tech press, and venture marketing, all of which report on companies that currently exist and are currently newsworthy.\n\n**Identify the survival filter (Step 2).** The visible sample is filtered twice over. First, a startup only becomes newsworthy or fundable once it has already cleared early hurdles — the ones that never raised, never launched, or quietly shut down inside the first year rarely generate a headline. Second, and specific to this wave, an enormous number of thin \"GPT-wrapper\" products were spun up on top of third-party model APIs; most never reached durable revenue, and many were rendered redundant the moment the underlying model vendor shipped the same feature natively. The population is *every AI startup that was ever attempted* in this period; the visible sample is the small residue that survived long enough to be counted. The base rate for the underlying category is brutal and well-established independent of the AI hype: the U.S. Bureau of Labor Statistics' Business Employment Dynamics series has for decades shown that roughly half of new U.S. businesses fail within five years, and venture-backed technology startups fail at higher rates than that baseline. The filter selects hard *against* failure being visible.\n\n**Construct the non-survivor hypothesis (Step 3).** The invisible graveyard is not a random subset of attempts. It is disproportionately made of companies that did many of the same things the celebrated winners did — assembled a team, shipped an LLM-powered product, chased the same \"AI-native\" positioning — and failed anyway, for reasons that had little to do with any single visible trait. Two structural forces make this graveyard especially large in 2023–2026. (1) **Platform risk:** wrapper startups whose entire value was a prompt-and-UI layer over a vendor model were exposed to being absorbed the instant the vendor expanded scope — a dynamic widely discussed as \"the model ate my startup.\" (2) **Cost and moat asymmetry:** the genuinely defensible frontier-model work required capital at a scale (multi-billion-dollar compute and training budgets) that almost no startup could raise, so the resource that actually explained the frontier winners' survival was unavailable to the very founders being told to imitate them. If the dead companies had looked like the survivors on the traits usually cited (\"AI-native,\" \"fast-moving,\" \"great demo\"), they would still be alive — they largely shared those traits and died regardless.\n\n**Re-estimate strength (Step 4).** Best case for the claim: some rare, hard-to-copy trait (proprietary data, deep distribution, elite research talent, or the capital to train frontier models) explains survival, and the non-survivors genuinely lacked it. Worst case: the celebrated traits were shared by the graveyard, and survival was driven by factors that are either non-replicable (being OpenAI/Anthropic-scale from the start) or partly stochastic (timing, a single anchor customer, one well-timed round before the funding window tightened). The evidence tilts toward the worst case for the *median* AI-startup bet: two widely reported facts discipline the hype. First, the market itself repriced the \"wrapper\" thesis in January 2025 when China's DeepSeek released a competitive model reportedly trained at a small fraction of frontier budgets, triggering a sharp single-day sell-off in AI-exposed equities (Nvidia's roughly $600 billion one-day market-value drop on January 27, 2025 was widely reported as the largest single-day loss in market history) — a public demonstration that the moat many startups assumed was durable was thinner than believed. Second, the capital concentrating in a handful of frontier labs (multi-billion-dollar rounds into OpenAI and Anthropic through 2024–2025) is the opposite signal from \"easy odds for a new entrant\"; it is evidence that survival at the frontier is gated by resources the median founder cannot access. Neither fact says AI startups can't win — it says the *rate* implied by staring at the winners is an artifact of the filter.\n\n**Correct or mark (Step 5).** The fix is structural, not motivational: recover the denominator before quoting the numerator. Anchor the prior on the population base rate for startups in the relevant category (BLS/venture failure rates), then ask what *specific, non-shared, hard-to-copy* advantage a given AI startup has that the graveyard lacked — proprietary data, real distribution, switching costs, or genuine research/compute scale — rather than any trait the dead companies also had. Where that population data is unavailable for a private, fast-moving market, mark the \"AI startups are winning\" conclusion as conditional on the survivor sample and explicitly not a statement about a new entrant's odds. This is the pure modern analogue of Wald's bombers: the returning planes are the funded unicorns everyone photographs; the armor belongs on the parts of the plane that correspond to where the missing startups were hit — platform risk, absent moat, and the capital wall — precisely the regions the survivor sample can never show you.\n\nThe mapped steps:\n\n1. State the claim: \"AI startups have high odds of success,\" sourced from the visible, newsworthy, currently-funded AI companies.\n2. Identify the survival filter: only startups that cleared early hurdles become visible/fundable, and thin GPT-wrappers were culled en masse (often absorbed by the model vendors); population = all attempts, sample = the survivors. Baseline failure is high — roughly half of new U.S. businesses fail within five years (BLS), higher for venture-backed tech.\n3. Construct the non-survivor hypothesis: the graveyard shared the celebrated traits (\"AI-native,\" fast, great demo) and died anyway, mostly from platform risk and a capital/moat wall the winners' imitators could not scale.\n4. Re-estimate strength: worst case dominates for the median bet — the DeepSeek-triggered repricing (Jan 27, 2025, Nvidia's ~$600B single-day drop) and the concentration of capital into a few frontier labs both show the moat is thin for entrants and survival at the frontier is resource-gated, not trait-explained.\n5. Correct or mark: reset the prior to the startup population base rate, then require a specific non-shared advantage; absent population data, mark \"AI startups are winning\" as conditional on survivors, not as an entrant's odds.\n\n*Sources: U.S. Bureau of Labor Statistics, Business Employment Dynamics — survival rates of establishments (roughly half of new U.S. businesses survive five years), https://www.bls.gov/bdm/ . Reuters and Associated Press reporting on the January 27, 2025 AI-stock sell-off following DeepSeek's model release, including Nvidia's record single-day market-capitalization decline (widely reported at approximately $600 billion). OpenAI and Anthropic public funding announcements, 2023–2025. Thiel, P. & Masters, B. (2014). *Zero to One.* Crown Business — on why \"competition\" and copied traits mislead startup reasoning.*\n\nFile v1.0.4:examples/mutual-fund-survivorship-and-reported-returns-1996.md\n\n# Method in Action: Mutual Fund Survivorship and Reported Returns (1971–1996)\n\n> *Example for the [survivorship-bias](../SKILL.md) skill.*\n\nThrough the early 1990s, the standard evidence for \"mutual funds deliver strong long-run returns\" came from commercial fund databases and industry performance tables. The claim looked data-driven: average the annual returns of the funds in the database over a decade or two, and the industry's track record appears solid. The problem is Step 1 of the Process in miniature — a confident conclusion drawn from a sample nobody had audited for its filter.\n\n**The survival filter (Step 2).** Fund databases of the era typically listed funds *currently in existence*. When a fund performed badly, its sponsor liquidated it or merged it into a better-performing sibling — and it silently dropped out of the historical tables. The population was every fund that ever operated over the sample period; the visible sample was only the funds that lasted to the end. Attrition was not rare: in Burton Malkiel's data on equity funds from 1971 to 1991, a substantial fraction of all funds that ever existed had disappeared before the end of the period, and disappearance was strongly associated with poor prior performance.\n\n**The non-survivor hypothesis (Step 3).** The vanished funds were not a random subset. They were disproportionately the losers — funds that trailed the market, bled assets, and were shut down or merged away. If they had looked like the survivors, sponsors would have had no reason to kill them. So the visible sample's average return must overstate the average return an investor choosing a fund *at the start* of the period would actually have earned.\n\n**Re-estimating the claim (Step 4).** Two studies made the correction quantitative. Malkiel (1995) reconstructed the full population of equity funds, including the dead ones, and found that the average annual return of surviving funds materially exceeded the average across all funds — a gap on the order of a percentage point or more per year during the 1980s. Elton, Gruber & Blake (1996) took the cleanest design: they froze the set of funds existing in 1977 and followed *every one of them* forward through 1993, tracking merged funds into their successors so nothing could exit the sample. They found that survivorship bias inflates measured performance by an amount that grows with the length of the sample period — roughly a percentage point per year over long horizons — enough to turn apparent stock-picking skill into underperformance after costs.\n\n**The correction (Step 5).** The fix was structural, not rhetorical: rebuild the sample so the filter cannot operate. Follow-forward designs and survivor-bias-free databases (such as the CRSP mutual fund database) became the methodological standard in academic finance. Performance claims computed on survivor-only samples are now treated as upward-biased by construction, and the burden of proof sits on anyone citing them.\n\nThe episode is the pure financial analogue of Wald's bombers, with one difference that makes it more insidious: nobody shot the losing funds down in public. They were quietly merged away by the same institutions whose track records benefited from their disappearance — the filter and the beneficiary of the filter were the same party.\n\nThe mapped steps:\n\n1. State the claim: \"mutual funds earn strong long-run returns,\" sourced from databases of currently existing funds.\n2. Identify the survival filter: poorly performing funds were liquidated or merged and dropped from historical tables; the sample is conditioned on lasting to the end of the period.\n3. Construct the non-survivor hypothesis: the dead funds were disproportionately the underperformers; had they resembled survivors, they would not have been closed.\n4. Re-estimate strength: follow-forward reconstructions (Malkiel 1995; Elton, Gruber & Blake 1996) show survivor-only averages overstate returns by roughly a percentage point per year, growing with horizon length.\n5. Correct or mark: adopt survivor-bias-free samples that track every fund from inception forward, merging targets included; treat survivor-only performance tables as conditional evidence only.\n\nPrimary sources: Elton, E. J., Gruber, M. J., & Blake, C. R. (1996). \"Survivorship Bias and Mutual Fund Performance.\" *Review of Financial Studies*, 9(4), 1097–1120. Malkiel, B. G. (1995). \"Returns from Investing in Equity Mutual Funds 1971 to 1991.\" *Journal of Finance*, 50(2), 549–572.\n\nFile v1.0.4:skill-card.md\n\n## Description: <br>\nSurvivorship Bias helps agents challenge claims drawn from survivor samples by identifying the survival filter, hypothesizing about missing non-survivors, and marking conclusions as corrected or conditional. <br>\n\nThis skill is ready for commercial/non-commercial use. <br>\n\n## Publisher: <br>\n[deciqai](https://clawhub.ai/user/deciqai) <br>\n\n### License/Terms of Use: <br>\nMIT-0 <br>\n\n\n## Use Case: <br>\nDevelopers, analysts, and AI agents use this skill to evaluate claims based on visible winners or surviving samples, including strategy advice, investment returns, startup narratives, treatment outcomes, and career guidance. <br>\n\n### Deployment Geography for Use: <br>\nGlobal <br>\n\n## Known Risks and Mitigations: <br>\nRisk: The skill can steer analysis by prompting survivorship-bias checks even when a claim already uses selection-corrected population data. <br>\nMitigation: Apply the skill only when the sample is survivor-filtered or the broader population claim is not already corrected. <br>\nRisk: The skill may surface publisher and citation links while guiding analysis. <br>\nMitigation: Treat links as references, review them before relying on them, and do not grant operational access based on this markdown-only skill. <br>\n\n\n## Reference(s): <br>\n- [ClawHub Skill Page](https://clawhub.ai/deciqai/skills/survivorship-bias) <br>\n- [Primary Sources](references/sources.md) <br>\n- [Abraham Wald Example](examples/abraham-wald-and-the-statistical-research-group-1943.md) <br>\n- [Mutual Fund Survivorship Example](examples/mutual-fund-survivorship-and-reported-returns-1996.md) <br>\n- [AI Startup Survivorship Example](examples/ai-startup-survivorship-2023-2026.md) <br>\n\n\n## Skill Output: <br>\n**Output Type(s):** [Markdown, Guidance] <br>\n**Output Format:** [Markdown analysis template and stepwise reasoning guidance] <br>\n**Output Parameters:** [1D] <br>\n**Other Properties Related to Output:** [Produces structured survivorship-bias analysis; no executable code or operational access is requested.] <br>\n\n## Skill Version(s): <br>\n1.0.4 (source: server-resolved release metadata) <br>\n\n## Ethical Considerations: <br>\nUsers should evaluate whether this skill is appropriate for their environment, review any generated or modified files before relying on them, and apply their organization's safety, security, and compliance requirements before deployment. <br>\n\nArchive v1.0.3: 6 files, 11806 bytes\n\nFiles: examples/abraham-wald-and-the-statistical-research-group-1943.md (7443b), examples/mutual-fund-survivorship-and-reported-returns-1996.md (4526b), references/sources.md (1806b), skill-card.md (2275b), SKILL.md (6901b), _meta.json (136b)\n\nFile v1.0.3:SKILL.md\n\n---\nname: survivorship-bias\ndescription: \"Activate when: user says 'look at what winners/billionaires/champions did,' investment returns or fund performance are being cited, a strategy is justified by pointing to companies that succeeded, historical data is treated as representative of all cases, or someone says 'this works because X did it.'\n  Do NOT activate when: population data is available and already selection-corrected; analysis is explicitly about survivors only with no claim about the broader population.\"\n---\n\n# Survivorship Bias\n\n## Overview\n\n**Survivorship bias** is drawing conclusions from a sample pre-filtered by survival — treating survivor traits as the *cause* of survival when non-survivors (absent from data by definition) may have had identical traits and still failed.\n\nCanon: Wald (1943) reversed the Navy's bomber-armor recommendation — returning planes showed damage where hits were survivable; the missing planes (shot down) were hit where returning planes showed no damage. Armor the gaps, not the hits.\n\nComposes with `bayesian-reasoning` (prior = population, not survivors), `critical-thinking` (what would non-survivors say?), `first-principles` (population is bedrock), and `abductive-reasoning` (\"winners have trait Y\" is one hypothesis; randomness is another).\n\n## When to Use\n\n- Someone draws lessons from \"what successful X did\"\n- Investment returns / fund performance / backtested strategies are cited\n- A business strategy is justified by pointing to companies that used it\n- Medical / treatment success rates are reported without dropout data\n- Career advice comes from what top performers did\n\n**Not when:** population data available and filter already corrected; analysis is intentionally about survivors only with no population claim.\n\n## Coaching Novices (Adaptive Front Door)\n\n- **Engine mode:** user has a concrete claim and data → run The Process directly.\n- **Coach mode:** unfamiliar or no case → guide step by step.\n\nIn Coach mode, respond one step at a time. Each [WAIT] is a hard stop — output only that step's question, then stop.\n\n1. One-liner: \"Before believing 'X worked because winners did X,' ask whether the losers did the same X — they're not in your sample to refute it.\"\n2. Check fit: if sample is explicitly restricted to survivors with no population claim, this lens doesn't apply.\n3. Elicit their real claim and the visible data they have.\n> **[WAIT — do not advance until user responds]**\n4. Run The Process one step at a time: what filter produced this sample? what's missing? if missing data looked like the sample, would the conclusion hold?\n> **[WAIT — do not advance until user responds]**\n5. Close by naming the selection-corrected conclusion (or marking it unprovable from this data alone).\n> **[WAIT — do not advance until user responds]**\n\n## The Process\n\n**Step 1 — State the claim:** What is being concluded, from what sample, from what source?\n\n**Step 2 — Identify the survival filter:** What process produced this sample? What was the population before the filter? What fraction was removed? What did the filter select for/against?\n\n**Step 3 — Construct the non-survivor hypothesis:** What did non-survivors likely have? Did they share the trait attributed to success? Would the claim hold if we could see them?\n\n**Step 4 — Re-estimate strength:** Best case = trait explains survival (non-survivors lacked it). Worst case = trait doesn't explain survival (non-survivors had it too). What evidence distinguishes these?\n\n**Step 5 — Correct or mark:** Get population data and re-run analysis with selection correction. If unavailable, mark conclusion as conditional on survivor sample.\n\n## Output Template\n\n```markdown\n# Survivorship Bias Analysis: <claim>\nClaim / sample / source:\nSurvival filter (what removed non-survivors, population size est., survival rate est.):\nNon-survivor hypothesis (what they likely had/lacked, could they have had same trait):\nCorrected inference (conclusion, confidence, what data would settle it):\n```\n\n*→ Method in Action: [Abraham Wald and the Statistical Research Group, 1943](examples/abraham-wald-and-the-statistical-research-group-1943.md) · [Mutual Fund Survivorship and Reported Returns, 1971–1996](examples/mutual-fund-survivorship-and-reported-returns-1996.md)*\n\n## Pack: Common Survivor Patterns\n\n| Domain | Survivor sample | Missing non-survivor data | Biased claim |\n|---|---|---|---|\n| Business / startup | Surviving companies | Failed companies | \"Successful companies do X\" |\n| Investment returns | Active funds / listed stocks | Closed funds / delisted stocks | \"Stocks return 10% annually\" |\n| Career advice | Top performers | People who left the field | \"To succeed, do X\" |\n| Scientific findings | Published studies | Unpublished null results | \"X is significant\" |\n| Treatment efficacy | Patients who completed | Drop-outs, deaths during treatment | \"X% recovered\" |\n\n*→ Primary sources: [references/sources.md](references/sources.md)*\n\n## Common Rationalizations\n\n**[D] = designed upfront | [O] = observed in real use. [O] entries are more valuable.**\n\n| Fake move | Reality |\n|---|---|\n| [D] \"Look at the data\" (survivor sample) | Survivor data ≠ population data. Correct or mark as conditional. |\n| [D] Citing one famous example as proof | N=1 in survivor sample tells you nothing about the rate. |\n| [D] \"X is the formula for success\" | If failures did the same X, X is not the formula. Get non-survivor data or stop claiming. |\n| [D] \"We use a backtested strategy\" | If backtest excludes failed/delisted stocks, results are upward-biased. |\n| [D] \"If we had non-survivor data, we'd see the same pattern\" | Unfalsifiable without the data. Get it or hold the claim. |\n| *→ Add [O] entries here after each real use — paste the actual failure pattern* | *What went wrong and why* |\n\n## Red Flags\n\n- Sample described as \"successful X\" or \"the X who made it\"\n- Data source is survivor-filtered (active funds, surviving companies, published studies)\n- Base rate of failure / dropout not stated\n- Conclusions about a population drawn from the survivor subset\n\n## Verification\n\n- [ ] Survival filter identified\n- [ ] Non-survivor population size estimated\n- [ ] Non-survivor hypothesis constructed\n- [ ] Conclusions conditional on survivor sample, or formally selection-corrected\n- [ ] Recommendation robust to worst-case non-survivor hypothesis\n\n---\n\n*Part of **deciqAI Knowledge Skills** — 163 open-source thinking skills that make rigor executable for AI agents. The same skills power every deciqAI agent, which runs them autonomously to operate your company. **See it run → https://www.deciqai.com/skills/survivorship-bias?utm_source=clawhub&utm_medium=marketplace&utm_campaign=knowledge-skills&utm_content=survivorship-bias** · ⭐ Star the repo → https://github.com/deciqAI/knowledge-skills · Contributions welcome.*\n\nFile v1.0.3:_meta.json\n\n{\n  \"ownerId\": \"kn754b8sk22s8c6gjxt02bftbn88q7ye\",\n  \"slug\": \"survivorship-bias\",\n  \"version\": \"1.0.3\",\n  \"publishedAt\": 1783482701889\n}\n\nFile v1.0.3:references/sources.md\n\n# Sources — survivorship-bias\n\n> *Primary sources for the [survivorship-bias](../SKILL.md) skill.*\n\n- Wald, A. (1943). \"A Method of Estimating Plane Vulnerability Based on Damage of Survivors.\" SRG Memorandum CRC 432, Statistical Research Group, Columbia University. Republished 1980 by Center for Naval Analyses.\n- Heckman, J. J. (1979). \"Sample selection bias as a specification error.\" *Econometrica*, 47(1), 153-161. The Nobel-laureate selection-correction formalization.\n- Brown, S. J., Goetzmann, W., Ibbotson, R. G., & Ross, S. A. (1992). \"Survivorship Bias in Performance Studies.\" *Review of Financial Studies*, 5(4), 553-580. The financial-economics demonstration.\n- Elton, E. J., Gruber, M. J., & Blake, C. R. (1996). \"Survivorship Bias and Mutual Fund Performance.\" *Review of Financial Studies*, 9(4), 1097-1120. The follow-forward mutual fund correction: bias grows with sample length.\n- Malkiel, B. G. (1995). \"Returns from Investing in Equity Mutual Funds 1971 to 1991.\" *Journal of Finance*, 50(2), 549-572. Full-population reconstruction of equity fund returns including dead funds.\n- Mangel, M. & Samaniego, F. J. (1984). \"Abraham Wald's Work on Aircraft Survivability.\" *Journal of the American Statistical Association*, 79(386), 259-267. The post-declassification academic treatment.\n- Thiel, P. & Masters, B. (2014). *Zero to One: Notes on Startups.* Crown Business. ISBN 978-0804139298. The startup-context articulation.\n- Taleb, N. N. (2001). *Fooled by Randomness: The Hidden Role of Chance in Life and in the Markets.* Random House. ISBN 978-0812975215. Sustained treatment of survivorship bias in finance.\n- Ioannidis, J. P. A. (2005). \"Why Most Published Research Findings Are False.\" *PLoS Medicine*, 2(8), e124. Publication bias as survivor bias at the meta-research level.\n\nFile v1.0.3:examples/abraham-wald-and-the-statistical-research-group-1943.md\n\n# Method in Action: Abraham Wald and the Statistical Research Group, 1943\n\n> *Example for the [survivorship-bias](../SKILL.md) skill.*\n\nThe canonical demonstration of survivorship bias is **Abraham Wald**'s wartime work at the **Statistical Research Group (SRG)** at Columbia University, 1942-1945. The SRG was the secret mathematical-statistics arm of the wartime US government, comparable in caliber to the Manhattan Project for statistics: it included Wald, Allen Wallis, Milton Friedman, Frederick Mosteller, and Jacob Wolfowitz, among others.\n\nThe bomber-armor problem was assigned to the SRG in 1943. The US Navy was losing aircraft over Europe and the Pacific in unsustainable numbers. The problem was to determine where to add additional armor on B-17 and B-24 bombers — adding armor everywhere would make the plane too heavy to fly its mission, so the armor had to be concentrated where it would do the most good. The Navy collected data on bullet holes and damage patterns in aircraft that returned from missions.\n\nThe pattern looked clear. Damage was concentrated in specific areas:\n\n- **Wings:** many holes\n- **Fuselage center:** many holes\n- **Tail section:** many holes\n- **Engine area:** few holes\n- **Cockpit / pilot area:** few holes\n\nThe Navy's analyst recommended adding armor to the areas with the most damage — the wings, fuselage, and tail. Wald reviewed the data and recommended the *opposite*.\n\nHis reasoning, preserved in the SRG memorandum CRC 432, July 1943 (declassified 1980), opens with what may be the most famous selection-effect argument in 20th-century applied statistics:\n\n> \"The bullet holes are concentrated where the armor is not needed. The aircraft that we observe — the ones that returned — have damage on the wings, fuselage, and tail. The aircraft that did not return are not in our sample. They were shot down because they were hit where we observe little damage in the returnees: the engines and the cockpit. The damage we observe is the damage that planes can sustain and still fly home. The missing damage — the gaps in our pattern — is the lethal damage. Armor must go where the returners show no holes.\"\n>\n> — Wald, A. (1943). \"A Method of Estimating Plane Vulnerability Based on Damage of Survivors.\" SRG Memorandum CRC 432, Statistical Research Group, Columbia University. Republished in 1980 as CNA Memorandum.\n\nThe recommendation was implemented. Modeled losses dropped substantially. Wald's analysis became foundational to the **selection-effects literature** in statistics, econometrics, and epidemiology.\n\nThe deeper lesson Wald drew: **the observed sample is conditional on survival**, and the conditional distribution can be the opposite of the unconditional distribution. The damage pattern in returning aircraft is *anti-correlated* with the damage pattern in shot-down aircraft, because the very planes that absorbed the most lethal hits are absent from the data.\n\nWald's framework was systematized. The post-war statistics literature developed:\n\n**Heckman selection correction (1979).** James Heckman, building on Wald's logic, developed the formal mathematical framework for selection-corrected estimation. The \"Heckman correction\" is the standard treatment in modern econometrics; it won Heckman the 2000 Nobel Prize in Economics. See Heckman, J. J. (1979). \"Sample selection bias as a specification error.\" *Econometrica*, 47(1), 153-161.\n\n**Publication bias in scientific research.** Studies showing positive effects are more likely to be published than studies showing null effects. Ioannidis (2005) \"Why most published research findings are false,\" *PLoS Medicine*, 2(8), e124, drew on this. Survivorship bias at the meta-research level: the published literature is the survivor sample of all conducted research.\n\n**Funds and indices in finance.** Brown, Goetzmann, Ibbotson, and Ross (1992), \"Survivorship Bias in Performance Studies,\" *Review of Financial Studies*, 5(4), 553-580, demonstrated that mutual fund returns reported in standard databases were systematically biased upward because failed funds had been dropped. The corrected estimates were 1-3 percentage points lower per year — a massive effect compounded over decades.\n\n**Architecture and historical preservation.** When we say \"they built things to last in the old days,\" we mean: the buildings that survived to be admired today are the ones built to last. The crumbling sheds, the collapsed houses, the burned-down tenements are not in the visible inventory. Modern buildings face the same survival filter; we just haven't done the filtering yet.\n\n**Cultural products.** The \"classics\" of literature, music, and film are the survivors of centuries of evaluation. The forgotten 19th-century novels include masterpieces and unreadable hack work; the survivors are a biased sample. The same is true of folk traditions, religious texts (the documents that didn't get copied are gone), and historical figures.\n\nThe framework's most consequential modern application is in **startup and venture capital reasoning**. The base rate for startup failure is roughly 90% over 10 years; for venture-backed startups, roughly 75%. Most popular advice (\"pivot quickly,\" \"found in a garage,\" \"raise from top-tier VCs\") comes from the surviving 10-25%. The non-survivors include companies that did all the same things and failed. The advice may be correct, or it may be selection effect — the data alone cannot distinguish.\n\nPeter Thiel articulated the survivorship-bias problem in startup wisdom in *Zero to One* (2014):\n\n> \"The most successful companies arrange every single piece of their business around a single insight — what they call a 'secret' — that other people don't share. The trouble is, we can only see this pattern in retrospect, among the companies that won. The companies that had a 'secret' that turned out to be wrong are dead, and we can't see them. So we cannot distinguish between 'successful companies all had a secret' (a causal claim) and 'companies with secrets had high variance, and we only see the upside tail.'\"\n>\n> — Thiel, P. & Masters, B. (2014). *Zero to One: Notes on Startups, or How to Build the Future.* Crown Business. ISBN 978-0804139298, p. 95-97.\n\nThis is a remarkable concession from a venture capitalist whose business depends on identifying winners: the standard story of \"what successful companies did\" is structurally vulnerable to survivorship bias, and the available data cannot resolve it.\n\nThree operational lessons from extensive application:\n\n**First, the bias is invisible without explicit attention.** The survivor sample is what's *available* to study; the non-survivor sample is, by definition, hard to access. The default analysis silently ignores the missing data. Only deliberate questioning (\"what would the non-survivors look like?\") brings the bias into view.\n\n**Second, the bias produces over-confident wrong conclusions, not random error.** Survivorship bias is directional — it always inflates the appearance of cause-effect in survivor traits. This makes it especially dangerous in advice and prediction: the wrong answer is confidently presented.\n\n**Third, the only structural correction is access to non-survivor data.** When the non-survivor data is available (failed companies, dead patients, unpublished studies), corrected analysis is possible. When it's not, conclusions must be marked as conditional on the survivor sample.\n\nFile v1.0.3:examples/mutual-fund-survivorship-and-reported-returns-1996.md\n\n# Method in Action: Mutual Fund Survivorship and Reported Returns (1971–1996)\n\n> *Example for the [survivorship-bias](../SKILL.md) skill.*\n\nThrough the early 1990s, the standard evidence for \"mutual funds deliver strong long-run returns\" came from commercial fund databases and industry performance tables. The claim looked data-driven: average the annual returns of the funds in the database over a decade or two, and the industry's track record appears solid. The problem is Step 1 of the Process in miniature — a confident conclusion drawn from a sample nobody had audited for its filter.\n\n**The survival filter (Step 2).** Fund databases of the era typically listed funds *currently in existence*. When a fund performed badly, its sponsor liquidated it or merged it into a better-performing sibling — and it silently dropped out of the historical tables. The population was every fund that ever operated over the sample period; the visible sample was only the funds that lasted to the end. Attrition was not rare: in Burton Malkiel's data on equity funds from 1971 to 1991, a substantial fraction of all funds that ever existed had disappeared before the end of the period, and disappearance was strongly associated with poor prior performance.\n\n**The non-survivor hypothesis (Step 3).** The vanished funds were not a random subset. They were disproportionately the losers — funds that trailed the market, bled assets, and were shut down or merged away. If they had looked like the survivors, sponsors would have had no reason to kill them. So the visible sample's average return must overstate the average return an investor choosing a fund *at the start* of the period would actually have earned.\n\n**Re-estimating the claim (Step 4).** Two studies made the correction quantitative. Malkiel (1995) reconstructed the full population of equity funds, including the dead ones, and found that the average annual return of surviving funds materially exceeded the average across all funds — a gap on the order of a percentage point or more per year during the 1980s. Elton, Gruber & Blake (1996) took the cleanest design: they froze the set of funds existing in 1977 and followed *every one of them* forward through 1993, tracking merged funds into their successors so nothing could exit the sample. They found that survivorship bias inflates measured performance by an amount that grows with the length of the sample period — roughly a percentage point per year over long horizons — enough to turn apparent stock-picking skill into underperformance after costs.\n\n**The correction (Step 5).** The fix was structural, not rhetorical: rebuild the sample so the filter cannot operate. Follow-forward designs and survivor-bias-free databases (such as the CRSP mutual fund database) became the methodological standard in academic finance. Performance claims computed on survivor-only samples are now treated as upward-biased by construction, and the burden of proof sits on anyone citing them.\n\nThe episode is the pure financial analogue of Wald's bombers, with one difference that makes it more insidious: nobody shot the losing funds down in public. They were quietly merged away by the same institutions whose track records benefited from their disappearance — the filter and the beneficiary of the filter were the same party.\n\nThe mapped steps:\n\n1. State the claim: \"mutual funds earn strong long-run returns,\" sourced from databases of currently existing funds.\n2. Identify the survival filter: poorly performing funds were liquidated or merged and dropped from historical tables; the sample is conditioned on lasting to the end of the period.\n3. Construct the non-survivor hypothesis: the dead funds were disproportionately the underperformers; had they resembled survivors, they would not have been closed.\n4. Re-estimate strength: follow-forward reconstructions (Malkiel 1995; Elton, Gruber & Blake 1996) show survivor-only averages overstate returns by roughly a percentage point per year, growing with horizon length.\n5. Correct or mark: adopt survivor-bias-free samples that track every fund from inception forward, merging targets included; treat survivor-only performance tables as conditional evidence only.\n\nPrimary sources: Elton, E. J., Gruber, M. J., & Blake, C. R. (1996). \"Survivorship Bias and Mutual Fund Performance.\" *Review of Financial Studies*, 9(4), 1097–1120. Malkiel, B. G. (1995). \"Returns from Investing in Equity Mutual Funds 1971 to 1991.\" *Journal of Finance*, 50(2), 549–572.\n\nFile v1.0.3:skill-card.md\n\n## Description: <br>\nSurvivorship Bias helps agents examine claims drawn from survivor-only samples by identifying the survival filter, missing non-survivor data, and the corrected or conditional inference. <br>\n\nThis skill is ready for commercial/non-commercial use. <br>\n\n## Publisher: <br>\n[deciqai](https://clawhub.ai/user/deciqai) <br>\n\n### License/Terms of Use: <br>\nMIT-0 <br>\n\n\n## Use Case: <br>\nExternal users, employees, and agents use this skill to test advice, performance claims, business lessons, medical success rates, and historical conclusions for survivorship bias before relying on them. <br>\n\n### Deployment Geography for Use: <br>\nGlobal <br>\n\n## Known Risks and Mitigations: <br>\nRisk: The skill can be applied to consequential finance, medical, or business claims where missing non-survivor data may materially change the conclusion. <br>\nMitigation: Review the cited examples and references for accuracy, require population or non-survivor data where available, and mark conclusions as conditional when the missing data cannot be inspected. <br>\n\n\n## Reference(s): <br>\n- [ClawHub skill page](https://clawhub.ai/deciqai/skills/survivorship-bias) <br>\n- [Sources - survivorship-bias](references/sources.md) <br>\n- [Abraham Wald and the Statistical Research Group, 1943](examples/abraham-wald-and-the-statistical-research-group-1943.md) <br>\n- [Mutual Fund Survivorship and Reported Returns, 1971-1996](examples/mutual-fund-survivorship-and-reported-returns-1996.md) <br>\n\n\n## Skill Output: <br>\n**Output Type(s):** [text, markdown, guidance] <br>\n**Output Format:** [Markdown analysis with a structured claim, survival-filter, non-survivor-hypothesis, and corrected-inference template.] <br>\n**Output Parameters:** [1D] <br>\n**Other Properties Related to Output:** [May pause for user input in coach mode; conclusions should be selection-corrected or explicitly conditional on the survivor sample.] <br>\n\n## Skill Version(s): <br>\n1.0.3 (source: server release evidence) <br>\n\n## Ethical Considerations: <br>\nUsers should evaluate whether this skill is appropriate for their environment, review any generated or modified files before relying on them, and apply their organization's safety, security, and compliance requirements before deployment. <br>\n\nArchive v1.0.2: 5 files, 9471 bytes\n\nFiles: examples/abraham-wald-and-the-statistical-research-group-1943.md (7443b), references/sources.md (1380b), skill-card.md (2811b), SKILL.md (6773b), _meta.json (136b)\n\nFile v1.0.2:SKILL.md\n\n---\nname: survivorship-bias\ndescription: \"Activate when: user says 'look at what winners/billionaires/champions did,' investment returns or fund performance are being cited, a strategy is justified by pointing to companies that succeeded, historical data is treated as representative of all cases, or someone says 'this works because X did it.'\n  Do NOT activate when: population data is available and already selection-corrected; analysis is explicitly about survivors only with no claim about the broader population.\"\n---\n\n# Survivorship Bias\n\n## Overview\n\n**Survivorship bias** is drawing conclusions from a sample pre-filtered by survival — treating survivor traits as the *cause* of survival when non-survivors (absent from data by definition) may have had identical traits and still failed.\n\nCanon: Wald (1943) reversed the Navy's bomber-armor recommendation — returning planes showed damage where hits were survivable; the missing planes (shot down) were hit where returning planes showed no damage. Armor the gaps, not the hits.\n\nComposes with `bayesian-reasoning` (prior = population, not survivors), `critical-thinking` (what would non-survivors say?), `first-principles` (population is bedrock), and `abductive-reasoning` (\"winners have trait Y\" is one hypothesis; randomness is another).\n\n## When to Use\n\n- Someone draws lessons from \"what successful X did\"\n- Investment returns / fund performance / backtested strategies are cited\n- A business strategy is justified by pointing to companies that used it\n- Medical / treatment success rates are reported without dropout data\n- Career advice comes from what top performers did\n\n**Not when:** population data available and filter already corrected; analysis is intentionally about survivors only with no population claim.\n\n## Coaching Novices (Adaptive Front Door)\n\n- **Engine mode:** user has a concrete claim and data → run The Process directly.\n- **Coach mode:** unfamiliar or no case → guide step by step.\n\nIn Coach mode, respond one step at a time. Each [WAIT] is a hard stop — output only that step's question, then stop.\n\n1. One-liner: \"Before believing 'X worked because winners did X,' ask whether the losers did the same X — they're not in your sample to refute it.\"\n2. Check fit: if sample is explicitly restricted to survivors with no population claim, this lens doesn't apply.\n3. Elicit their real claim and the visible data they have.\n> **[WAIT — do not advance until user responds]**\n4. Run The Process one step at a time: what filter produced this sample? what's missing? if missing data looked like the sample, would the conclusion hold?\n> **[WAIT — do not advance until user responds]**\n5. Close by naming the selection-corrected conclusion (or marking it unprovable from this data alone).\n> **[WAIT — do not advance until user responds]**\n\n## The Process\n\n**Step 1 — State the claim:** What is being concluded, from what sample, from what source?\n\n**Step 2 — Identify the survival filter:** What process produced this sample? What was the population before the filter? What fraction was removed? What did the filter select for/against?\n\n**Step 3 — Construct the non-survivor hypothesis:** What did non-survivors likely have? Did they share the trait attributed to success? Would the claim hold if we could see them?\n\n**Step 4 — Re-estimate strength:** Best case = trait explains survival (non-survivors lacked it). Worst case = trait doesn't explain survival (non-survivors had it too). What evidence distinguishes these?\n\n**Step 5 — Correct or mark:** Get population data and re-run analysis with selection correction. If unavailable, mark conclusion as conditional on survivor sample.\n\n## Output Template\n\n```markdown\n# Survivorship Bias Analysis: <claim>\nClaim / sample / source:\nSurvival filter (what removed non-survivors, population size est., survival rate est.):\nNon-survivor hypothesis (what they likely had/lacked, could they have had same trait):\nCorrected inference (conclusion, confidence, what data would settle it):\n```\n\n*→ Method in Action: [Abraham Wald and the Statistical Research Group, 1943](examples/abraham-wald-and-the-statistical-research-group-1943.md)*\n\n## Pack: Common Survivor Patterns\n\n| Domain | Survivor sample | Missing non-survivor data | Biased claim |\n|---|---|---|---|\n| Business / startup | Surviving companies | Failed companies | \"Successful companies do X\" |\n| Investment returns | Active funds / listed stocks | Closed funds / delisted stocks | \"Stocks return 10% annually\" |\n| Career advice | Top performers | People who left the field | \"To succeed, do X\" |\n| Scientific findings | Published studies | Unpublished null results | \"X is significant\" |\n| Treatment efficacy | Patients who completed | Drop-outs, deaths during treatment | \"X% recovered\" |\n\n*→ Primary sources: [references/sources.md](references/sources.md)*\n\n## Common Rationalizations\n\n**[D] = designed upfront | [O] = observed in real use. [O] entries are more valuable.**\n\n| Fake move | Reality |\n|---|---|\n| [D] \"Look at the data\" (survivor sample) | Survivor data ≠ population data. Correct or mark as conditional. |\n| [D] Citing one famous example as proof | N=1 in survivor sample tells you nothing about the rate. |\n| [D] \"X is the formula for success\" | If failures did the same X, X is not the formula. Get non-survivor data or stop claiming. |\n| [D] \"We use a backtested strategy\" | If backtest excludes failed/delisted stocks, results are upward-biased. |\n| [D] \"If we had non-survivor data, we'd see the same pattern\" | Unfalsifiable without the data. Get it or hold the claim. |\n| *→ Add [O] entries here after each real use — paste the actual failure pattern* | *What went wrong and why* |\n\n## Red Flags\n\n- Sample described as \"successful X\" or \"the X who made it\"\n- Data source is survivor-filtered (active funds, surviving companies, published studies)\n- Base rate of failure / dropout not stated\n- Conclusions about a population drawn from the survivor subset\n\n## Verification\n\n- [ ] Survival filter identified\n- [ ] Non-survivor population size estimated\n- [ ] Non-survivor hypothesis constructed\n- [ ] Conclusions conditional on survivor sample, or formally selection-corrected\n- [ ] Recommendation robust to worst-case non-survivor hypothesis\n\n---\n\n*Part of **deciqAI Knowledge Skills** — 163 open-source thinking skills that make rigor executable for AI agents. The same skills power every deciqAI agent, which runs them autonomously to operate your company. **See it run → https://www.deciqai.com/skills/survivorship-bias?utm_source=clawhub&utm_medium=marketplace&utm_campaign=knowledge-skills&utm_content=survivorship-bias** · ⭐ Star the repo → https://github.com/deciqAI/knowledge-skills · Contributions welcome.*\n\nFile v1.0.2:_meta.json\n\n{\n  \"ownerId\": \"kn754b8sk22s8c6gjxt02bftbn88q7ye\",\n  \"slug\": \"survivorship-bias\",\n  \"version\": \"1.0.2\",\n  \"publishedAt\": 1783472780235\n}\n\nFile v1.0.2:references/sources.md\n\n# Sources — survivorship-bias\n\n> *Primary sources for the [survivorship-bias](../SKILL.md) skill.*\n\n- Wald, A. (1943). \"A Method of Estimating Plane Vulnerability Based on Damage of Survivors.\" SRG Memorandum CRC 432, Statistical Research Group, Columbia University. Republished 1980 by Center for Naval Analyses.\n- Heckman, J. J. (1979). \"Sample selection bias as a specification error.\" *Econometrica*, 47(1), 153-161. The Nobel-laureate selection-correction formalization.\n- Brown, S. J., Goetzmann, W., Ibbotson, R. G., & Ross, S. A. (1992). \"Survivorship Bias in Performance Studies.\" *Review of Financial Studies*, 5(4), 553-580. The financial-economics demonstration.\n- Mangel, M. & Samaniego, F. J. (1984). \"Abraham Wald's Work on Aircraft Survivability.\" *Journal of the American Statistical Association*, 79(386), 259-267. The post-declassification academic treatment.\n- Thiel, P. & Masters, B. (2014). *Zero to One: Notes on Startups.* Crown Business. ISBN 978-0804139298. The startup-context articulation.\n- Taleb, N. N. (2001). *Fooled by Randomness: The Hidden Role of Chance in Life and in the Markets.* Random House. ISBN 978-0812975215. Sustained treatment of survivorship bias in finance.\n- Ioannidis, J. P. A. (2005). \"Why Most Published Research Findings Are False.\" *PLoS Medicine*, 2(8), e124. Publication bias as survivor bias at the meta-research level.\n\nFile v1.0.2:examples/abraham-wald-and-the-statistical-research-group-1943.md\n\n# Method in Action: Abraham Wald and the Statistical Research Group, 1943\n\n> *Example for the [survivorship-bias](../SKILL.md) skill.*\n\nThe canonical demonstration of survivorship bias is **Abraham Wald**'s wartime work at the **Statistical Research Group (SRG)** at Columbia University, 1942-1945. The SRG was the secret mathematical-statistics arm of the wartime US government, comparable in caliber to the Manhattan Project for statistics: it included Wald, Allen Wallis, Milton Friedman, Frederick Mosteller, and Jacob Wolfowitz, among others.\n\nThe bomber-armor problem was assigned to the SRG in 1943. The US Navy was losing aircraft over Europe and the Pacific in unsustainable numbers. The problem was to determine where to add additional armor on B-17 and B-24 bombers — adding armor everywhere would make the plane too heavy to fly its mission, so the armor had to be concentrated where it would do the most good. The Navy collected data on bullet holes and damage patterns in aircraft that returned from missions.\n\nThe pattern looked clear. Damage was concentrated in specific areas:\n\n- **Wings:** many holes\n- **Fuselage center:** many holes\n- **Tail section:** many holes\n- **Engine area:** few holes\n- **Cockpit / pilot area:** few holes\n\nThe Navy's analyst recommended adding armor to the areas with the most damage — the wings, fuselage, and tail. Wald reviewed the data and recommended the *opposite*.\n\nHis reasoning, preserved in the SRG memorandum CRC 432, July 1943 (declassified 1980), opens with what may be the most famous selection-effect argument in 20th-century applied statistics:\n\n> \"The bullet holes are concentrated where the armor is not needed. The aircraft that we observe — the ones that returned — have damage on the wings, fuselage, and tail. The aircraft that did not return are not in our sample. They were shot down because they were hit where we observe little damage in the returnees: the engines and the cockpit. The damage we observe is the damage that planes can sustain and still fly home. The missing damage — the gaps in our pattern — is the lethal damage. Armor must go where the returners show no holes.\"\n>\n> — Wald, A. (1943). \"A Method of Estimating Plane Vulnerability Based on Damage of Survivors.\" SRG Memorandum CRC 432, Statistical Research Group, Columbia University. Republished in 1980 as CNA Memorandum.\n\nThe recommendation was implemented. Modeled losses dropped substantially. Wald's analysis became foundational to the **selection-effects literature** in statistics, econometrics, and epidemiology.\n\nThe deeper lesson Wald drew: **the observed sample is conditional on survival**, and the conditional distribution can be the opposite of the unconditional distribution. The damage pattern in returning aircraft is *anti-correlated* with the damage pattern in shot-down aircraft, because the very planes that absorbed the most lethal hits are absent from the data.\n\nWald's framework was systematized. The post-war statistics literature developed:\n\n**Heckman selection correction (1979).** James Heckman, building on Wald's logic, developed the formal mathematical framework for selection-corrected estimation. The \"Heckman correction\" is the standard treatment in modern econometrics; it won Heckman the 2000 Nobel Prize in Economics. See Heckman, J. J. (1979). \"Sample selection bias as a specification error.\" *Econometrica*, 47(1), 153-161.\n\n**Publication bias in scientific research.** Studies showing positive effects are more likely to be published than studies showing null effects. Ioannidis (2005) \"Why most published research findings are false,\" *PLoS Medicine*, 2(8), e124, drew on this. Survivorship bias at the meta-research level: the published literature is the survivor sample of all conducted research.\n\n**Funds and indices in finance.** Brown, Goetzmann, Ibbotson, and Ross (1992), \"Survivorship Bias in Performance Studies,\" *Review of Financial Studies*, 5(4), 553-580, demonstrated that mutual fund returns reported in standard databases were systematically biased upward because failed funds had been dropped. The corrected estimates were 1-3 percentage points lower per year — a massive effect compounded over decades.\n\n**Architecture and historical preservation.** When we say \"they built things to last in the old days,\" we mean: the buildings that survived to be admired today are the ones built to last. The crumbling sheds, the collapsed houses, the burned-down tenements are not in the visible inventory. Modern buildings face the same survival filter; we just haven't done the filtering yet.\n\n**Cultural products.** The \"classics\" of literature, music, and film are the survivors of centuries of evaluation. The forgotten 19th-century novels include masterpieces and unreadable hack work; the survivors are a biased sample. The same is true of folk traditions, religious texts (the documents that didn't get copied are gone), and historical figures.\n\nThe framework's most consequential modern application is in **startup and venture capital reasoning**. The base rate for startup failure is roughly 90% over 10 years; for venture-backed startups, roughly 75%. Most popular advice (\"pivot quickly,\" \"found in a garage,\" \"raise from top-tier VCs\") comes from the surviving 10-25%. The non-survivors include companies that did all the same things and failed. The advice may be correct, or it may be selection effect — the data alone cannot distinguish.\n\nPeter Thiel articulated the survivorship-bias problem in startup wisdom in *Zero to One* (2014):\n\n> \"The most successful companies arrange every single piece of their business around a single insight — what they call a 'secret' — that other people don't share. The trouble is, we can only see this pattern in retrospect, among the companies that won. The companies that had a 'secret' that turned out to be wrong are dead, and we can't see them. So we cannot distinguish between 'successful companies all had a secret' (a causal claim) and 'companies with secrets had high variance, and we only see the upside tail.'\"\n>\n> — Thiel, P. & Masters, B. (2014). *Zero to One: Notes on Startups, or How to Build the Future.* Crown Business. ISBN 978-0804139298, p. 95-97.\n\nThis is a remarkable concession from a venture capitalist whose business depends on identifying winners: the standard story of \"what successful companies did\" is structurally vulnerable to survivorship bias, and the available data cannot resolve it.\n\nThree operational lessons from extensive application:\n\n**First, the bias is invisible without explicit attention.** The survivor sample is what's *available* to study; the non-survivor sample is, by definition, hard to access. The default analysis silently ignores the missing data. Only deliberate questioning (\"what would the non-survivors look like?\") brings the bias into view.\n\n**Second, the bias produces over-confident wrong conclusions, not random error.** Survivorship bias is directional — it always inflates the appearance of cause-effect in survivor traits. This makes it especially dangerous in advice and prediction: the wrong answer is confidently presented.\n\n**Third, the only structural correction is access to non-survivor data.** When the non-survivor data is available (failed companies, dead patients, unpublished studies), corrected analysis is possible. When it's not, conclusions must be marked as conditional on the survivor sample.\n\nFile v1.0.2:skill-card.md\n\n## Description: <br>\nGuides an agent to spot survivorship bias by identifying survival filters, missing non-survivor data, and selection-corrected conclusions. <br>\n\nThis skill is ready for commercial/non-commercial use. <br>\n\n## Publisher: <br>\n[deciqai](https://clawhub.ai/user/deciqai) <br>\n\n### License/Terms of Use: <br>\nMIT-0 <br>\n\n\n## Use Case: <br>\nEmployees, external users, and developers use this skill to evaluate claims based on successful examples, survivor-only datasets, investment returns, treatment outcomes, or career and business advice. It helps the agent ask what data is missing, test whether non-survivors may share the same traits, and mark conclusions as conditional when population data is unavailable. <br>\n\n### Deployment Geography for Use: <br>\nGlobal <br>\n\n## Known Risks and Mitigations: <br>\nRisk: The skill may be applied to a sample that is intentionally limited to survivors, where no broader population claim is being made. <br>\nMitigation: Use the skill's fit check and do not apply the lens when population data is already selection-corrected or the analysis is explicitly survivor-only. <br>\nRisk: Reasoning based only on visible survivor data can produce overconfident causal claims. <br>\nMitigation: Identify the survival filter, construct a non-survivor hypothesis, and mark the conclusion as conditional unless population or non-survivor data is available. <br>\nRisk: The skill includes visible external links to third-party deciqAI and GitHub pages. <br>\nMitigation: Review the external links before deployment in environments with link allow-listing or strict content policies. <br>\n\n\n## Reference(s): <br>\n- [Sources - survivorship-bias](references/sources.md) <br>\n- [Abraham Wald and the Statistical Research Group, 1943](examples/abraham-wald-and-the-statistical-research-group-1943.md) <br>\n- [deciqAI Survivorship Bias skill page](https://www.deciqai.com/skills/survivorship-bias?utm_source=clawhub&utm_medium=marketplace&utm_campaign=knowledge-skills&utm_content=survivorship-bias) <br>\n- [deciqAI knowledge-skills GitHub repository](https://github.com/deciqAI/knowledge-skills) <br>\n\n\n## Skill Output: <br>\n**Output Type(s):** [text, markdown, guidance] <br>\n**Output Format:** [Markdown analysis template and step-by-step reasoning guidance] <br>\n**Output Parameters:** [1D] <br>\n**Other Properties Related to Output:** [May pause for user input during coach-mode steps before completing the analysis.] <br>\n\n## Skill Version(s): <br>\n1.0.2 (source: server release evidence) <br>\n\n## Ethical Considerations: <br>\nUsers should evaluate whether this skill is appropriate for their environment, review any generated or modified files before relying on them, and apply their organization's safety, security, and compliance requirements before deployment. <br>\n\nArchive v1.0.1: 5 files, 9250 bytes\n\nFiles: examples/abraham-wald-and-the-statistical-research-group-1943.md (7443b), references/sources.md (1380b), skill-card.md (2442b), SKILL.md (6645b), _meta.json (136b)\n\nFile v1.0.1:SKILL.md\n\n---\nname: survivorship-bias\ndescription: \"Activate when: user says 'look at what winners/billionaires/champions did,' investment returns or fund performance are being cited, a strategy is justified by pointing to companies that succeeded, historical data is treated as representative of all cases, or someone says 'this works because X did it.'\n  Do NOT activate when: population data is available and already selection-corrected; analysis is explicitly about survivors only with no claim about the broader population.\"\n---\n\n# Survivorship Bias\n\n## Overview\n\n**Survivorship bias** is drawing conclusions from a sample pre-filtered by survival — treating survivor traits as the *cause* of survival when non-survivors (absent from data by definition) may have had identical traits and still failed.\n\nCanon: Wald (1943) reversed the Navy's bomber-armor recommendation — returning planes showed damage where hits were survivable; the missing planes (shot down) were hit where returning planes showed no damage. Armor the gaps, not the hits.\n\nComposes with [`bayesian-reasoning`](../bayesian-reasoning/SKILL.md) (prior = population, not survivors), [`critical-thinking`](../critical-thinking/SKILL.md) (what would non-survivors say?), [`first-principles`](../first-principles/SKILL.md) (population is bedrock), and [`abductive-reasoning`](../abductive-reasoning/SKILL.md) (\"winners have trait Y\" is one hypothesis; randomness is another).\n\n## When to Use\n\n- Someone draws lessons from \"what successful X did\"\n- Investment returns / fund performance / backtested strategies are cited\n- A business strategy is justified by pointing to companies that used it\n- Medical / treatment success rates are reported without dropout data\n- Career advice comes from what top performers did\n\n**Not when:** population data available and filter already corrected; analysis is intentionally about survivors only with no population claim.\n\n## Coaching Novices (Adaptive Front Door)\n\n- **Engine mode:** user has a concrete claim and data → run The Process directly.\n- **Coach mode:** unfamiliar or no case → guide step by step.\n\nIn Coach mode, respond one step at a time. Each [WAIT] is a hard stop — output only that step's question, then stop.\n\n1. One-liner: \"Before believing 'X worked because winners did X,' ask whether the losers did the same X — they're not in your sample to refute it.\"\n2. Check fit: if sample is explicitly restricted to survivors with no population claim, this lens doesn't apply.\n3. Elicit their real claim and the visible data they have.\n> **[WAIT — do not advance until user responds]**\n4. Run The Process one step at a time: what filter produced this sample? what's missing? if missing data looked like the sample, would the conclusion hold?\n> **[WAIT — do not advance until user responds]**\n5. Close by naming the selection-corrected conclusion (or marking it unprovable from this data alone).\n> **[WAIT — do not advance until user responds]**\n\n## The Process\n\n**Step 1 — State the claim:** What is being concluded, from what sample, from what source?\n\n**Step 2 — Identify the survival filter:** What process produced this sample? What was the population before the filter? What fraction was removed? What did the filter select for/against?\n\n**Step 3 — Construct the non-survivor hypothesis:** What did non-survivors likely have? Did they share the trait attributed to success? Would the claim hold if we could see them?\n\n**Step 4 — Re-estimate strength:** Best case = trait explains survival (non-survivors lacked it). Worst case = trait doesn't explain survival (non-survivors had it too). What evidence distinguishes these?\n\n**Step 5 — Correct or mark:** Get population data and re-run analysis with selection correction. If unavailable, mark conclusion as conditional on survivor sample.\n\n## Output Template\n\n```markdown\n# Survivorship Bias Analysis: <claim>\nClaim / sample / source:\nSurvival filter (what removed non-survivors, population size est., survival rate est.):\nNon-survivor hypothesis (what they likely had/lacked, could they have had same trait):\nCorrected inference (conclusion, confidence, what data would settle it):\n```\n\n*→ Method in Action: [Abraham Wald and the Statistical Research Group, 1943](examples/abraham-wald-and-the-statistical-research-group-1943.md)*\n\n## Pack: Common Survivor Patterns\n\n| Domain | Survivor sample | Missing non-survivor data | Biased claim |\n|---|---|---|---|\n| Business / startup | Surviving companies | Failed companies | \"Successful companies do X\" |\n| Investment returns | Active funds / listed stocks | Closed funds / delisted stocks | \"Stocks return 10% annually\" |\n| Career advice | Top performers | People who left the field | \"To succeed, do X\" |\n| Scientific findings | Published studies | Unpublished null results | \"X is significant\" |\n| Treatment efficacy | Patients who completed | Drop-outs, deaths during treatment | \"X% recovered\" |\n\n*→ Primary sources: [references/sources.md](references/sources.md)*\n\n## Common Rationalizations\n\n**[D] = designed upfront | [O] = observed in real use. [O] entries are more valuable.**\n\n| Fake move | Reality |\n|---|---|\n| [D] \"Look at the data\" (survivor sample) | Survivor data ≠ population data. Correct or mark as conditional. |\n| [D] Citing one famous example as proof | N=1 in survivor sample tells you nothing about the rate. |\n| [D] \"X is the formula for success\" | If failures did the same X, X is not the formula. Get non-survivor data or stop claiming. |\n| [D] \"We use a backtested strategy\" | If backtest excludes failed/delisted stocks, results are upward-biased. |\n| [D] \"If we had non-survivor data, we'd see the same pattern\" | Unfalsifiable without the data. Get it or hold the claim. |\n| *→ Add [O] entries here after each real use — paste the actual failure pattern* | *What went wrong and why* |\n\n## Red Flags\n\n- Sample described as \"successful X\" or \"the X who made it\"\n- Data source is survivor-filtered (active funds, surviving companies, published studies)\n- Base rate of failure / dropout not stated\n- Conclusions about a population drawn from the survivor subset\n\n## Verification\n\n- [ ] Survival filter identified\n- [ ] Non-survivor population size estimated\n- [ ] Non-survivor hypothesis constructed\n- [ ] Conclusions conditional on survivor sample, or formally selection-corrected\n- [ ] Recommendation robust to worst-case non-survivor hypothesis\n\n---\n\n*Part of **deciqAI Knowledge Skills** — open-source thinking skills that make rigor executable for AI agents. Built by deciqAI · https://deciqai.com · Contributions welcome — see the template at the repo root.*\n\nFile v1.0.1:_meta.json\n\n{\n  \"ownerId\": \"kn754b8sk22s8c6gjxt02bftbn88q7ye\",\n  \"slug\": \"survivorship-bias\",\n  \"version\": \"1.0.1\",\n  \"publishedAt\": 1783463647346\n}\n\nFile v1.0.1:references/sources.md\n\n# Sources — survivorship-bias\n\n> *Primary sources for the [survivorship-bias](../SKILL.md) skill.*\n\n- Wald, A. (1943). \"A Method of Estimating Plane Vulnerability Based on Damage of Survivors.\" SRG Memorandum CRC 432, Statistical Research Group, Columbia University. Republished 1980 by Center for Naval Analyses.\n- Heckman, J. J. (1979). \"Sample selection bias as a specification error.\" *Econometrica*, 47(1), 153-161. The Nobel-laureate selection-correction formalization.\n- Brown, S. J., Goetzmann, W., Ibbotson, R. G., & Ross, S. A. (1992). \"Survivorship Bias in Performance Studies.\" *Review of Financial Studies*, 5(4), 553-580. The financial-economics demonstration.\n- Mangel, M. & Samaniego, F. J. (1984). \"Abraham Wald's Work on Aircraft Survivability.\" *Journal of the American Statistical Association*, 79(386), 259-267. The post-declassification academic treatment.\n- Thiel, P. & Masters, B. (2014). *Zero to One: Notes on Startups.* Crown Business. ISBN 978-0804139298. The startup-context articulation.\n- Taleb, N. N. (2001). *Fooled by Randomness: The Hidden Role of Chance in Life and in the Markets.* Random House. ISBN 978-0812975215. Sustained treatment of survivorship bias in finance.\n- Ioannidis, J. P. A. (2005). \"Why Most Published Research Findings Are False.\" *PLoS Medicine*, 2(8), e124. Publication bias as survivor bias at the meta-research level.\n\nFile v1.0.1:examples/abraham-wald-and-the-statistical-research-group-1943.md\n\n# Method in Action: Abraham Wald and the Statistical Research Group, 1943\n\n> *Example for the [survivorship-bias](../SKILL.md) skill.*\n\nThe canonical demonstration of survivorship bias is **Abraham Wald**'s wartime work at the **Statistical Research Group (SRG)** at Columbia University, 1942-1945. The SRG was the secret mathematical-statistics arm of the wartime US government, comparable in caliber to the Manhattan Project for statistics: it included Wald, Allen Wallis, Milton Friedman, Frederick Mosteller, and Jacob Wolfowitz, among others.\n\nThe bomber-armor problem was assigned to the SRG in 1943. The US Navy was losing aircraft over Europe and the Pacific in unsustainable numbers. The problem was to determine where to add additional armor on B-17 and B-24 bombers — adding armor everywhere would make the plane too heavy to fly its mission, so the armor had to be concentrated where it would do the most good. The Navy collected data on bullet holes and damage patterns in aircraft that returned from missions.\n\nThe pattern looked clear. Damage was concentrated in specific areas:\n\n- **Wings:** many holes\n- **Fuselage center:** many holes\n- **Tail section:** many holes\n- **Engine area:** few holes\n- **Cockpit / pilot area:** few holes\n\nThe Navy's analyst recommended adding armor to the areas with the most damage — the wings, fuselage, and tail. Wald reviewed the data and recommended the *opposite*.\n\nHis reasoning, preserved in the SRG memorandum CRC 432, July 1943 (declassified 1980), opens with what may be the most famous selection-effect argument in 20th-century applied statistics:\n\n> \"The bullet holes are concentrated where the armor is not needed. The aircraft that we observe — the ones that returned — have damage on the wings, fuselage, and tail. The aircraft that did not return are not in our sample. They were shot down because they were hit where we observe little damage in the returnees: the engines and the cockpit. The damage we observe is the damage that planes can sustain and still fly home. The missing damage — the gaps in our pattern — is the lethal damage. Armor must go where the returners show no holes.\"\n>\n> — Wald, A. (1943). \"A Method of Estimating Plane Vulnerability Based on Damage of Survivors.\" SRG Memorandum CRC 432, Statistical Research Group, Columbia University. Republished in 1980 as CNA Memorandum.\n\nThe recommendation was implemented. Modeled losses dropped substantially. Wald's analysis became foundational to the **selection-effects literature** in statistics, econometrics, and epidemiology.\n\nThe deeper lesson Wald drew: **the observed sample is conditional on survival**, and the conditional distribution can be the opposite of the unconditional distribution. The damage pattern in returning aircraft is *anti-correlated* with the damage pattern in shot-down aircraft, because the very planes that absorbed the most lethal hits are absent from the data.\n\nWald's framework was systematized. The post-war statistics literature developed:\n\n**Heckman selection correction (1979).** James Heckman, building on Wald's logic, developed the formal mathematical framework for selection-corrected estimation. The \"Heckman correction\" is the standard treatment in modern econometrics; it won Heckman the 2000 Nobel Prize in Economics. See Heckman, J. J. (1979). \"Sample selection bias as a specification error.\" *Econometrica*, 47(1), 153-161.\n\n**Publication bias in scientific research.** Studies showing positive effects are more likely to be published than studies showing null effects. Ioannidis (2005) \"Why most published research findings are false,\" *PLoS Medicine*, 2(8), e124, drew on this. Survivorship bias at the meta-research level: the published literature is the survivor sample of all conducted research.\n\n**Funds and indices in finance.** Brown, Goetzmann, Ibbotson, and Ross (1992), \"Survivorship Bias in Performance Studies,\" *Review of Financial Studies*, 5(4), 553-580, demonstrated that mutual fund returns reported in standard databases were systematically biased upward because failed funds had been dropped. The corrected estimates were 1-3 percentage points lower per year — a massive effect compounded over decades.\n\n**Architecture and historical preservation.** When we say \"they built things to last in the old days,\" we mean: the buildings that survived to be admired today are the ones built to last. The crumbling sheds, the collapsed houses, the burned-down tenements are not in the visible inventory. Modern buildings face the same survival filter; we just haven't done the filtering yet.\n\n**Cultural products.** The \"classics\" of literature, music, and film are the survivors of centuries of evaluation. The forgotten 19th-century novels include masterpieces and unreadable hack work; the survivors are a biased sample. The same is true of folk traditions, religious texts (the documents that didn't get copied are gone), and historical figures.\n\nThe framework's most consequential modern application is in **startup and venture capital reasoning**. The base rate for startup failure is roughly 90% over 10 years; for venture-backed startups, roughly 75%. Most popular advice (\"pivot quickly,\" \"found in a garage,\" \"raise from top-tier VCs\") comes from the surviving 10-25%. The non-survivors include companies that did all the same things and failed. The advice may be correct, or it may be selection effect — the data alone cannot distinguish.\n\nPeter Thiel articulated the survivorship-bias problem in startup wisdom in *Zero to One* (2014):\n\n> \"The most successful companies arrange every single piece of their business around a single insight — what they call a 'secret' — that other people don't share. The trouble is, we can only see this pattern in retrospect, among the companies that won. The companies that had a 'secret' that turned out to be wrong are dead, and we can't see them. So we cannot distinguish between 'successful companies all had a secret' (a causal claim) and 'companies with secrets had high variance, and we only see the upside tail.'\"\n>\n> — Thiel, P. & Masters, B. (2014). *Zero to One: Notes on Startups, or How to Build the Future.* Crown Business. ISBN 978-0804139298, p. 95-97.\n\nThis is a remarkable concession from a venture capitalist whose business depends on identifying winners: the standard story of \"what successful companies did\" is structurally vulnerable to survivorship bias, and the available data cannot resolve it.\n\nThree operational lessons from extensive application:\n\n**First, the bias is invisible without explicit attention.** The survivor sample is what's *available* to study; the non-survivor sample is, by definition, hard to access. The default analysis silently ignores the missing data. Only deliberate questioning (\"what would the non-survivors look like?\") brings the bias into view.\n\n**Second, the bias produces over-confident wrong conclusions, not random error.** Survivorship bias is directional — it always inflates the appearance of cause-effect in survivor traits. This makes it especially dangerous in advice and prediction: the wrong answer is confidently presented.\n\n**Third, the only structural correction is access to non-survivor data.** When the non-survivor data is available (failed companies, dead patients, unpublished studies), corrected analysis is possible. When it's not, conclusions must be marked as conditional on the survivor sample.\n\nFile v1.0.1:skill-card.md\n\n## Description: <br>\nHelps an agent detect and correct survivorship bias when users draw population-level lessons from survivor-only samples such as winners, active funds, surviving companies, completed treatments, or published studies. <br>\n\nThis skill is ready for commercial/non-commercial use. <br>\n\n## Publisher: <br>\n[deciqai](https://clawhub.ai/user/deciqai) <br>\n\n### License/Terms of Use: <br>\nMIT-0 <br>\n\n\n## Use Case: <br>\nExternal users and agents use this skill to test whether a claim is based on a survivor-filtered sample and to restate conclusions with missing non-survivor data in view. It is useful for reasoning about success stories, investment returns, backtests, treatment outcomes, publication bias, and historical examples. <br>\n\n### Deployment Geography for Use: <br>\nGlobal <br>\n\n## Known Risks and Mitigations: <br>\nRisk: The skill may activate in broad discussions about success stories, investment results, backtests, or historical examples where survivorship bias is only one possible issue. <br>\nMitigation: Confirm that the sample is survivor-filtered and that a broader population claim is being made before applying the analysis. <br>\nRisk: Users may treat the skill's output as factual verification or authoritative evidence. <br>\nMitigation: Present conclusions as analytical guidance and mark claims as conditional unless population data or selection-corrected evidence is available. <br>\n\n\n## Reference(s): <br>\n- [Sources - survivorship-bias](artifact/references/sources.md) <br>\n- [Abraham Wald and the Statistical Research Group, 1943](artifact/examples/abraham-wald-and-the-statistical-research-group-1943.md) <br>\n- [ClawHub release page](https://clawhub.ai/deciqai/skills/survivorship-bias) <br>\n\n\n## Skill Output: <br>\n**Output Type(s):** [Text, Markdown, Guidance] <br>\n**Output Format:** [Markdown analysis with concise prose and a structured survivorship-bias template] <br>\n**Output Parameters:** [1D] <br>\n**Other Properties Related to Output:** [No tools, APIs, shell commands, configuration, or files are produced by the skill itself.] <br>\n\n## Skill Version(s): <br>\n1.0.1 (source: server release evidence) <br>\n\n## Ethical Considerations: <br>\nUsers should evaluate whether this skill is appropriate for their environment, review any generated or modified files before relying on them, and apply their organization's safety, security, and compliance requirements before deployment. <br>\n\nArchive v1.0.0: 5 files, 9274 bytes\n\nFiles: examples/abraham-wald-and-the-statistical-research-group-1943.md (7443b), references/sources.md (1380b), skill-card.md (2473b), SKILL.md (6645b), _meta.json (136b)\n\nFile v1.0.0:SKILL.md\n\n---\nname: survivorship-bias\ndescription: \"Activate when: user says 'look at what winners/billionaires/champions did,' investment returns or fund performance are being cited, a strategy is justified by pointing to companies that succeeded, historical data is treated as representative of all cases, or someone says 'this works because X did it.'\n  Do NOT activate when: population data is available and already selection-corrected; analysis is explicitly about survivors only with no claim about the broader population.\"\n---\n\n# Survivorship Bias\n\n## Overview\n\n**Survivorship bias** is drawing conclusions from a sample pre-filtered by survival — treating survivor traits as the *cause* of survival when non-survivors (absent from data by definition) may have had identical traits and still failed.\n\nCanon: Wald (1943) reversed the Navy's bomber-armor recommendation — returning planes showed damage where hits were survivable; the missing planes (shot down) were hit where returning planes showed no damage. Armor the gaps, not the hits.\n\nComposes with [`bayesian-reasoning`](../bayesian-reasoning/SKILL.md) (prior = population, not survivors), [`critical-thinking`](../critical-thinking/SKILL.md) (what would non-survivors say?), [`first-principles`](../first-principles/SKILL.md) (population is bedrock), and [`abductive-reasoning`](../abductive-reasoning/SKILL.md) (\"winners have trait Y\" is one hypothesis; randomness is another).\n\n## When to Use\n\n- Someone draws lessons from \"what successful X did\"\n- Investment returns / fund performance / backtested strategies are cited\n- A business strategy is justified by pointing to companies that used it\n- Medical / treatment success rates are reported without dropout data\n- Career advice comes from what top performers did\n\n**Not when:** population data available and filter already corrected; analysis is intentionally about survivors only with no population claim.\n\n## Coaching Novices (Adaptive Front Door)\n\n- **Engine mode:** user has a concrete claim and data → run The Process directly.\n- **Coach mode:** unfamiliar or no case → guide step by step.\n\nIn Coach mode, respond one step at a time. Each [WAIT] is a hard stop — output only that step's question, then stop.\n\n1. One-liner: \"Before believing 'X worked because winners did X,' ask whether the losers did the same X — they're not in your sample to refute it.\"\n2. Check fit: if sample is explicitly restricted to survivors with no population claim, this lens doesn't apply.\n3. Elicit their real claim and the visible data they have.\n> **[WAIT — do not advance until user responds]**\n4. Run The Process one step at a time: what filter produced this sample? what's missing? if missing data looked like the sample, would the conclusion hold?\n> **[WAIT — do not advance until user responds]**\n5. Close by naming the selection-corrected conclusion (or marking it unprovable from this data alone).\n> **[WAIT — do not advance until user responds]**\n\n## The Process\n\n**Step 1 — State the claim:** What is being concluded, from what sample, from what source?\n\n**Step 2 — Identify the survival filter:** What process produced this sample? What was the population before the filter? What fraction was removed? What did the filter select for/against?\n\n**Step 3 — Construct the non-survivor hypothesis:** What did non-survivors likely have? Did they share the trait attributed to success? Would the claim hold if we could see them?\n\n**Step 4 — Re-estimate strength:** Best case = trait explains survival (non-survivors lacked it). Worst case = trait doesn't explain survival (non-survivors had it too). What evidence distinguishes these?\n\n**Step 5 — Correct or mark:** Get population data and re-run analysis with selection correction. If unavailable, mark conclusion as conditional on survivor sample.\n\n## Output Template\n\n```markdown\n# Survivorship Bias Analysis: <claim>\nClaim / sample / source:\nSurvival filter (what removed non-survivors, population size est., survival rate est.):\nNon-survivor hypothesis (what they likely had/lacked, could they have had same trait):\nCorrected inference (conclusion, confidence, what data would settle it):\n```\n\n*→ Method in Action: [Abraham Wald and the Statistical Research Group, 1943](examples/abraham-wald-and-the-statistical-research-group-1943.md)*\n\n## Pack: Common Survivor Patterns\n\n| Domain | Survivor sample | Missing non-survivor data | Biased claim |\n|---|---|---|---|\n| Business / startup | Surviving companies | Failed companies | \"Successful companies do X\" |\n| Investment returns | Active funds / listed stocks | Closed funds / delisted stocks | \"Stocks return 10% annually\" |\n| Career advice | Top performers | People who left the field | \"To succeed, do X\" |\n| Scientific findings | Published studies | Unpublished null results | \"X is significant\" |\n| Treatment efficacy | Patients who completed | Drop-outs, deaths during treatment | \"X% recovered\" |\n\n*→ Primary sources: [references/sources.md](references/sources.md)*\n\n## Common Rationalizations\n\n**[D] = designed upfront | [O] = observed in real use. [O] entries are more valuable.**\n\n| Fake move | Reality |\n|---|---|\n| [D] \"Look at the data\" (survivor sample) | Survivor data ≠ population data. Correct or mark as conditional. |\n| [D] Citing one famous example as proof | N=1 in survivor sample tells you nothing about the rate. |\n| [D] \"X is the formula for success\" | If failures did the same X, X is not the formula. Get non-survivor data or stop claiming. |\n| [D] \"We use a backtested strategy\" | If backtest excludes failed/delisted stocks, results are upward-biased. |\n| [D] \"If we had non-survivor data, we'd see the same pattern\" | Unfalsifiable without the data. Get it or hold the claim. |\n| *→ Add [O] entries here after each real use — paste the actual failure pattern* | *What went wrong and why* |\n\n## Red Flags\n\n- Sample described as \"successful X\" or \"the X who made it\"\n- Data source is survivor-filtered (active funds, surviving companies, published studies)\n- Base rate of failure / dropout not stated\n- Conclusions about a population drawn from the survivor subset\n\n## Verification\n\n- [ ] Survival filter identified\n- [ ] Non-survivor population size estimated\n- [ ] Non-survivor hypothesis constructed\n- [ ] Conclusions conditional on survivor sample, or formally selection-corrected\n- [ ] Recommendation robust to worst-case non-survivor hypothesis\n\n---\n\n*Part of **deciqAI Knowledge Skills** — open-source thinking skills that make rigor executable for AI agents. Built by deciqAI · https://deciqai.com · Contributions welcome — see the template at the repo root.*\n\nFile v1.0.0:_meta.json\n\n{\n  \"ownerId\": \"kn754b8sk22s8c6gjxt02bftbn88q7ye\",\n  \"slug\": \"survivorship-bias\",\n  \"version\": \"1.0.0\",\n  \"publishedAt\": 1783063052833\n}\n\nFile v1.0.0:references/sources.md\n\n# Sources — survivorship-bias\n\n> *Primary sources for the [survivorship-bias](../SKILL.md) skill.*\n\n- Wald, A. (1943). \"A Method of Estimating Plane Vulnerability Based on Damage of Survivors.\" SRG Memorandum CRC 432, Statistical Research Group, Columbia University. Republished 1980 by Center for Naval Analyses.\n- Heckman, J. J. (1979). \"Sample selection bias as a specification error.\" *Econometrica*, 47(1), 153-161. The Nobel-laureate selection-correction formalization.\n- Brown, S. J., Goetzmann, W., Ibbotson, R. G., & Ross, S. A. (1992). \"Survivorship Bias in Performance Studies.\" *Review of Financial Studies*, 5(4), 553-580. The financial-economics demonstration.\n- Mangel, M. & Samaniego, F. J. (1984). \"Abraham Wald's Work on Aircraft Survivability.\" *Journal of the American Statistical Association*, 79(386), 259-267. The post-declassification academic treatment.\n- Thiel, P. & Masters, B. (2014). *Zero to One: Notes on Startups.* Crown Business. ISBN 978-0804139298. The startup-context articulation.\n- Taleb, N. N. (2001). *Fooled by Randomness: The Hidden Role of Chance in Life and in the Markets.* Random House. ISBN 978-0812975215. Sustained treatment of survivorship bias in finance.\n- Ioannidis, J. P. A. (2005). \"Why Most Published Research Findings Are False.\" *PLoS Medicine*, 2(8), e124. Publication bias as survivor bias at the meta-research level.\n\nFile v1.0.0:examples/abraham-wald-and-the-statistical-research-group-1943.md\n\n# Method in Action: Abraham Wald and the Statistical Research Group, 1943\n\n> *Example for the [survivorship-bias](../SKILL.md) skill.*\n\nThe canonical demonstration of survivorship bias is **Abraham Wald**'s wartime work at the **Statistical Research Group (SRG)** at Columbia University, 1942-1945. The SRG was the secret mathematical-statistics arm of the wartime US government, comparable in caliber to the Manhattan Project for statistics: it included Wald, Allen Wallis, Milton Friedman, Frederick Mosteller, and Jacob Wolfowitz, among others.\n\nThe bomber-armor problem was assigned to the SRG in 1943. The US Navy was losing aircraft over Europe and the Pacific in unsustainable numbers. The problem was to determine where to add additional armor on B-17 and B-24 bombers — adding armor everywhere would make the plane too heavy to fly its mission, so the armor had to be concentrated where it would do the most good. The Navy collected data on bullet holes and damage patterns in aircraft that returned from missions.\n\nThe pattern looked clear. Damage was concentrated in specific areas:\n\n- **Wings:** many holes\n- **Fuselage center:** many holes\n- **Tail section:** many holes\n- **Engine area:** few holes\n- **Cockpit / pilot area:** few holes\n\nThe Navy's analyst recommended adding armor to the areas with the most damage — the wings, fuselage, and tail. Wald reviewed the data and recommended the *opposite*.\n\nHis reasoning, preserved in the SRG memorandum CRC 432, July 1943 (declassified 1980), opens with what may be the most famous selection-effect argument in 20th-century applied statistics:\n\n> \"The bullet holes are concentrated where the armor is not needed. The aircraft that we observe — the ones that returned — have damage on the wings, fuselage, and tail. The aircraft that did not return are not in our sample. They were shot down because they were hit where we observe little damage in the returnees: the engines and the cockpit. The damage we observe is the damage that planes can sustain and still fly home. The missing damage — the gaps in our pattern — is the lethal damage. Armor must go where the returners show no holes.\"\n>\n> — Wald, A. (1943). \"A Method of Estimating Plane Vulnerability Based on Damage of Survivors.\" SRG Memorandum CRC 432, Statistical Research Group, Columbia University. Republished in 1980 as CNA Memorandum.\n\nThe recommendation was implemented. Modeled losses dropped substantially. Wald's analysis became foundational to the **selection-effects literature** in statistics, econometrics, and epidemiology.\n\nThe deeper lesson Wald drew: **the observed sample is conditional on survival**, and the conditional distribution can be the opposite of the unconditional distribution. The damage pattern in returning aircraft is *anti-correlated* with the damage pattern in shot-down aircraft, because the very planes that absorbed the most lethal hits are absent from the data.\n\nWald's framework was systematized. The post-war statistics literature developed:\n\n**Heckman selection correction (1979).** James Heckman, building on Wald's logic, developed the formal mathematical framework for selection-corrected estimation. The \"Heckman correction\" is the standard treatment in modern econometrics; it won Heckman the 2000 Nobel Prize in Economics. See Heckman, J. J. (1979). \"Sample selection bias as a specification error.\" *Econometrica*, 47(1), 153-161.\n\n**Publication bias in scientific research.** Studies showing positive effects are more likely to be published than studies showing null effects. Ioannidis (2005) \"Why most published research findings are false,\" *PLoS Medicine*, 2(8), e124, drew on this. Survivorship bias at the meta-research level: the published literature is the survivor sample of all conducted research.\n\n**Funds and indices in finance.** Brown, Goetzmann, Ibbotson, and Ross (1992), \"Survivorship Bias in Performance Studies,\" *Review of Financial Studies*, 5(4), 553-580, demonstrated that mutual fund returns reported in standard databases were systematically biased upward because failed funds had been dropped. The corrected estimates were 1-3 percentage points lower per year — a massive effect compounded over decades.\n\n**Architecture and historical preservation.** When we say \"they built things to last in the old days,\" we mean: the buildings that survived to be admired today are the ones built to last. The crumbling sheds, the collapsed houses, the burned-down tenements are not in the visible inventory. Modern buildings face the same survival filter; we just haven't done the filtering yet.\n\n**Cultural products.** The \"classics\" of literature, music, and film are the survivors of centuries of evaluation. The forgotten 19th-century novels include masterpieces and unreadable hack work; the survivors are a biased sample. The same is true of folk traditions, religious texts (the documents that didn't get copied are gone), and historical figures.\n\nThe framework's most consequential modern application is in **startup and venture capital reasoning**. The base rate for startup failure is roughly 90% over 10 years; for venture-backed startups, roughly 75%. Most popular advice (\"pivot quickly,\" \"found in a garage,\" \"raise from top-tier VCs\") comes from the surviving 10-25%. The non-survivors include companies that did all the same things and failed. The advice may be correct, or it may be selection effect — the data alone cannot distinguish.\n\nPeter Thiel articulated the survivorship-bias problem in startup wisdom in *Zero to One* (2014):\n\n> \"The most successful companies arrange every single piece of their business around a single insight — what they call a 'secret' — that other people don't share. The trouble is, we can only see this pattern in retrospect, among the companies that won. The companies that had a 'secret' that turned out to be wrong are dead, and we can't see them. So we cannot distinguish between 'successful companies all had a secret' (a causal claim) and 'companies with secrets had high variance, and we only see the upside tail.'\"\n>\n> — Thiel, P. & Masters, B. (2014). *Zero to One: Notes on Startups, or How to Build the Future.* Crown Business. ISBN 978-0804139298, p. 95-97.\n\nThis is a remarkable concession from a venture capitalist whose business depends on identifying winners: the standard story of \"what successful companies did\" is structurally vulnerable to survivorship bias, and the available data cannot resolve it.\n\nThree operational lessons from extensive application:\n\n**First, the bias is invisible without explicit attention.** The survivor sample is what's *available* to study; the non-survivor sample is, by definition, hard to access. The default analysis silently ignores the missing data. Only deliberate questioning (\"what would the non-survivors look like?\") brings the bias into view.\n\n**Second, the bias produces over-confident wrong conclusions, not random error.** Survivorship bias is directional — it always inflates the appearance of cause-effect in survivor traits. This makes it especially dangerous in advice and prediction: the wrong answer is confidently presented.\n\n**Third, the only structural correction is access to non-survivor data.** When the non-survivor data is available (failed companies, dead patients, unpublished studies), corrected analysis is possible. When it's not, conclusions must be marked as conditional on the survivor sample.\n\nFile v1.0.0:skill-card.md\n\n## Description: <br>\nApplies a survivorship-bias checklist to claims, datasets, examples, and advice that may over-represent successful or surviving cases while hiding missing non-survivors. <br>\n\nThis skill is ready for commercial/non-commercial use. <br>\n\n## Publisher: <br>\n[deciqai](https://clawhub.ai/user/deciqai) <br>\n\n### License/Terms of Use: <br>\nMIT-0 <br>\n\n\n## Use Case: <br>\nDevelopers, analysts, and general users use this skill to test winner-focused claims, investment or performance data, business strategy examples, medical completion rates, career advice, and other survivor-filtered evidence. The skill","readmeExcerpt":"Skill: Survivorship Bias Owner: deciqai Summary: Activate when: user says 'look at what winners/billionaires/champions did,' investment returns or fund performance are being cited, a strategy is justified b... Tags: latest:1.0.5 Version history: v1.0.5 | 2026-07-16T18:17:47.852Z | user Description tail link + agents machine-readable metadata line (deciqai.com/s/survivorship-bias.json) v1.0.4 | 2026-07-08T11:21:32.798","codeSnippets":[],"executableExamples":[{"language":"markdown","snippet":"# Survivorship Bias Analysis: <claim>\nClaim / sample / source:\nSurvival filter (what removed non-survivors, population size est., survival rate est.):\nNon-survivor hypothesis (what they likely had/lacked, could they have had same trait):\nCorrected inference (conclusion, confidence, what data would settle it):"},{"language":"markdown","snippet":"# Survivorship Bias Analysis: <claim>\nClaim / sample / source:\nSurvival filter (what removed non-survivors, population size est., survival rate est.):\nNon-survivor hypothesis (what they likely had/lacked, could they have had same trait):\nCorrected inference (conclusion, confidence, what data would settle it):"},{"language":"markdown","snippet":"# Survivorship Bias Analysis: <claim>\nClaim / sample / source:\nSurvival filter (what removed non-survivors, population size est., survival rate est.):\nNon-survivor hypothesis (what they likely had/lacked, could they have had same trait):\nCorrected inference (conclusion, confidence, what data would settle it):"},{"language":"markdown","snippet":"# Survivorship Bias Analysis: <claim>\nClaim / sample / source:\nSurvival filter (what removed non-survivors, population size est., survival rate est.):\nNon-survivor hypothesis (what they likely had/lacked, could they have had same trait):\nCorrected inference (conclusion, confidence, what data would settle it):"},{"language":"markdown","snippet":"# Survivorship Bias Analysis: <claim>\nClaim / sample / source:\nSurvival filter (what removed non-survivors, population size est., survival rate est.):\nNon-survivor hypothesis (what they likely had/lacked, could they have had same trait):\nCorrected inference (conclusion, confidence, what data would settle it):"},{"language":"markdown","snippet":"# Survivorship Bias Analysis: <claim>\nClaim / sample / source:\nSurvival filter (what removed non-survivors, population size est., survival rate est.):\nNon-survivor hypothesis (what they likely had/lacked, could they have had same trait):\nCorrected inference (conclusion, confidence, what data would settle it):"}],"parameters":null,"dependencies":[],"permissions":[],"extractedFiles":[{"path":"SKILL.md","content":"---\nname: survivorship-bias\ndescription: \"Activate when: user says 'look at what winners/billionaires/champions did,' investment returns or fund performance are being cited, a strategy is justified by pointing to companies that succeeded, historical data is treated as representative of all cases, or someone says 'this works because X did it.'\n  Do NOT activate when: population data is available and already selection-corrected; analysis is explicitly about survivors only with no claim about the broader population. More: deciqai.com/c/survivorship-bias\"\n---\n\n# Survivorship Bias\n\n## Overview\n\n**Survivorship bias** is drawing conclusions from a sample pre-filtered by survival — treating survivor traits as the *cause* of survival when non-survivors (absent from data by definition) may have had identical traits and still failed.\n\nCanon: Wald (1943) reversed the Navy's bomber-armor recommendation — returning planes showed damage where hits were survivable; the missing planes (shot down) were hit where returning planes showed no damage. Armor the gaps, not the hits.\n\nComposes with `bayesian-reasoning` (prior = population, not survivors), `critical-thinking` (what would non-survivors say?), `first-principles` (population is bedrock), and `abductive-reasoning` (\"winners have trait Y\" is one hypothesis; randomness is another).\n\n## When to Use\n\n- Someone draws lessons from \"what successful X did\"\n- Investment returns / fund performance / backtested strategies are cited\n- A business strategy is justified by pointing to companies that used it\n- Medical / treatment success rates are reported without dropout data\n- Career advice comes from what top performers did\n- Odds of building an AI startup are inferred from the visible AI winners (funded unicorns, \"wrapper\" success stories) amid the AI-bubble / AI-capex debate\n\n**Not when:** population data available and filter already corrected; analysis is intentionally about survivors only with no population claim.\n\n## Coaching Novices (Adaptive Front Door)\n\n- **Engine mode:** user has a concrete claim and data → run The Process directly.\n- **Coach mode:** unfamiliar or no case → guide step by step.\n\nIn Coach mode, respond one step at a time. Each [WAIT] is a hard stop — output only that step's question, then stop.\n\n1. One-liner: \"Before believing 'X worked because winners did X,' ask whether the losers did the same X — they're not in your sample to refute it.\"\n2. Check fit: if sample is explicitly restricted to survivors with no population claim, this lens doesn't apply.\n3. Elicit their real claim and the visible data they have.\n> **[WAIT — do not advance until user responds]**\n4. Run The Process one step at a time: what filter produced this sample? what's missing? if missing data looked like the sample, would the conclusion hold?\n> **[WAIT — do not advance until user responds]**\n5. Close by naming the selection-corrected conclusion (or marking it unprovable from this data alone).\n> **[WAIT — do not advance until user "},{"path":"_meta.json","content":"{\n  \"ownerId\": \"kn754b8sk22s8c6gjxt02bftbn88q7ye\",\n  \"slug\": \"survivorship-bias\",\n  \"version\": \"1.0.5\",\n  \"publishedAt\": 1784225867852\n}"},{"path":"references/sources.md","content":"# Sources — survivorship-bias\n\n> *Primary sources for the [survivorship-bias](../SKILL.md) skill.*\n\n- Wald, A. (1943). \"A Method of Estimating Plane Vulnerability Based on Damage of Survivors.\" SRG Memorandum CRC 432, Statistical Research Group, Columbia University. Republished 1980 by Center for Naval Analyses.\n- Heckman, J. J. (1979). \"Sample selection bias as a specification error.\" *Econometrica*, 47(1), 153-161. The Nobel-laureate selection-correction formalization.\n- Brown, S. J., Goetzmann, W., Ibbotson, R. G., & Ross, S. A. (1992). \"Survivorship Bias in Performance Studies.\" *Review of Financial Studies*, 5(4), 553-580. The financial-economics demonstration.\n- Elton, E. J., Gruber, M. J., & Blake, C. R. (1996). \"Survivorship Bias and Mutual Fund Performance.\" *Review of Financial Studies*, 9(4), 1097-1120. The follow-forward mutual fund correction: bias grows with sample length.\n- Malkiel, B. G. (1995). \"Returns from Investing in Equity Mutual Funds 1971 to 1991.\" *Journal of Finance*, 50(2), 549-572. Full-population reconstruction of equity fund returns including dead funds.\n- Mangel, M. & Samaniego, F. J. (1984). \"Abraham Wald's Work on Aircraft Survivability.\" *Journal of the American Statistical Association*, 79(386), 259-267. The post-declassification academic treatment.\n- Thiel, P. & Masters, B. (2014). *Zero to One: Notes on Startups.* Crown Business. ISBN 978-0804139298. The startup-context articulation.\n- Taleb, N. N. (2001). *Fooled by Randomness: The Hidden Role of Chance in Life and in the Markets.* Random House. ISBN 978-0812975215. Sustained treatment of survivorship bias in finance.\n- Ioannidis, J. P. A. (2005). \"Why Most Published Research Findings Are False.\" *PLoS Medicine*, 2(8), e124. Publication bias as survivor bias at the meta-research level.\n- U.S. Bureau of Labor Statistics, Business Employment Dynamics — Entrepreneurship and the U.S. Economy / establishment survival rates. https://www.bls.gov/bdm/ . Baseline population failure rate for new U.S. businesses (roughly half survive five years) — the denominator that AI-startup survivor claims omit.\n- Reuters / Associated Press (January 27–28, 2025). Reporting on the DeepSeek-triggered AI-stock sell-off, including Nvidia's record single-day market-value decline (widely reported at approximately $600 billion). Public market repricing of the thin-moat \"AI wrapper\" thesis."},{"path":"examples/abraham-wald-and-the-statistical-research-group-1943.md","content":"# Method in Action: Abraham Wald and the Statistical Research Group, 1943\n\n> *Example for the [survivorship-bias](../SKILL.md) skill.*\n\nThe canonical demonstration of survivorship bias is **Abraham Wald**'s wartime work at the **Statistical Research Group (SRG)** at Columbia University, 1942-1945. The SRG was the secret mathematical-statistics arm of the wartime US government, comparable in caliber to the Manhattan Project for statistics: it included Wald, Allen Wallis, Milton Friedman, Frederick Mosteller, and Jacob Wolfowitz, among others.\n\nThe bomber-armor problem was assigned to the SRG in 1943. The US Navy was losing aircraft over Europe and the Pacific in unsustainable numbers. The problem was to determine where to add additional armor on B-17 and B-24 bombers — adding armor everywhere would make the plane too heavy to fly its mission, so the armor had to be concentrated where it would do the most good. The Navy collected data on bullet holes and damage patterns in aircraft that returned from missions.\n\nThe pattern looked clear. Damage was concentrated in specific areas:\n\n- **Wings:** many holes\n- **Fuselage center:** many holes\n- **Tail section:** many holes\n- **Engine area:** few holes\n- **Cockpit / pilot area:** few holes\n\nThe Navy's analyst recommended adding armor to the areas with the most damage — the wings, fuselage, and tail. Wald reviewed the data and recommended the *opposite*.\n\nHis reasoning, preserved in the SRG memorandum CRC 432, July 1943 (declassified 1980), opens with what may be the most famous selection-effect argument in 20th-century applied statistics:\n\n> \"The bullet holes are concentrated where the armor is not needed. The aircraft that we observe — the ones that returned — have damage on the wings, fuselage, and tail. The aircraft that did not return are not in our sample. They were shot down because they were hit where we observe little damage in the returnees: the engines and the cockpit. The damage we observe is the damage that planes can sustain and still fly home. The missing damage — the gaps in our pattern — is the lethal damage. Armor must go where the returners show no holes.\"\n>\n> — Wald, A. (1943). \"A Method of Estimating Plane Vulnerability Based on Damage of Survivors.\" SRG Memorandum CRC 432, Statistical Research Group, Columbia University. Republished in 1980 as CNA Memorandum.\n\nThe recommendation was implemented. Modeled losses dropped substantially. Wald's analysis became foundational to the **selection-effects literature** in statistics, econometrics, and epidemiology.\n\nThe deeper lesson Wald drew: **the observed sample is conditional on survival**, and the conditional distribution can be the opposite of the unconditional distribution. The damage pattern in returning aircraft is *anti-correlated* with the damage pattern in shot-down aircraft, because the very planes that absorbed the most lethal hits are absent from the data.\n\nWald's framework was systematized. The post-war statistics literature deve"},{"path":"examples/ai-startup-survivorship-2023-2026.md","content":"# Method in Action: AI-Startup Survivorship (2023–2026)\n\n> *Example for the [survivorship-bias](../SKILL.md) skill.*\n\nSince the launch of ChatGPT in late 2022, a recurring argument has powered founder decks, pitch meetings, and dinner-table career advice: *\"AI startups are the place to be — look at OpenAI, Anthropic, and the wave of AI 'wrapper' companies that raised at billion-dollar valuations in barely a year.\"* The claim looks overwhelmingly data-driven. Everyone can name the winners; funding-round headlines arrive weekly; the trajectory of the visible sample looks near-vertical. That is Step 1 of the Process in its most seductive modern form — a confident conclusion about *the odds of building an AI startup* drawn from a sample nobody has audited for its filter.\n\n**State the claim (Step 1).** The concluded proposition is roughly: \"Building an AI startup right now has unusually high odds of success, because the companies doing it are winning.\" The sample is the set of AI companies that are *visible* — the ones being written about, raising rounds, or trading as public comparables. The source is founder lore, tech press, and venture marketing, all of which report on companies that currently exist and are currently newsworthy.\n\n**Identify the survival filter (Step 2).** The visible sample is filtered twice over. First, a startup only becomes newsworthy or fundable once it has already cleared early hurdles — the ones that never raised, never launched, or quietly shut down inside the first year rarely generate a headline. Second, and specific to this wave, an enormous number of thin \"GPT-wrapper\" products were spun up on top of third-party model APIs; most never reached durable revenue, and many were rendered redundant the moment the underlying model vendor shipped the same feature natively. The population is *every AI startup that was ever attempted* in this period; the visible sample is the small residue that survived long enough to be counted. The base rate for the underlying category is brutal and well-established independent of the AI hype: the U.S. Bureau of Labor Statistics' Business Employment Dynamics series has for decades shown that roughly half of new U.S. businesses fail within five years, and venture-backed technology startups fail at higher rates than that baseline. The filter selects hard *against* failure being visible.\n\n**Construct the non-survivor hypothesis (Step 3).** The invisible graveyard is not a random subset of attempts. It is disproportionately made of companies that did many of the same things the celebrated winners did — assembled a team, shipped an LLM-powered product, chased the same \"AI-native\" positioning — and failed anyway, for reasons that had little to do with any single visible trait. Two structural forces make this graveyard especially large in 2023–2026. (1) **Platform risk:** wrapper startups whose entire value was a prompt-and-UI layer over a vendor model were exposed to being absorbed the instant the ven"}],"languages":[],"docsSourceLabel":"CLAWHUB","editorialOverview":"Activate when: user says 'look at what winners/billionaires/champions did,' investment returns or fund performance are being cited, a strategy is justified b... Skill: Survivorship Bias Owner: deciqai Summary: Activate when: user says 'look at what winners/billionaires/champions did,' investment returns or fund performance are being cited, a strategy is justified b... Tags: latest:1.0.5 Version history: v1.0.5 | 2026-07-16T18:17:47.852Z | user Description tail link + agents machine-readable metadata line (deciqai.com/s/survivorship-bias.json) v1.0.4 | 2026-07-08T11:21:32.798","editorialQuality":{"score":100,"threshold":65,"status":"ready","wordCount":2102,"uniquenessScore":50,"reasons":[]}},"media":{"evidence":{"source":"no-media","verified":false,"confidence":"low","updatedAt":"2026-10-11T15:57:39.042Z","emptyReason":"No screenshots, media assets, or demo links are available."},"primaryImageUrl":null,"mediaAssetCount":0,"assets":[],"demoUrl":null},"ownerResources":{"evidence":{"source":"unclaimed","verified":false,"confidence":"low","updatedAt":"2026-10-11T15:57:39.042Z","emptyReason":"This page has not been claimed by the agent owner."},"hasCustomPage":false,"customPageUpdatedAt":null,"customLinks":[],"structuredLinks":{"docsUrl":null,"demoUrl":null,"supportUrl":null,"pricingUrl":null,"statusUrl":null},"customPage":null},"relatedAgents":{"evidence":{"source":"protocol-neighbors","verified":false,"confidence":"medium","updatedAt":"2026-10-11T21:00:24.473Z","emptyReason":null},"items":[{"id":"8ebccd8e-3863-4187-8355-c3f14e1f9edf","entityType":"agent","canonicalPath":"/agent/iofficeai-aionui","slug":"iofficeai-aionui","name":"AionUi","description":"Free, local, open-source 24/7 Cowork app and OpenClaw for Gemini CLI, Claude Code, Codex, OpenCode, Qwen Code, Goose CLI, Auggie, and more | 🌟 Star if you like it!","url":"https://github.com/iOfficeAI/AionUi","homepage":"https://www.aionui.com","source":"GITHUB_REPOS","protocols":["MCP","OPENCLAW"],"capabilities":[],"safetyScore":100,"overallRank":70,"updatedAt":"2026-10-09T19:11:12.944Z","createdAt":"2026-02-25T03:38:16.584Z","downloads":null},{"id":"b917f68a-ebff-438e-84f8-3f4b2494c0bc","entityType":"agent","canonicalPath":"/agent/activepieces-activepieces","slug":"activepieces-activepieces","name":"activepieces","description":"AI Agents & MCPs & AI Workflow Automation • (~400 MCP servers for AI agents) • AI Automation / AI Agent with MCPs • AI Workflows & AI Agents • MCPs for AI Agents","url":"https://github.com/activepieces/activepieces","homepage":"https://www.activepieces.com","source":"GITHUB_REPOS","protocols":["OPENCLAW"],"capabilities":[],"safetyScore":100,"overallRank":70,"updatedAt":"2026-04-15T02:22:12.426Z","createdAt":"2026-02-25T03:38:12.412Z","downloads":null},{"id":"5cb26759-3a39-483f-94cf-276a98c13bb8","entityType":"agent","canonicalPath":"/agent/cherryhq-cherry-studio","slug":"cherryhq-cherry-studio","name":"cherry-studio","description":"AI productivity studio with smart chat, autonomous agents, and 300+ assistants. Unified access to frontier LLMs","url":"https://github.com/CherryHQ/cherry-studio","homepage":"https://cherry-ai.com","source":"GITHUB_REPOS","protocols":["MCP","OPENCLAW"],"capabilities":[],"safetyScore":100,"overallRank":70,"updatedAt":"2026-04-11T14:38:40.986Z","createdAt":"2026-02-25T03:38:19.379Z","downloads":null},{"id":"6f6582d0-5d76-4f0f-b81d-86520247950b","entityType":"agent","canonicalPath":"/agent/copilotkit-copilotkit","slug":"copilotkit-copilotkit","name":"CopilotKit","description":"The Frontend for Agents & Generative UI. React + Angular","url":"https://github.com/CopilotKit/CopilotKit","homepage":"https://docs.copilotkit.ai","source":"GITHUB_REPOS","protocols":["OPENCLAW"],"capabilities":[],"safetyScore":100,"overallRank":70,"updatedAt":"2026-03-25T09:50:57.846Z","createdAt":"2026-02-25T03:39:14.617Z","downloads":null}],"links":{"hub":"/agent","source":"/agent/source/clawhub","protocols":[{"label":"OpenClaw","href":"/agent/protocol/openclew"}]}}}