How we count failures, post-mortems and capital loss

Every record in the corpus is classified by entity type (startup, scale-up, public company, bank, crypto protocol, subsidiary…) and event type (what actually happened). Only a subset of event types is treated as a “true failure” for the purpose of survival, lifespan and failure-rate statistics.

A valuation drop on a still-operating public company is not the same event as a Chapter 7 liquidation. We separate them.

We deliberately avoid calling every record a “post-mortem”. Instead we classify each entry into one of four evidence levels, and only the top two levels are counted as post-mortems in our headline numbers.

Headline hierarchy we consider defensible:
1,094 documented failure events · 340 verified case studies · 50 full post-mortems · 0 forensic analyses.

Aggregate capital-loss numbers are the most abused metric in failure databases. Assets, valuations, funding, market cap and customer funds are not the same thing and must never be summed together.

Every record has a single sum-safe field, estimated_net_economic_loss, expressed in USD millions. Only entries where this field is populated and confidence is medium or high contribute to the total.

Loss-method examples

• FTX — customer_shortfall. We count the unrecovered customer deposits (~$8B at time of collapse), not the $32B peak valuation.
• WeWork — funding_sunk. We count the ~$15B of equity written down through Chapter 11, not the peak private valuation of $47B.
• Silicon Valley Bank — equity_writedown. We count the ~$1.9B in equity destroyed at receivership, not the $209B in bank assets (which were largely transferred, not destroyed).
• Rivian / Grab / Didi — no contribution to the total. These are still-operating public companies; market-cap declines are not accounting losses.

The corpus is a sample of publicly documented failures. That sample is not representative of the universe of all companies. We over-represent:

• Mega-failures with mainstream press coverage
• US and English-language markets
• Venture-backed startups with public funding data
• Crypto, fintech, and consumer categories that attract journalism
• Spectacular fraud and collapse events

And we under-represent:

• Bootstrapped micro-businesses
• Silent shutdowns with no press release
• Non-English markets
• Small distressed sales and never-funded ventures

Claims we do not make

• We do not claim “validation reduces failure risk by 7.5×”. That would require a controlled cohort study with treated and control groups. We have not published one.
• We do not claim “predictive signals”. What we surface are recurring patterns observed retrospectively, not signals validated on a holdout set.
• We do not claim to be the “largest post-mortem corpus”. We describe this as one of the largest publicly indexed corpora of documented failures.
• Third-party statistics (e.g. CB Insights’ “42% no market need”) are labelled as external benchmarks and never conflated with distributions computed on the IdeaProof corpus.

Individual case studies attribute their key facts to specific source types. Where a fact carries an interpretation from IdeaProof Research rather than a primary source, it is labelled as such.

Per-field source attribution is being progressively rolled out across the highest-signal cases first (FTX, SVB, WeWork, Theranos, Wirecard, Byju’s and roughly 60 others). Records without per-field attribution fall back to a consolidated source list at the bottom of the case study.

• 2026-07-17 — Honest metrics refactor
  Introduced entity_type, event_type, evidence_level and sum-safe estimated_net_economic_loss. Retired the aggregated “$1T capital destroyed” claim. Retired “7.5× lower risk validated” and “predictive signals” framing. Added this page.