How the failure corpus is built
241 healthcare startups in this corpus were verified dead using a 5-signal composite: domain status, LinkedIn employee dispersal, Crunchbase closed flag, press coverage, and funding cessation. Each failure is then Gioia-coded with a 3-level causal chain — first-order observations, second-order themes, and a structural failure node. Minimum 3 signals required for inclusion.
Gioia coding methodology
Each corpus entry is coded using the Gioia three-level qualitative methodology, adapted from Gioia, Corley & Hamilton (2013) for structural failure analysis. The three levels:
Level 1 — First-order codes
Raw observations from primary sources: founder post-mortems, investor write-ups, press coverage, SEC filings, SBIR award databases. Direct claims and factual statements only.
Level 2 — Second-order themes
First-order codes grouped into structural patterns through constant comparison. A pattern of "buyers demanded proof we couldn't produce" codes to clinical validation gap. A pattern of "sales cycles exceeded runway" codes to CAC payback crisis.
Level 3 — Aggregate dimensions (13 nodes)
Themes mapped to one of 13 structural failure nodes. Each entry records the full Gioia chain for auditability. The taxonomy is append-only — new nodes require documented anchor entries and dual validation.
Coding sessions are date-stamped and linked to primary source URLs. Any entry whose primary source becomes unavailable is flagged for re-verification at the next audit cycle. The audit log is maintained in version control alongside the corpus.
Entries are tagged with provenance: human_validated, auto_coded_verified, or auto_coded_unverified. Bayesian weights: 1.0× / 1.0× / 0.7× respectively.
Coding Consistency
Classification consistency is assessed using a synthetic inter-rater reliability protocol: the same 30 thesis cases are run through the structural classification pipeline twice and the primary node assignments are compared using Cohen's κ.
Method
Synthetic IRR — pipeline-vs-pipeline comparison (two independent classification runs on the same input). The structural matcher and pre-classifier contain no LLM calls and are deterministically reproducible at temperature 0. This is not human-vs-human coding agreement.
Metric
Cohen's κ on primary node agreement across 30 held-out thesis cases spanning ≥8 sub-verticals.
Target
κ ≥ 0.75 (substantial agreement). Marginal: 0.60–0.74. Fail: <0.60.
Framework alignment
Methodology adapted from Cantamessa et al. (2018) startup failure taxonomy approach, which applies structured qualitative coding with inter-rater verification to failure mode classification.
IRR results are logged to data/irr-log.json and re-run after every corpus expansion of ≥20 cases.
Coverage Distribution
Shannon entropy measures how broadly failure cases are distributed across the 13 taxonomy nodes. It is a measure of coverage breadth — not coding accuracy, analytical validity, or predictive power.
A corpus with high entropy covers the failure landscape broadly (no single pattern dominates). A corpus with low entropy has cases concentrated in a few nodes — useful for that sub-domain, but not representative of the full failure space. Current Shannon entropy ratio: 0.87 (87% of theoretical maximum for 13 nodes). Directionally consistent with CB Insights 2024 failure category coverage.
Per-node case count — all 13 failure nodes
Evidence tiers: Robust ≥25 cases, Developing 10–24, Thin <10. The engine notes thin-node evidence on results pages. Thin nodes do not prevent analysis — they widen the uncertainty band.
Death verification — 5-signal composite
A company is not included as a confirmed failure unless at least 3 of 5 signals are present:
- Domain expired or redirects to a parking page
- LinkedIn headcount dropped to zero or shows dispersal to other employers
- Crunchbase or PitchBook status marked as "closed" or "defunct"
- Press coverage of shutdown, acquisition failure, or pivot away from original mission
- No funding activity for ≥24 months with no product activity
No living companies appear in the failure corpus. Active companies, struggling companies, and acqui-hires with positive outcomes are excluded from all failure calculations.
Quality controls
Node concentration ceiling (25%)
No single failure node may exceed 25% of all confirmed failures. This prevents the corpus from over-representing any one structural pattern. A pre-commit hook blocks corpus changes that would violate this ceiling.
Template placeholder guard
Pre-commit hooks block any corpus entry containing unresolved template strings ({{ }}). Every field must be a real value before an entry can be committed.
Discovery agent verification
Automated discovery agents (SBIR Phase I, EDGAR, press archives) feed a staging pipeline. Every staged entry is reviewed before merging into the live corpus. The pipeline never auto-merges.
What is NOT in the corpus
Scope limitations are part of the methodology. The following are explicitly excluded:
- —
UK/NHS dynamics: The corpus is US-focused. UK NHS procurement structures, commissioning pathways, and NICE approval processes differ structurally from US payer dynamics.
- —
Active companies: No company currently operating appears in the failure corpus, regardless of financial health.
- —
Pivoted companies: Companies that pivoted to an entirely different market are excluded. The original thesis is coded, not the pivot.
- —
Non-structural failures: Fraud, deliberate misrepresentation, and founder misconduct are excluded. These are non-structural causes with no generalizable pattern.
- —
Thin sub-verticals: Sub-verticals with fewer than 5 confirmed failures produce THIN evidence depth classifications. Healthcare B2B (Other) is the most saturated sub-vertical with 33 entries.
Corpus summary: 241 confirmed failures · 15 sub-verticals · 13 structural failure nodes · 5-signal death verification · Gioia three-level coding · 25% node concentration ceiling enforced at commit time. Data sources: SBIR Phase I awards, Crunchbase, PitchBook, SEC EDGAR, Rock Health, CB Insights, founder interviews, press archives.
Related methodology