Limitations

1. Survivorship bias — the corpus contains only failures

The corpus documents confirmed dead companies — not near-failures, pivots, or companies that faced identical structural conditions and survived. This means the structural risk score measures pattern similarity to known failures, but cannot account for structurally similar companies that succeeded anyway. A high score reflects pattern proximity to the failure corpus, not a failure rate among all companies that ever exhibited those structural dimensions. Survivor data exists in a separate corpus used for defense evidence, but survivor counts remain small relative to the failure corpus.

2. Funding data from public disclosures only

Funding figures in the corpus come from Crunchbase, PitchBook, SBIR databases, and press disclosures. Many healthcare startups — especially bootstrapped companies and those without institutional venture backing — raise capital through undisclosed rounds, SBIR Phase II grants, strategic investments, or revenue financing that never appears in public databases. The funding figures used in capital-at-risk calculations may significantly understate actual capital destroyed for companies in these categories.

3. Single primary node per entry — reality is multi-causal

The Gioia coding methodology assigns each company to one primary failure node. This constraint produces a clean taxonomy suitable for statistical analysis, but it compresses the genuine multi-causal complexity of real company failures. Most failed companies exhibited signals from two or more failure mechanisms. The primary node captures the dominant structural cause — the mechanism that, if resolved, would most likely have changed the outcome. Secondary causes are noted in the coding chain but do not contribute to corpus match calculations.

4. Narrative depth varies across entries

Some corpus entries are backed by multiple primary sources — detailed founder post-mortems, investor retrospectives, press investigations, and court filings. Others, particularly SBIR-funded early-stage companies that shut down quietly, have thin public records. The extraction completeness score in each entry reflects this variance. Entries with low extraction completeness are weighted at 0.7× in Bayesian calculations when provenance is auto-coded and unverified. This does not eliminate the information gap, but it reduces the influence of thinly-documented entries on prosecution charges.

5. Sub-verticals below 20 entries — pattern distributions may shift

Sub-verticals with fewer than 20 confirmed failures have thin pattern distributions. The top failure node in a sub-vertical with 8 entries may reflect idiosyncratic characteristics of those specific companies rather than a structural pattern in the space. As more entries are added, the dominant failure mode for below-threshold sub-verticals may change substantially. The THIN evidence depth classification makes this limitation explicit whenever a thesis falls into one of these categories.