Method
This is the reasoning behind every El Foso case. What evidence is admitted, how each claimed moat is scored against its proof bar, which tests it has to survive, and why the output is founder questions and kill conditions instead of a score. Read it once and you can audit any case on the site, line by line.
The staged workflow
Seven stages, with a human checkpoint at every hand-off. The evidence ledger is written before any assessment. Nothing reaches the output without a source and an access date.
- Intake. Scope and an arm's-length conflict screen.
- Evidence sweep. The citation ledger, written first.
- Claim mapping. Each claimed network effect against its proof bar.
- Structural diagnostics. Seven tests, forced to record at least one honest weakness.
- AI-erosion test. Each moat component through a decision tree, every leaf with a falsifier.
- Graduation lens. What must be true for Seed to Series A.
- Report. Verdict null, evidence coverage, cost line, three founder questions, kill conditions.
Ratings, not scores
Each claimed network effect gets an ordinal, never a number. A claim with zero proof-bar evidence is capped at the third rating; that cap is the gate.
| Rating | What it requires |
|---|---|
| Proof-bar evidence found and cited; the effect is demonstrated, not asserted. | |
| Some proof-bar items met, others missing; direction supported, magnitude not. | |
| The company claims it; no proof-bar evidence found. The gate cap. | |
| Evidence found that cuts against the claim. |
The one surviving number is evidence coverage (found / missing / contradicted). It measures the diligence, never the company. The cost line measures the leverage. Neither is a verdict.
The AI-erosion test
Each moat component is classified as it meets LLM-era commoditisation. Every call carries the single observation that would flip it.
| Outcome | When |
|---|---|
| A workflow an AI-native team rebuilds cheaply, or a public-web-derivable dataset. | |
| Regulated access, physical supply density, contractual exclusivity, embedded trust. | |
| A real network effect that LLMs make cheaper to multi-tenant around. | |
| Proprietary data that compounds with usage and model quality. |
Limitations
- The instrument is blind to what was never public. It cannot aggregate tacit, local knowledge.
- Seed-stage companies leave a thin public trail; a high "missing" count is often honest, not a defect.
- The network-effects taxonomy is NFX's, paraphrased and credited once in the repo.
- The measurement lens, how effect strength surfaces in acquisition, retention, platform, monetisation and unit-economics KPIs, follows the fund's own published framework; the seven diagnostics map to it, credited once in the repo.
- The erosion tree encodes 2026 model-economics assumptions that will age; the falsifier fields are where that shows first.