# CLAUDE.md — Thematic supply-chain research framework
# BEP Research · Ben Pouladian · bepresearch.substack.com
# Latest version: tools.bepresearch.com/framework

Drop this file in the root of a project folder and Claude Code (or any agent that reads
`CLAUDE.md`) will pick it up automatically. Point it at filings, transcripts, decks, or your
own notes and it will analyse them through this lens.

Shared free by BEP Research, where this method gets applied to a live buildout every week.
It encodes a **way of looking at a market**, not a portfolio. There are no positions, price
targets, or recommendations in this file, and you should not treat its output as advice.

---

## What this framework is for

Finding the durable layer in a technology buildout — the part every buyer has to pass
through regardless of which product, model, or company ends up winning. It was built on the
AI datacenter buildout (memory, optics, packaging, power) and then carried into AI-assisted
drug discovery. It should port to any capital cycle with a physical supply chain underneath
a hyped layer.

## Core principles

**1. The bottleneck always migrates.** Markets fixate on the glamorous layer and misprice the
constraint. In datacenters the constraint was never really the GPU; it moved to memory, to
advanced packaging, to power. Ask where the constraint is *now*, and where it goes next when
this one is relieved. The answer is rarely where attention is.

**2. Own the toll, not the coin flip.** Prefer businesses paid on the *attempt* over
businesses paid on the *outcome*. A vendor selling instruments and consumables gets paid
whether or not the customer's product succeeds. Volume of attempts tends to rise through a
cycle even when the success rate does not.

**3. Separate capital expenditure from consumables.** This is the distinction most analysis
collapses, and it is expensive. Capex is approved once by a committee on a budget cycle and
is deferrable in a lean year, so it is genuinely cyclical. Consumables are spent per unit of
activity and reordered continuously. Two businesses can sit in the same "tools" bucket and
run on completely different clocks. Check which one you actually own.

**4. What diffuses cannot be a moat.** Software, models, and published methods spread. If a
capability can be replicated by reading a paper, renting compute, or hiring three people, it
will not hold pricing power. Durable positions rest on things that take a decade to
replicate: process physics, manufacturing know-how, regulatory acceptance, installed base,
switching cost, and accumulated proprietary data.

**5. Scarcity is the input nobody can synthesize.** In any AI-adjacent stack, find the input
the model cannot generate for itself. Usually it is measured, physical, proprietary data —
and whoever owns the instrument that produces it owns the scarce thing.

## The layer-scoring checklist

Score any layer 1–5 on how hard it is to replace, then ask these in order:

- **Replicability.** Could a well-funded competitor stand this up in three years? If yes, cap
  the score at 2.
- **Recurrence.** Is revenue consumed and reordered, or booked once?
- **Switching cost.** What does the customer lose by leaving — data, validation history,
  regulatory qualification, workflow?
- **Position in the flow.** Does everyone upstream and downstream have to pass through it?
- **Independence from outcomes.** Does it get paid when the customer fails?
- **Concentration honesty.** If several holdings share one macro driver, that is one bet
  wearing several names. Count it as one.

## Falsification

A thesis without a falsification test is a story. For each one, hold three things:

1. The cleanest way the thesis is wrong.
2. The **tripwire**: a specific, observable number that would tell you it is happening.
3. What you would do if it fired.

Rank the risks by how cleanly they break the thesis, not by how easy they are to rebut.

## Two distinctions worth holding

**Structural versus cyclical.** They run on different clocks. A weak quarter does not settle
a decade-long question about who can physically build a thing, and a strong one does not
prove it either. Know which of the two any given data point actually speaks to.

**What the thesis does not require.** Write down the things you are *not* underwriting. It
disciplines the claim and tells you exactly how much has to go right.

## Deliberately not in this file

Positions, weights, price targets, entry levels, or any specific recommendation. This is a
way of thinking, not a book. Build your own conclusions and size your own risk.

---

*BEP Research · bepresearch.substack.com · Research and commentary, not investment advice.*