# Coding rules

Written before any company was coded, and amended only by dated entry at the
bottom.

---

## The idea under test

There is a thesis that runs through most practical startup advice: that you can
tell in advance which startups are likely to succeed by looking at **what shape
the product is**. Marketplaces work. Things that let ordinary people earn money
work. Things that make an expensive service cheap work. The thesis is usually
delivered as a list of product categories, and the list is arrived at by looking
at companies that reached a billion dollars and noticing what they have in
common.

This paper does not evaluate any one person's list. It evaluates **the method
that produces such lists**, because the defect, if there is one, is in the
method and would survive any change of list.

---

## The scheme, and where it came from

Seven product shapes. They were not taken from anyone's published taxonomy. They
were written by reading the product descriptions in the failure corpus until the
recurring shapes stopped changing, which happened at seven, and they are the
same shapes that any reader of this literature will recognise, which is the
point rather than a coincidence. If the argument here is right, a different set
of seven would behave the same way.

| code | shape | the test applied |
|---|---|---|
| A | two-sided marketplace | matches independent buyers and sellers in a market where the matching was manual |
| B | earning platform | lets ordinary people earn money from their time or an asset, with little upfront work or qualification |
| C | labour substitution | replaces skilled work the customer would otherwise do or pay for, at a small fraction of the cost |
| D | cheaper or easier substitute | the same thing people already buy, materially cheaper or more convenient, at no loss of quality |
| E | localised replication | an established model rebuilt for a region that did not have it |
| F | invitation software | growth comes from users bringing in other users, to be paid, to sign, or to join a conversation |
| G | access to a gated asset | opens an asset class or service previously closed by wealth, licence or geography |

Two further codes, both real and both used:

- **`none`**: the company is outside all seven. Used, not avoided.
- **`insufficient`**: the corpus entry quotes the founder's goodbye letter and
  never says what the product did. These cannot be coded either way. They are
  counted, published, and excluded from the denominator rather than guessed at.

---

## The rule

**One coder, coding from the description alone.** Each company is coded from the
text of its own entry and nothing else: no searching for what the company later
became, no inference from the company name, no use of anything the coder happens
to know about it. This is deliberate. It holds the evidence for each company to
roughly what was knowable about it as a young company, and it stops an outcome
the coder already knows leaking into a judgement about product shape.

**A shape is assigned when the description states the mechanism.** Not when the
company could be argued into it. A drug developer is not code C because
medicine saves labour; C is about replacing work the customer would otherwise
buy.

**Multiple shapes may be assigned.** A ride-hailing company is genuinely A, B
and D at once, and forcing a single code would hide the thing the paper is
measuring, which is how much of the space these shapes cover.

---

## The three tests, fixed before the data was touched

**Test 1, coverage.** Population: the complete unicorn list. Seven shapes that
all describe consumer or business software cannot describe a company that makes
a rocket, a battery or a drug. The share of the $1B+ population in industries
where a software product is the exception is a ceiling on how much of that
population any such scheme can reach. Computed from the list's own industry
labels; no coding judgement enters it.

**Test 2, discrimination.** Population: the startup failure post-mortem
compilation. Frame: every entry carrying a description of what the product did.
Sample: sixty companies drawn with `random.Random(20260901)`, so the same sixty
come out of every run and the sample cannot be tuned after the fact. The
question is what share of companies that went to zero also satisfy the shapes.

**The decision rule, written before the count.** For product shape to carry
information about outcome, its rate among the dead has to be materially lower
than its rate among the living. Test 1 supplies the ceiling on the living. If
the rate among the dead comes within ten points of that ceiling, the scheme is
describing software startups rather than successful ones, and the thesis fails
on its own terms.

**Test 3, stability.** Population: the unicorn list, split by the year each
company entered it. Any rule fitted to winners is fitted to the winners of some
period. If the composition of that population turns over between cohorts, a rule
fitted to one is describing a moment rather than a mechanism, and its shelf life
is the length of the cohort. Computed; no coding judgement.

---

## What this design cannot do

The winner side is **not** coded for product shape. There are no product
descriptions on the unicorn list, only names, industries, valuations, countries
and entry dates, and coding fourteen hundred companies from names would import
exactly the outcome knowledge that the rule above exists to keep out. So the
paper does not report a winner-side rate. It reports the ceiling from test 1 and
compares the dead against that, which is a weaker comparison than a matched pair
of samples and is stated as such in the limitations.

---

## Amendments

*None. If the coding is revised, the revision is recorded here with a date and a
reason, and the original figure stays in the paper beside the new one.*