Method: how the ripeness rule was tested

The rule: two numbers, multiplied

1. Capital burned proving demand. Add up what funded companies that tried the category and died had raised or lost. Investors do not put hundreds of millions into something nobody wants, so a deep graveyard is evidence of demand.
2. How far the killing constraint has moved. Name the specific thing that stopped them (a cost, a law, a capability, an install base) and compare its value now with its value when they died.

A category scores high only when both are large. Multiplying, not adding, is deliberate: location check-in apps shared ride-hailing's enabler (cheap GPS in phones) and produced nothing durable, because nobody had burned money proving people wanted it. Graveyard depth was four for GPS-dispatched delivery and zero for check-in.

What the backtests found

These results come from the project's research runs of 31 July to 6 August 2026. The underlying scripts were not published with them, so treat the numbers as reported, not reproducible here.

• 39 hand-verified cases. The company that won a category was founded a median of 0.0 years after its binding constraint crossed, 80% within -2 to +3 years. The well-funded attempts that died were a median of 15 years early. Only about half of famous "too early" stories survived scrutiny; the rest had a competitor succeeding at the same time.
• Frozen-date backtest on 110 categories (enumerated by constraint, 13 freeze years from 1996 to 2020, 192 observations, constraint values hard-cut at each freeze year): capital × lift reached an AUC of 0.719 (cluster-bootstrap 95% interval 0.60-0.82). It was 0.773 on 23 categories; the drop is what more data does to a small-sample result.
• Label permutation: none of 2,000 outcome shuffles reached the real AUC (p < 0.0005).
• Temporal holdout: a threshold fitted only on 1996-2010 selected 2012-2020 observations at 77% precision (Wilson 95% interval 50-92%), 2.0× the test base rate.
• Abstaining: 49 observations where the named constraint had not measurably improved formed at 68%, above the base rate. The wrong constraint had been named. Scoring them as zero dragged the AUC to 0.583, so the rule now abstains on them. On this board that is the verdict "I don't know".
• Unresolved: whether the timing half (lift) adds anything over capital alone. The difference in AUC has an interval from about -0.08 to +0.26.

What failed, and is published because it failed

• A linguistic index of founders' "why now" claims scored a perfect AUC of 1.000 on synthetic text and exactly 0.500, chance, on 34 real S-1 filings. Dead.
• A pre-registered test of risk-factor language was falsified: 0.556 once the one case the lexicon could have been tuned on was removed.
• A 2,000-candidate scale run produced nothing distinguishable from its null model, because matching companies to constraints by tags carries no causal information.
• A language model forecasting from history scored an AUC of 0.88, but about half of that edge vanished when identifying details were removed. Much of it was memory, not forecasting. Only calls registered before they resolve can show skill, which is why the ledger exists.

Verdicts

Ripe

The constraint that killed the category has crossed, and a named, sourced graveyard proves demand. A cohort-level graveyard ("the 2010s microgrid startups") is not enough.

Trigger pending

A dated public event (a final rule, a pivotal trial, a data release) will decide it. The board watches the source.

Blocked on a check

One named, priced check decides it. Unknown is not the same as false.

No-go

The blocker has not moved, or has moved the wrong way.

I don't know

The named blocker never moved, or it crossed and the category died anyway. Either way the named constraint was not the binding one, and the board abstains.

Formed

History: the constraint crossed and the category formed. These are the backtest, not predictions.

Not a timing failure

A famous "too early" story that fails a test: a competitor succeeded at the same time, the unit economics never worked, fraud, a cost-curve inversion, a coordination failure, or capture by the owner of the missing piece.

Enabler classes: E1 cost curve, E2 bandwidth, E3 device install base, E4 labour structure, E5 legal permission, E6 behavioural norm, E7 algorithmic step, E8 capital availability, E9 protocol invention. E5 crossings are the only kind scheduled in public, which is why the board keeps a regulatory calendar and a slip corpus of how often those dates move.

How calls are made, frozen and resolved

• Every call states a probability, a resolution date, a resolver and a falsifier. Resolvers are automatic where the outcome is in a public API (the Federal Register, FRED, ClinicalTrials.gov) and manual otherwise, with the method written down in advance.
• Calls, verdict changes and resolutions are appended to a SHA-256 hash chain. A registered call cannot be edited; a changed view becomes a new call. The deploy job refuses to publish a ledger that does not extend the one already live, and your browser can re-verify every hash.
• Automatic resolvers run weekly. Manual calls that come due are resolved by a scheduled review that appends the outcome and its evidence.
• The score is the Brier score over resolved calls. Until calls resolve, the board shows no score.

Limits

• Small samples, assembled by one research process. Every interval above is wide.
• Graveyards are only as good as their sources. Capital is shown only where a stable source discloses it, so totals are floors.
• Metric series are proxies for constraints. Live series refresh automatically; hand-collected ones are labelled and refreshed by review.
• The slip corpus is biased toward dates that moved, so its base rates are upper bounds.

Not advice

This is research about categories, not companies. Nothing on this site is investment, legal, tax or medical advice, a rating of any security, or a recommendation to fund, buy or sell anything.