The same forecast, delivered two ways.
Change the numbers below and press run. One simulation happens, in your browser, on your machine. Then we show you that single run twice: once as a glass box, with the method, the seed, the uncertainty band and a sealed receipt you can check — and once as a black box, which is the identical run with the information taken out and only a “contact sales” sign left standing.
Glass box
This is what LEXUN returnsEverything needed to disagree with us.
—
Probability the business still has cash at the horizon, as a band
0%50%100%
The shaded region is the answer. The thin tick is the raw point estimate — shown, but never led with.
What moves the answer
Survival by scenario
Sealed receipt — SHA-256 over the canonical bytes of this run
—
—
The receipt hashes everything except the wall-clock timestamp, which necessarily differs between two runs. Paste this run into the reproducer and you will get the same band on your own machine — or verify a sealed record without trusting anything on this page.
Black box
Same run, information removedEverything needed to agree with us, and nothing else.
—
Survival probability
Model: proprietary.
Method: not disclosed.
Uncertainty: not shown.
Seed: not shown.
Inputs: not returned.
Reproducible: no.
Resolution date: none set.
Track record: not published.
Illustrative only — not a live control. A black-box vendor ends the page here; the glass panel beside it already gave you the whole answer.
Read this before you accuse us of a strawman. The number above is not a competitor's output and is not invented. It is the median of our own simulation, the one running in the glass panel, printed to one decimal place with the context stripped out. One run, two deliveries. We could not construct a fairer comparison if we tried, because there is only one set of numbers on this page.
A number you cannot inspect is not a forecast. It is an assertion with a decimal point.
What the black panel took away — and what each thing was worth
Every row below is a real property of the run in the glass panel. The black panel has all of them internally; it simply does not give them to you.
| What you lose | What that costs you |
|---|---|
| The uncertainty band | You cannot tell a 55% you should act on from a 55% that is really "somewhere between 40 and 70". Our band comes from the binomial standard error of the run — — at 5,000 paths — widened when the inputs are mostly assumptions. |
| The seed | Without it nobody, including us, can prove the answer was not cherry-picked from a dozen runs. With it, the same inputs return the same result forever. |
| The method | You cannot check whether the model is appropriate for your situation, or find the assumption you disagree with. Ours is written out on the Accuracy Centre. |
| The inputs and their classification | You cannot see which numbers were facts you supplied and which were assumptions someone chose. That distinction is what our confidence level is computed from — it is structural, not a vibe. |
| The sensitivity ranking | You get a verdict but no lever. The glass panel tells you which single input would move the answer most, so you know where to spend your attention. |
| The resolution criterion and date | Without a frozen criterion and a date, the forecast can never be marked wrong. Unfalsifiable results are comfortable for the vendor and useless for you. |
| The sealed receipt | Nothing stops the number being quietly edited afterwards. A SHA-256 seal over the canonical bytes means any later change is detectable by a stranger. |
| The false precision, in reverse | The black panel adds something too: a decimal place. At 5,000 paths that decimal is noise. Precision beyond the standard error is decoration that reads as authority. |
Why we built this instead of writing an argument
The case for transparency is easy to state and easy to nod along with, which makes it easy to ignore. Watching a single result get worse in front of you is harder to ignore. Nothing was faked to make the point: press run, change the inputs, run it again, and both panels move together, because they are one run.
The honest counter-argument is that a published method is copyable and a hidden one is not. We think that trade is worth making, and we have set out why — including the parts of it that are genuinely weak — on why this is hard to copy and how we compare on the axes that matter.
The biases this engine is built against
Every structural choice in the engine is a countermeasure to a documented failure of human judgement. Named, with the countermeasure and where the evidence comes from:
The planning fallacy. People systematically underestimate time and cost and overestimate success odds, even when they know the base rates (Kahneman & Tversky’s planning-fallacy work; Buehler et al.’s replications). Countermeasure: the outside view is shown beside every UK runway result — the ONS business-demography survival rate, dated and sourced — and a very high band draws a dismissible note naming the fallacy.
Overconfidence. Untrained probability estimates are reliably too narrow. Countermeasure: the engine never returns a point estimate, only a band; input volatility defaults are placeholders that say so; and the calibration test lets anyone measure their own gap between confidence and accuracy in ninety seconds.
Status-quo bias. A randomised study of people genuinely torn over major life changes (Levitt, Review of Economic Studies, 2021) found those who made the change reported higher happiness six months later — population-level evidence that “stay put” is over-chosen. Countermeasure: the career model prices the cost of waiting explicitly, and cites that study as population evidence, never as advice about your case.
Noise. The same person given the same case twice gives different answers (Kahneman, Sibony & Sunstein, Noise, 2021); structured process reduces the scatter. Countermeasure: identical inputs and seed always return the identical band — reproducibility is a property of the engine, checked by the release gate, so the noise a human would add is exactly the part the machine refuses to add.
These notes are educational and population-level. They never change a computed result, and none of them is advice about any individual decision.
Do the same thing with your own decision
This page runs one model on one question. The platform runs the same machinery on business runway, career changes and major purchases, keeps the receipts, and scores the forecasts against what actually happened — including the misses. It is free, it runs in your browser, and your decision data does not leave your device.
Run a simulation Free, in your browser. No account, no card.
See the method and the accuracy record · Reproduce a result yourself · Verify a sealed record