Method

How the score is worked out

Most tools like this hand you a number and no way to check it. Here is exactly where ours comes from, and what it can't tell you.

The model doesn't choose the number

Asking a language model for "a score out of 10" gets you a different answer every time you ask. So it isn't asked. It scores six named dimensions separately, each with a written justification, and the final number is arithmetic done in our code.

Demand30%
Is the problem real, urgent and already costing someone money?
Unit economics22%
Does the margin on one sale survive the cost of winning that customer?
Competition18%
How crowded, and how beatable are the people already there?
Defensibility16%
Is there anything here that is still hard to copy in three years?
Execution risk14%
Regulatory, logistical and technical walls, scored inverted.
Founder–market fitadjustment
Only applied when you have told us about the team, so nobody is punished for leaving it out.

score = Σ (weight × dimension) , then ±1 point at most for founder–market fit when it is known. Rounded to one decimal, clamped to 0–10.

The weights are a judgement call, not a law of nature. They are written here so you can disagree with them out loud rather than wonder what the black box was thinking.

The confidence band is not decoration

A score of 7.0 built from a two-line description is not the same claim as a 7.0 built from a description with a price, a named customer and existing traction. So each dimension is tagged as stated (you told us) or inferred (the model filled it in), and the more inferring it had to do, the wider the band gets.

If you see ±2.0, the honest reading is "somewhere between 5 and 9, and we are mostly guessing". Narrow it by writing more.

Unsupported figures are labelled

Market sizes, growth rates and cost estimates are the easiest thing in the world for a language model to invent, because an invented number reads exactly like a real one. Where a figure can be tied to a source, the source is linked. Where it can't, it is marked as an estimate in the report rather than quietly presented as a fact.

It will refuse to score a non-idea

If your description is too thin to evaluate — no customer, no problem, no way it makes money — you get questions instead of a report. A confident-looking analysis of four words is worse than no analysis, because you might believe it.

What this score is not

  • It is not a prediction. Nothing here has been tested against whether high-scoring ideas actually succeed. Until that study exists, treat the number as a structured opinion, not a forecast.
  • It is not advice. It is a language model's reading of what you wrote. Check anything you plan to spend money on.
  • It has not met you. Founder–market fit and execution ability decide most outcomes, and a text box cannot see either.
  • It rewards clear writing. A well-described mediocre idea will out-score a badly described good one. That is a real limitation, not a feature.

Same idea, same score

The model runs at temperature zero against a fixed schema, so repeating an identical description returns the same score and the same units. Earlier versions of this tool did not do that, and the market size could move by half its own value between two runs thirty seconds apart. If you ever see that happen again, it is a bug — please tell us.