A rigorous account of how MakeOrKillIt turns a paragraph of ambition into a defensible verdict — by refusing to average, scoring in ranges, and arguing against the idea before it lets it pass.
The conventional scorecard rates an idea on several dimensions, averages the result, and treats the number as judgment. It is a comfortable ritual and a dangerous one: a single disqualifying weakness — an idea that is illegal, or has no market, or cannot be built — is quietly offset by strengths elsewhere. The mean survives; the idea should not have.
MakeOrKillIt is built on a different premise. Some flaws are fatal and some are merely weak, and the two must never be summed on the same scale. So the engine separates them: a binary gate first, banded scoring second, and a deterministic rule to combine what remains.
One number per dimension, collapsed into a mean. A nine on market cancels a one on legality. Precision without honesty.
The fatal flaw is caught before scoring begins. Survivors are measured in confidence-weighted ranges and combined by a fixed rule.
Before any score is computed, the idea is put to five must-meet criteria. A clear failure on any one is an instant verdict of kill; the scoring stage is never reached. Where the evidence is genuinely insufficient, the gate does not speculate — it returns the question to you rather than guessing an idea to death.
No dimension is ever a single number. Each is a range whose width is governed by the model's confidence: the less grounded the estimate, the wider the band. When the model would otherwise have to invent a market size or a growth rate, it lowers its confidence instead — and the widening band carries that uncertainty forward into the verdict.
This is the quiet centre of the design. A wide band signals uncertainty and argues for gathering cheap evidence. A band that is narrow but low signals something else entirely: a confidently mediocre idea, to be reshaped or abandoned. The two are different recommendations, and only a range can tell them apart.
Show a range and a confidence, never a false 73 out of 100.
The band is not decoration around the score — the band is the argument. Its position relative to the thresholds, and its width, are what produce the recommendation.
Six weighted dimensions, summing to one. These weights are a starting position, not a fixed truth: the calibration loop revises them as real outcomes accrue.
The scoring model is instructed to be skeptical by default. For each dimension it must produce the single strongest piece of disconfirming evidence; for the idea as a whole it must construct the most persuasive kill case it can, whether or not it believes it. These are surfaced to you directly, under the heading of the strongest arguments against the idea.
It is the oldest defense in the book — Socratic, adversarial — against the most common failure in the room: an author who has already fallen in love with the plan.
The aggregate band, on a scale of one hundred, is read against two lines — a make line at 65 and a kill line at 40. What the band does relative to those lines, and how wide it is, decides the call.
Even the worst realistic case clears the bar. Proceed.
Wide: reduce uncertainty with evidence. Narrow but middling: reshape the idea, or let it go.
Even the best realistic case is too weak to justify. Stop.
Each verdict is stored as a prediction. When the real result becomes known it is recorded and paired back to its forecast. Once enough outcomes have accumulated, the weights and thresholds are re-fit against what actually happened — proposed for human approval, and versioned so any change can be undone. Without this loop the system is merely an articulate opinion. With it, the opinion is answerable to reality — which is the only place its credibility can come from.
Describe the problem, who it is for, and how it earns. The gate, the bands, and the verdict resolve in front of you.
Start the audit