Guide · Layer 4

Validation

Written for: seed / Series A scientist-founder

Before any sensitivity, confirm the number is even plausible. It’s a quick check, and worth doing first. Come at it from three angles:

• Validate by a different route — for an electrified process, “power cost × ~2” gets you most of the way; rebuilt a genuinely different way, it should agree to within a factor of ~2
• Find comparables — set your number beside the same metric for real, characterized plants. A plausibility test and a read on how you stack up, in one move
• Compare the cost structure — break your cost down on a percentage basis and set it against a real comp or an existing paper. If equipment is normally ~72% of cost in this industry and your model shows 40%, something’s off: either a real advantage, or a unit operation you forgot. Finding where that delta sits is the most useful thing this check does

Two habits keep it honest:

• Keep a sanity band in mind — an ~11% margin looks plausible; a 55% margin is a no-free-lunch flag to zoom in on, not a victory
• Show your work inside the model — record where each number came from right in the sheet (“vendor call: 35–37”). It builds credibility and the team’s memory

A factored estimate is ±30–50%, so a waterfall and a few bracketed assumptions are easy to put in front of someone who knows the field. You can’t Google your final answer, so outside eyes are the test that’s left: walk an advisor through it and ask what looks wrong.

🧭 Coach’s Read

Don’t run a sensitivity on a number you haven’t gut-checked — a tornado on a 10×-wrong figure is precise nonsense, and the usual culprit is a unit slip the model will never flag itself. Recompute by a genuinely different path; the same spreadsheet with one cell changed just reproduces your own error. Benchmark like-for-like — same boundary, cost year, capacity basis — and never benchmark against the incumbent figure you seeded the model from. That check is circular and tests nothing.