Explain the risk-adjusted ROI of an AI project with uncertain feasibility.

Bindrose is Corvantis Biosciences' AI molecule screener. It reads a target protein's shape and ranks candidate compounds by how likely each one is to bind it, before a single molecule gets made in a lab. Bindrose has a real record on 40 ordinary targets. Auger isn't ordinary: a shallow, mostly featureless pocket that no lab has ever found a working small-molecule binder for, by any method. Wenke Delacoste owns Bindrose's numbers. Osei Bramwell, who runs portfolio strategy, needs a fundable case for screening Auger anyway, and won't take a number with no arithmetic behind it.

How to answer this, stage by stage

Nobody's grading whether you can name a probability out loud. They're grading whether you'll multiply instead of hope, and whether you know a confident single number is a warning sign when nobody's ever tried the thing before.

Let's learn

Here's what happens when a research committee asks a scientist for one number, and gets handed a range instead.

Bindrose is Corvantis Biosciences' AI molecule screener. Give it a target protein's shape, and it ranks candidate compounds by how likely each one is to bind that target, weeks before anyone synthesizes a single molecule to test in a lab.

Before Bindrose, a target like this got triaged by hand: a small team of chemists working from literature analogies and gut feel, picking maybe 500 candidates to actually make and test, at real cost and real time, with a hit rate under 3%. Bindrose changed that math on the 40 targets it's screened so far. It ranks about 2 million virtual candidates down to a shortlist worth testing in a real lab, and roughly 30% of its top-ranked candidates go on to validate as genuine binders.

Forty targets in, Bindrose's edge on a normal, well-defined pocket isn't a guess anymore. It's a track record.

Auger is not a normal target. Every target Bindrose has ever screened has a deep, well-defined pocket, a clear place for a small molecule to grab onto. Auger's pocket is shallow and mostly featureless, the kind of shape the field calls undruggable, because no lab, using any method, has ever found a validated small-molecule binder for a pocket like it.

Bindrose's 30% hit rate belongs to the left panel. Nobody has evidence it belongs to the right one too.

Here's the turn. The real problem with Auger was never that Bindrose might fail on it. Plenty of ambitious science fails, and a portfolio built on nothing but sure things isn't much of a portfolio. The real problem is what happens when a comfortable, unsourced number stands in for that risk in front of the people deciding whether to spend real money. Eighteen months earlier, a different team pitched a different novel target at 70% feasible, no stated source, because 70% sounded fundable and 15% didn't. Corvantis spent $4.1 million over nine months. Zero validated hits. Nobody could ever say afterward where the 70% had come from.

What it costs at its worst: a fundable-sounding number that turns out to have no arithmetic under it doesn't just waste a budget line. It burns the one thing a small computational team actually trades on, which is being believed the next time they ask for money.

What I'd leave alone: Bindrose's ordinary, well-defined-pocket targets. With 40 of them behind it and a real hit rate to show, feasibility there is close to certain. Building an elaborate range for a number that's already solid isn't rigor, it's theater.

The lesson: an honest range costs you comfort in the room, not credibility. The number that survives a hard question isn't the one that sounds most sure of itself. It's the one that shows its work.

Now here is the same thing as a story

Read the short version above when you're in the room. Read this one when you want to feel why a number with no source is a bigger risk than the science it's supposed to describe.

Every Tuesday afternoon, Wenke Delacoste opens the same tracker: forty rows, one per target Bindrose has ever screened, every single one with a real, well-defined pocket to grab onto. Ask Wenke which of those forty nearly missed a real binder, and they can recite the assay number from memory.

For most of the last two years, that tracker was the whole job. The committee stopped asking Wenke to walk the arithmetic on an ordinary target; they trusted the topline number on the slide, because the topline number had earned it, target after target.

The business cases got shorter to match. The first ones ran a full page: assumptions, sources, a range with reasons. A year in, most had shrunk to a paragraph. By the time Auger came up, the draft in front of Wenke was a single line: "Feasibility: high. Recommend proceeding."

Then came the hallway conversation. Osei Bramwell caught Wenke near the elevator, not with a data pull, not with an audit, just a remark. "You know what happened to Ferro, right? Nobody in that room wants that number again."

Wenke knew exactly what had happened to Ferro. Everyone did. A single feasibility figure, 70%, that had felt right in the room because the science team was excited, and that nobody could actually source when the money was already gone. Nine months. $4.1 million. Zero.

Not a guess about four months. A plan for what each month buys, and what it costs to find out.

So Wenke deleted the single line and started over, the slow way. Nine prior attempts at a first tractable hit on a genuinely novel target, company-wide, any method: one success. That's an 11% base rate, sourced, checkable, unflattering. Then the one piece of real evidence Bindrose actually had: a backtest against the two closest analog shallow-pocket structures in public data, where the model's ranking did better than random, not proof of anything, but a real, if weak, positive signal. Blend the two honestly and you land at 18%, not 70%, and not a comfortable round 50% either.

The old decision, the one Wenke would take back, wasn't a design choice on Bindrose at all. It was what happened in the room where Ferro's number got set. Someone asked, "where does 70% come from," and nobody had a real answer, and nobody pushed, because pushing back on an excited science team felt like betting against the science itself. The number stood because no one made it earn its place.

Run that meeting again, with one rule added: every number states its source out loud, or it doesn't go on the slide. Same Auger opportunity, same pressure to move fast. This time, when Osei asks where 18% comes from, Wenke has it in one sentence, and the range on the slide reads $4.8 million to $18.8 million, not a single confident figure with nothing under it.

One habit let a number's confidence stand in for its evidence. The other makes the evidence show up as a range, even when the range is a little uncomfortable to look at.

What Wenke would tell their past self, watching that Ferro meeting happen and saying nothing: an unsourced number isn't a smaller risk than a range. It's the same risk. It just waits until the bill comes due to show itself.

BOUND, or how to price a bet nobody's placed before

Not a story wearing a framework's clothes. This is an estimation problem where the hardest number in the equation is also the one nobody at the company has real data on, and BOUND is what turns "it's uncertain" into a figure a portfolio committee can actually act on.

Three real numbers and one habit. The habit, using a range instead of a point on the number that's actually uncertain, is the whole method.

The whole answer to this question in one picture. A number that's only true if you're lucky was never the same thing as a number that's true either way.

Feasibility sits exactly where you don't want your biggest lever to sit: least grounded, most consequential.

Three things worth stating directly, since this is where the real judgment sits. The AI-specific failure mode is silent out-of-distribution overconfidence: Bindrose's scoring model was trained almost entirely on deep, well-defined pockets, so nothing in its architecture flags "this input doesn't look like anything I've seen" when it's handed Auger's shallow, featureless one. It can still hand back a confident-looking score built on spurious geometric similarity rather than any real binding signal. The guardrail is simple and non-negotiable: no Bindrose score counts as evidence for Auger on its own. Every one of the top 250 candidates has to clear a real wet-lab binding assay before it's called a hit, and the pilot tracks calibration against the two held-out analog structures the whole way through, not just the final ranking. And the quality-versus-cost trade-off is real: Wenke could have screened a wider 10-million-candidate pool and validated a cheaper 80-candidate batch to save budget, but a smaller wet-lab sample gives a noisier read on the true hit rate exactly when the feasibility range needs to be trustworthy. The larger, fully-validated 250-candidate batch costs more per candidate. It's also the only version of this pilot whose answer the committee can actually act on.

And if you want to be sure it really works, try it somewhere else

Same five letters, a battery lab instead of a wet lab, and this time the unproven thing isn't a binding pocket. It's a whole class of chemistry nobody's gotten to hold together yet.

Ionvale is a battery-materials company. Its screening model, Cellwright, ranks candidate solid electrolyte formulations by predicted ionic conductivity and mechanical stability, the two properties a real cell needs to actually work. Cellwright has a solid record on Ionvale's existing electrolyte families. What it's never been tried on is a genuinely new solid-state chemistry class the R&D team wants to move to, one that promises a real jump in energy density if it can be made to hold together at all.

Same BOUND, a different lab, a different kind of unproven. Both times, the model's architecture itself is the thing nobody's tested yet.

Sigbjorn Rooijakkers owns Cellwright's ROI case, and ran the same five letters Wenke did. Break it down: probability the model can usefully screen this new chemistry class, times the payoff if it can, minus the cost of a physical pilot. Own the numbers: 2 successes in 7 prior "new chemistry class" screening attempts, about 28% feasible, against a $45 million payoff, mostly avoided bench trial-and-error and a faster path to a production-viable cell, minus a $1.1 million pilot to fabricate and test the top 40 candidate formulations. Use a range: 20% to 40%, putting the risk-adjusted value between $7.9 million and $16.9 million, not the $43.9 million a naive sheet would show.

Same rank, different lever: Ionvale's fix isn't a bigger physical-testing team. It's the same habit, run on Cellwright's own numbers: source the base rate, range the number nobody has real precedent for, and watch feasibility, not the payoff estimate, as the assumption that actually decides the answer.

Swap the trigger and it still runs.
Speed: an interviewer caps you at ninety seconds. Skip straight to it: multiply probability by payoff, subtract the cost of finding out, range the number that's genuinely uncertain, and fund the small bet before the big one.
Cost: there's no time this quarter to build a sourced base rate from scratch. Ship the honest version anyway, a wider range built from whatever real data exists, revisited once the pilot itself produces better numbers.
The model got better, for real: say Bindrose's backtest signal on shallow pockets turns out stronger than expected. The range narrows and shifts up, but the habit doesn't change. A good result updates the number. It doesn't excuse skipping the range next time.

Where people run it wrong.
They report a single confident-sounding probability because a range reads as less certain, exactly the mistake that cost Program Ferro $4.1 million.
They range every term evenly instead of separating the well-grounded numbers from the genuinely uncertain one.
They compare the risk-adjusted number to nothing, so nobody in the room can tell if $10.4 million is a strong return or a red flag.

How to use it live. Ask the source question before naming a number: "is this probability estimate coming from a real base rate and real evidence, or is it a feeling that happens to sound fundable?" That question alone usually tells you whether you're about to defend a real number or repeat someone else's Program Ferro.

Flashcards (tap any card to flip it)

Check yourself Score: 0 / 0