The Right to Stop

The practical difference between the two categories, risk and uncertainty, is that in the former the distribution of the outcome in a group of instances is known… while in the case of uncertainty this is not true, the reason being in general that it is impossible to form a group of instances, because the situation dealt with is in a high degree unique.

— Frank Knight, Risk, Uncertainty and Profit, 1921

This essay is written for the seat that signs the capital release: chair, CFO, or BD head. The argument generalizes; the cases are pharma.

The allocator has seen hundreds of these slides. The eNPV deck is up, and she knows before the analyst finishes the build that the only honest number on the page is the discount rate. Nobody in the room believes the distribution; they believe it the way you believe a weather forecast you have already decided to ignore. The partnership is one of a kind, the comparators unlike each other, the capability being rented certain to be repriced by a model release nobody in the building controls before the contract is a year old. The discount rate is real, the milestones are real, the signature ninety seconds away is real. The distribution is invented to make the meeting move. She looks for the one row she would actually sign for: the line that says when they stop. There is no such row. No date for the kill, no name beside it. She signs anyway. Everyone does.

---

Essay 01 was about what you are pricing. This one is about a worse problem: what you do when the thing cannot be priced at all. The answer is not a better number but a design for the one part of the bet the firm controls, the condition under which it stops. Frank Knight named the difficulty in 1921. Risk is where the distribution of outcomes is known: the molecule whose Phase 2 readout is priced against a success rate of roughly one in three, measured across thousands of drug-development paths. Uncertainty is where the distribution itself is unknown, the reference class too thin or too heterogeneous to estimate. Keynes put it without the taxonomy: of the prospects that decide a firm’s future, “there is no scientific basis on which to form any calculable probability whatever. We simply do not know” (1937). His conclusion was not paralysis. We act anyway, on convention; the danger is mistaking the convention for knowledge.

An AI partnership signed by a top-twenty pharma in 2026 sits on the uncertainty side for three reasons that compound. It is a partnership of one, a bet that alters the firm’s own conditions so completely no reference class can form. Its comparators are too heterogeneous to pool, though not in the way the trade press assumes: model architectures have largely converged, and what differs deal to deal is the data rights, the evaluation harness, and who owns what the training produces. And the dominant uncertainty is not the molecule the partnership might produce but the capability frontier it rents, which reprices faster than the deal amortizes.

The standard ROI machinery treats uncertainty as a measurement problem: estimate the distribution, discount it. When the distribution is known, the math is honest. When it is unknown, the math is not wrong so much as not applicable, and I should be careful with that phrase, because it is the strongest claim here. A thin distribution is not no distribution. A bad number a board anchors and stress-tests may still beat walking in with only a stopping rule, which makes the slide dangerous rather than useless. I press the stronger charge because the failure mode I keep watching is treating the scrap as an answer. But the CFO who refuses to concede the distribution is fully unknown is not wrong, and what follows has to survive his version. A board that signs a ten-figure partnership against that slide has not priced a bet. It has priced a forecast.

This matters more in 2026 than in any prior decade, because the cliff is forcing capital decisions on a clock the science is not setting. Roughly three hundred billion dollars of branded revenue goes off patent across the six years to 2030, cumulatively, with the worst single year around a hundred and four billion in 2028. That is about a sixth of the industry’s revenue, and the largest such concentration it has faced. Every AI partnership signed in these two years is sized against that hole by firms simultaneously running the largest BD spree of the cycle, which argues for slower, smaller stages rather than faster, larger ones.

---

The substitute, when the distribution is unknown, is a prespecified rule that stops the spending without asking the spender’s permission. Every serious capital domain already runs a version, and pharma runs the best version of all, everywhere except the line this essay is about.

Start inside your own house. A Phase 3 trial does not rely on the sponsor’s judgment about when to quit. The protocol carries a prespecified futility boundary written before the first patient is enrolled, and an independent Data Monitoring Committee reads the data against it under a charter, seeing unblinded data the sponsor cannot see. Formally the DMC advises and the sponsor decides. The advice is hard to refuse because it does not arrive privately: the recommendation is visible to investigators, to institutional review boards, and in substance to the regulator, so overruling it is an act the sponsor has to defend to three constituencies it does not control. Criterion written in advance, read by a body with no career stake in continuation, and a refusal that costs something in public. That is the entire argument of this essay, and pharma has run it as routine for decades. Note what makes it work, because it is also what makes it hard to copy: a regulator, a charter and a patient are all standing outside the sponsor, and none of them can be talked round in a management meeting. Everything that follows is about what happens as those outsiders drop away.

Now widen the frame, because the pattern is not pharma’s.

Defense procurement has run a statutory version since 1982. Under the Nunn-McCurdy provisions, a program whose unit cost breaches twenty-five percent over its current baseline is presumed terminated unless the Secretary of Defense certifies otherwise to Congress. Banking supervision writes prespecified triggers that suspend a firm’s own discretion and oblige a named executive to act on the record. Project finance calls a default on covenant breach, which stops the facility until the lenders choose to waive it. Venture simply stops paying when the round runs out, and continuing requires a new investor to reprice the company. In every one of them the default halts and somebody has to act affirmatively to continue. On pharma’s capability line the default is renewal, the party who decides is the party who signed, and nobody outside has to price anything.

And here is the honesty in it. Defense has run the strongest version of this instrument for more than forty years, statutory and quantitative, default-to-terminate, a cabinet officer on the certification, congressional visibility. It mostly does not terminate anything: of the thirty-seven major programs the Congressional Research Service counted with significant or critical breaches since 2007, four were terminated. The rest were certified and restructured.

That record does not argue against writing kill triggers. It is the strongest evidence for the claim this essay makes: writing the trigger is necessary and nowhere near sufficient. A statutory stop rule, with a default that terminates and a cabinet officer who must sign to continue, mostly does not terminate. The presumption bites only on the critical breaches inside that count, so the true ratio is kinder than four in thirty-seven. It is not kind enough to matter. Anyone who thinks a clause does the work alone should read it twice.

Which sharpens what the venture comparison is for. Venture is not disciplined because venture partners have unstable judgment; the evidence runs the other way, and returns persist strongly across funds raised by the same partnership (Kaplan and Schoar, 2005). Venture is disciplined because of which way the default points. When the round is out, the money stops, and continuing requires a repricing: a new investor, or the existing ones on fresh terms. Either way the number is set by what somebody will pay now, not by anyone’s prior conviction. Pharma’s capability line inverts every one of those properties. The default is renewal. The party who decides is the party who signed. Nobody outside has to price anything.

At top-twenty scale the discipline has to be built deliberately: a kill trigger named in the board-approved memo, fired against a dated observable, with an officer’s name attached, and not the officer who signed the deal.

Real-options theory priced the right to stop three decades ago, and McGrath and Nerkar tested that reasoning on pharma R&D in 2004 and found it holds. Annie Duke’s Quit charts the individual version. What neither settles is the institutional problem: who fires the trigger when the decider is a committee spending money that is not its own. Staw settled the empirics. The person most likely to fund a failing course of action is the one personally responsible for starting it, and self-justification rather than new evidence drives the escalation (Staw, 1976). A stop owned by the signer is the stop least likely to fire.

So the trigger goes elsewhere, and “elsewhere” is not one committee, which is where most versions of this proposal, including my own earlier one, go wrong. Audit is the reflex answer because it is independent of the sponsor. But audit is staffed with financially qualified directors and chartered to give assurance over management’s assertions rather than to hold operating decisions. No audit committee on earth can adjudicate whether a model’s top-decile targets survived orthogonal validation. Hand it a scientific observable and the trigger dies of incompetence rather than capture.

The routing that works uses machinery boards already have. Four bodies, four jobs. The science committee adjudicates the observable: did the evidence arrive, against the threshold, by the date. It also holds the right to amend the criterion, on a minuted resolution and never at the sponsor’s initiative, which is the provision the rest of this depends on. Audit attests the control: not whether the science passed, but whether the trigger existed, was tested and was reported. It is the only committee with an independent staff arm, which makes it the right verifier and the wrong owner. The chair escalates, so a missed trigger reaches the full board rather than dying inside the committee that owns the program. Compensation carries the consequence, which is allocative rather than disciplinary: no top-twenty board has fired an officer over a missed AI kill trigger, and no industry dismisses an executive over one missed metric. It sits in the long-term incentive scorecard, which is the compensation committee’s to write. Not the clawback policy, which recovers pay after a restatement and cannot be fired by a missed operational trigger. The scorecard can.

Seen this way the kill trigger stops being a penalty and becomes a scheduled transfer of the decision right: from the sponsor, on a date named in advance, to a body that did not sign the deal. The contractual vehicles already exist: the option-to-acquire structure pharma has used on biotech for a decade, and the structured platform license with staged, milestone-gated terminations. And every one of these agreements can already be terminated, for material breach, for insolvency, for change of control, usually for convenience on notice. That is not the gap. The gap is that termination is written throughout as a right the firm may exercise, and nowhere as a default the firm has to argue its way out of on the evidence. No clause anywhere says what would have to be observed, by when, for the deal to end without anyone deciding to end it.

The trigger has to fire against the right observable, so the deal has to be honest about what capability it is renting. A foundation-model partnership in 2026 rents against one of three constraints.

Structure and design (AlphaFold descendants, RFdiffusion, generative chemistry) rents time-to-credible-candidate. The naive trigger fires on whether candidates entered the firm’s translational pipeline by month nine, which is a throughput measure on the one constraint where throughput was never binding. It needs a quality leg: how many survived the firm’s own development-candidate gate. A trigger that counts arrivals passes on a flood of compounds that die internally.

Target identification and biology reasoning (multi-modal models on omics, cell painting, perturb-seq) rents prior on target tractability. The trigger fires on whether the model’s top-decile targets survived orthogonal validation on the firm’s named assay panel at a rate above the firm’s own historical base rate. “Entered validation campaigns within twelve months” is a bar designed not to fire.

Clinical and real-world-evidence integration rents patient-stratification fidelity. “Different, defensible enrollment criteria” is not signable. What a biostatistician will sign is that the model changed the powering assumptions in a filed protocol or statistical analysis plan.

Each observable needs a threshold, not only a date, calibrated against the firm’s own attrition history rather than borrowed from the model’s benchmark scores. A trigger with a date but no threshold is a meeting, not a gate.

And now the part that makes this specifically hard for AI, which I did not see until I tried to write the triggers down. The evaluation harness is endogenous. It is co-designed with, and often trained against, the party it is supposed to measure. Escalation of commitment in Staw’s classical form shows up as a sponsor defending a failing program. In a capability partnership it shows up one layer up, and it is far harder to see: the sponsor does not defend the program, the sponsor revises the benchmark. Nobody argues the evidence was good enough. Somebody merely observes that the harness was the wrong harness, that the field has moved, that the 2024 metric does not capture what the 2026 model does. Each of those will sometimes be true, which is what makes it the perfect instrument of escalation.

Hence the one clause here with no analogue in the molecule world. Call it the harness clause: the harness is specified by the firm and versioned at signing rather than adopted from the partner’s benchmark suite; a held-out slice is never shipped to the partner, and it is rotated between determinations, because the partner learns your gate from the pattern of what you accept even when no file ever changes hands. Withholding the file is easy. Withholding the function is not, and a firm should know which of the two it has actually bought. And revising the harness requires the trigger officer, not the sponsor, as a governed act with reasons minuted. That last provision generalizes past AI. If escalation travels through criterion revision rather than criterion violation, the governed object is not the criterion. It is the right to amend the criterion.

None of this is hypothetical. Pharma already writes prewritten, quantitative go/no-go criteria on its own molecules, and has run them at portfolio scale for two decades. What it has never done is point that machinery at the capability line.

So why does a discipline this well proven stop at the molecule boundary? Not because anyone exempted the AI bet on purpose. Because enforcement decays, along one axis: who is allowed to amend the criterion.

An in-licensed molecule on staged payments stops reasonably well. You cannot amend the criterion. It is written into somebody else’s contract and the next tranche is somebody else’s money.

An internal program stops badly, and it stops badly despite carrying a written criterion. Here I have to correct something I have said loosely for years. Programs do not run past their target product profile. A TPP is a design aspiration, and the falsifiable stop inside it is the minimum acceptable profile. What happens is worse and narrower: at the readout, the profile gets revised. The minimum acceptable becomes acceptable. Nobody violates anything. The criterion and the authority to amend it sit in the same organization, usually the same function, and amendment is quiet, reasonable and documented. The practitioner’s version, having sat on both sides of it: killing a live internal program takes a mountain. Evidence, air cover, and someone senior willing to spend capital on a negative. Revising the criterion takes a meeting.

A capability investment has no criterion to amend, and the frontier amends it for you. A molecule has an agreed object of failure: it meets the profile or it does not, and everyone in the room knows which. A capability has no such object, and when it is hyped the object moves on purpose: any evidence of underdelivery can be answered with the next model release. The reputational asymmetry compounds it. Killing a molecule is normal science and no one’s career turns on it. Killing your AI partnership in 2026 reads to the market as falling behind. This is why the AI bet arrives dressed as a capability purchase, booked against a procurement mental model that never asks for a kill trigger, rather than as the R&D bet under uncertainty it is.

Three tiers, one axis. The capability line, where amendment is unconstrained, is where the largest unhedged commitments of this cycle are being written.

The obvious rebuttal is that the deal sheet already solves this. Lilly’s collaboration with Insilico put roughly four percent of headline value up front, so almost everything in it is contingent. And contingent value does sometimes fail to vest: Bristol Myers Squibb’s collaboration with Schrödinger walked down from a two-point-seven-billion milestone ceiling to four hundred and eighty-two million on a single remaining neurology target, visible only in Schrödinger’s quarterly filings. Things do get killed. What followed the kill is the tell. In August 2026, the same day that filing appeared, the two companies announced an expanded AI collaboration disclosing no financial terms, no milestones, and no termination criterion at all. The sponsor did not revise the criterion; it re-signed the relationship in a form that has none. Escalation has a third form, after violating the criterion and moving it: re-contracting out from under it, onto the capability line, where there is no criterion to amend.

But look at what those structures buy. A milestone gate is an option to stop. A kill trigger is a commitment to stop, and what it removes is the sponsor’s discretion not to. That distinction is the whole argument: discretion is what fails. A milestone gate leaves the decision with the sponsor, which is exactly where Staw says it will not get made. Sometimes it does anyway. Across the largest partnerships of this cycle, no disclosed term says what would oblige them to. The public record does hold one clean instance of the difference. When Recursion halted REC-2282 in May 2025, the trial’s lead cohort had passed its prespecified futility analysis. The company discontinued anyway, citing the totality of the data and other programs more deserving of the resources. The criterion did not fail. It was simply not binding, because nothing made it so. The question is why nobody was willing to bind themselves to it.

Which exposes a tension in the title. A right is an option, and I have just argued options do not get exercised. It is worth having only in the form that removes your own future discretion, so the honest name for it is closer to an obligation the firm grants to someone else while it can still say what failure would look like. The instrument as usually imagined fires at signature, but escalation is a renewal phenomenon, when the sponsor has the most sunk conviction and the least appetite for a public reversal. A trigger not re-armed at renewal expires exactly when it becomes necessary. AbbVie’s Calico partnership was renewed in 2018 and again in 2021 and terminated in 2025 on a failed readout. Three decision points, and no criterion at any of them.

Every capability that lowers the cost of running an experiment lowers something else, and it is not the carrying cost of a program. FTEs, tox, CMC and the portfolio slot did not get cheaper. What collapsed is the cost of plausible continuation. Another benchmark, another cut of the data, another model version weeks away: the supply of respectable arguments for not stopping has risen faster than anything else in the system.

So the number of options a portfolio can rationally hold rises while the number it can retire does not. The binding constraint stops being the capital to fund the option and becomes the governance capacity to kill it. Capital at this scale is replenishable. Board attention is not. A firm that can hold forty live bets and retire none does not have a portfolio. It has a backlog it is still paying for.

That is measurable, and the crudeness is the point. Call it the kill-capacity ratio:

Decompose it by line: molecule, internal program, capability. Then by moment, readout versus renewal. Every input comes from data the company already holds. Most boards have never seen it, and the first to put it on the board pack will discover in about ninety seconds that its capability line has a denominator of zero.

This is also where the ergodicity argument belongs, and it is usually pointed at the wrong ledger. When losses compound multiplicatively, the average outcome across many firms tells you nothing about the fate of the one firm living a single path through time (Peters, 2019). But no single AI partnership is a ruin event for a top-twenty pharma. AbbVie put at least one and three-quarter billion into Calico and its balance sheet is entirely fine. The absorbing barrier is not the balance sheet but the governance system: once unretired commitments exceed what the board can adjudicate, the firm has lost the ability to stop anything, and that state is absorbing.

The whole problem compresses to one line. Stopping is nearly free for the firm and very expensive for the person who has to do it. Every device in this essay, the separate officer, the science committee, the disclosure, the LTI hook, is that sentence applied four times.

I have argued one side of this instrument, and anyone who has negotiated one of these agreements has been waiting for the other. Three costs are real.

The counterparty prices it in. A partner who knows you can walk at month nine will charge for that optionality, in a higher upfront or a tighter exclusivity. That is an argument for treating the price of the walk-away right as a line item rather than a surprise.

The observable becomes a target. Name a metric, attach money, and it gets staffed to. This is Goodhart with a contract attached, and it is why the harness clause holds the observable out of the partner’s hands. A trigger the counterparty can optimize always passes, which is worse than no trigger because it launders continuation as evidence.

Disclosure is a different animal from precommitment, and I have been sloppy about the difference. A trigger named in a board-approved memo is an internal governance act. A trigger announced in a press release is a market event: a dated catalyst for short sellers, a materiality question every quarter, a public reservation point in a live negotiation. Everything here is satisfied by the board-approved version, and that is the one I mean.

The counterparty’s case is not a courtesy either. From the founder’s seat a dated public observable on the cap table is a detonator, and an instrument written only for the buyer will anti-select: the partners willing to sign will be the ones with nothing to lose. So the instrument has to be symmetric, and the symmetric version is worth naming: the right to be stopped. If the firm walks at the trigger, the partner leaves with unencumbered rights to the target class rather than a dead program and an encumbrance; the trigger is disclosed to two boards and to no exchange; the observable is frozen at signing or held by a third party, so neither side can staff to it or quietly revise it. A partner’s willingness to be bound by that is the cheapest diligence signal available.

Set against those costs is the objection that all of this under-bets an industry-defining technology. It is the right objection, and the answer runs the other way: a firm whose downside is bounded by design can size the front stage larger, because the exposure is capped by structure rather than by nerve. The stop is what licenses concentration. Andy Grove’s version, written about Intel’s pivot out of memory, is the one boards should hear: at a strategic inflection point, the bet you do not make is itself a commitment, by omission, to a model the inflection is making obsolete. Staging buys the next bet.

The staging unit is a single allocation memo with three columns, and the order is the argument. First, the bet: the partnership, the named therapeutic thesis, the headline number. Second, the stop, carrying two things the standard memo never writes down: the prespecified kill trigger, written against the constraint the deal is actually renting and re-armed at every renewal; and the opportunity cost, meaning the internal program this release displaces and the BD line that now does not get funded. Every Commit at this scale is a kill of something the firm is already doing, and a memo that does not name that kill will not survive a real board. Third, the accretion: what the firm keeps if the bet dies at the trigger. The accretion is real, though whether it compounds or decays is exactly the question the frontier’s repricing raises, and either way it is the byproduct of disciplined staging rather than the reason to stage. A firm that funds the third column and skips the second has bought a substrate and called it a strategy.

Monday morning, the reader walks into the next AI allocation review and asks three questions in this order:

1. What is the kill trigger, what date does it fire, and who may amend it? If the answer is we will know it when we see it, there is no kill trigger. The deal is drift dressed as a Commit.
2. What does this stop us from doing, and have we named it? If the answer is it is incremental, the memo is hiding the opportunity cost, and the board cannot weigh it.
3. What do we keep if the bet dies, and whose name is on it? If the answer is different teams, different cost centers, no single sponsor, the firm is buying a subscription and reporting it as an investment. Then ask all three again from the other side of the table.

The discipline reduces to one rule a reader can carry into the meeting alone: no signature without a written right to stop. The chair’s version is a standing question to the science committee, attested by audit: which of our AI partnerships carry board-approved kill triggers, when was each last re-armed, have we tested that they would fire, and what is our kill-capacity ratio on the capability line?

In most boardrooms the honest answer is that none do. One count puts the first quarter of 2026 at thirty-six major pharma AI partnerships against eighteen a year earlier. They disclose milestone schedules. Almost none discloses a prewritten kill trigger. The claim has to stay narrow, because private contracts may carry termination provisions the press releases omit. But that gap is exactly what a board has to read past the slide to find, and deal flow doubling in a year without the discipline appearing is itself the finding.

One number in my own source list gets read against me. BCG puts Phase 1 success for AI-native pipelines at eighty to ninety percent, far above the historic rate. The same analysis puts Phase 2 at roughly forty percent and calls it comparable to historic industry averages. The advantage evaporates exactly where biology is tested, which is where the right to stop earns its keep.

The larger doubt is whether this machinery fits all of these bets, or only the ones where the unknown is still the distribution rather than the regime the distribution lives in. AI-bio outcomes in 2026 may sit in a third regime, where the trigger’s underlying observable has been redefined by the time the trigger fires. If that regime is real the discipline survives, because write a kill trigger still holds, but the trigger written today may not mean what it meant by the time the world is ready to fire it. The instrument I am least sure of is the one I have leaned on hardest.

And I should turn this on my own layer before someone else does. Advisory work of the kind I do is sold on judgment and renewed on relationship, which is to say it runs without a criterion too. The rule applied to my own seat would read: an adviser who cannot say in advance what would make the engagement not worth renewing is selling the drift this essay has spent nine pages describing. I have not always written that down either.

So, a forecast, scored on partner filings rather than press releases, because public disclosure is the tape this essay has just told you to distrust. Take the first fifteen to twenty new top-twenty AI partnerships signed after the first of September 2026, and read them through to the second quarter of 2028. Either they wind down, if they wind down, through quiet milestone non-pursuit with no evidence of a prespecified trigger. Or one to three show a dated, criterion-linked termination, and a vanguard has emerged. I expect the second. I will be wrong in a direction the record will make legible.

---

Pricing under genuine uncertainty is neither aggression nor caution. It is the discipline of writing down what would make you stop before you start, and then putting the pen out of your own reach. The firm that does this is not the firm that bets least; it is the firm that still has the right to bet after the first three bets are wrong. That right is the asset the eNPV slide cannot show, because the slide prices a forecast and the right to stop prices a decision.

Every capital domain that allocates under uncertainty has learned this and written it down: defense in statute, banking in regulation, project finance in covenants, venture in the structure of the round, pharma in the charter of a data monitoring committee. The capability line is the last one where nobody has. It is also, this decade, the one carrying the most money.

A firm that prices uncertainty this way is measuring itself against something, and which measurement still lives on the board pack is where the discipline meets its next test. Productivity is the wrong scoreboard, and that is the next essay.

The arc this series traces says some decisions compound into a lead that cannot be bought back. The right to stop is how you keep from making the one you cannot take back. It is also the criterion the next generation of pharma CEOs will be judged on by their boards: not whether they could sign a foundation-model partnership, but whether they wrote the right to stop into it, and whether anybody other than the signer was allowed to change it.