From Stage Gates to Belief Flow: Continuous Decisions and Dynamic Pricing for R&D in the Age of AI
Treat decisions as investments and a stage gate is an option that can only be exercised on a few dates. AI lets beliefs update almost continuously with the evidence, but updating more often doesn't create value by itself; value comes from a fuller state, better-timed decisions and better-suited holders, net of the costs and frictions they bring.
science
Decisions in biopharma R&D cluster at a few points: candidate nomination, the IND filing, the start and readout of each clinical phase, the marketing application. Between those points data keep accumulating, but a project’s valuation and resourcing usually stay put until the next review re-prices it. One intuitive way to put it: R&D decisions today are like a “coarse-grained Markov process”, in which projects jump between a few stages; AI looks as if it could turn this into a continuous Markov process. This piece tests that intuition. The question is: if judgments about a project could update almost continuously with the evidence, what would change about the value of R&D assets and the way decisions are made — and what would it mean for investment behaviour?
It is worth asking now for two reasons. First, Bayesian methods and statistical decision theory are mature, and adaptive designs and platform trials now have regulatory precedent. Second, AI is lowering the human cost of organising evidence, maintaining models and updating judgments; things that used to be done “once a period” because attention was limited can start to be done continuously.
The article has two halves. The first sets out the structure of the problem, reviews the existing theory and practice in real options, sequential decisions, continuous-time asset pricing and clinical trial design, and explains which parts of the intuition above hold and which need correcting. The second proposes a framework, which this piece calls belief-flow pricing, gives several propositions about it, and then discusses its effects on investment and dealmaking, its risks and limits, the organisational mechanism, the AI infrastructure and a path to implementation. The two numerical examples come from illustrative models we wrote ourselves; they show mechanism and direction only and do not represent data from any real project.
Part one · 1. Why R&D decisions come in blocks
The stage-gate management that drug companies widely use was first systematised by Robert Cooper for new-product development: a project is divided into stages, and at the end of each a review decides whether to continue, stop or rework (Cooper 1990, Business Horizons[1]). In drug development these “gates” correspond roughly to target validation, lead optimisation, candidate nomination, IND, Phases I to III and the marketing application.
- Stage gate: dividing development into stages, with review points only between them, at which a review decides whether to commit the next stage’s resources. Its implicit premise is that important information arrives at the end of stages, and that no decisions are needed between reviews.
Behind this structure lies a set of high-risk, long-horizon numbers. By BIO, Informa and QLS’s analysis of clinical programmes from 2011 to 2020, the overall probability of approval from entering Phase I is about 7.9%, taking about 10.5 years on average (BIO 2021[2]). Wong, Siah and Lo, estimating from 2000–2015 data with a different method, found an overall success rate of 13.8%, and only 3.4% in oncology (Wong et al. 2019, Biostatistics[3]). The former takes drug–indication pairs as its unit and multiplies phase-by-phase transition rates; the latter tracks development paths one by one and imputes missing phase information, so the two have different denominators. By DiMasi and colleagues’ estimate, the R&D cost per approved new drug, counting failures and the time cost of capital, is about $2.56 billion (2013 dollars, excluding post-approval R&D) (DiMasi et al. 2016, J Health Econ[4]). That even the industry’s most-used success-rate benchmarks differ this much shows that the “prior” itself carries considerable uncertainty.
Decisions come in blocks for at least four reasons. They differ in nature, and the rest of the piece treats them separately.
First, information itself comes in blocks. The most important efficacy evidence in a clinical trial appears at pre-specified analysis points; event-driven endpoints only become statistically meaningful once enough events have accrued. This blockiness comes from biology and trial design, and doesn’t disappear by looking more often.
Second, cognitive and organisational bandwidth is limited. Herbert Simon pointed out that real decision makers, limited in their capacity to process information, seek what is “good enough” rather than optimal (Simon 1955, QJE[5]). Christopher Sims modelled attention as an information channel with limited capacity, through which decision makers rationally choose to look less, and more coarsely (Sims 2003, J Monetary Econ[6]). Periodic reviews are a sensible arrangement under attention constraints: they concentrate scarce senior time on a few points.
Third, statistical and regulatory integrity. Repeated significance testing on accumulating data inflates the type I error rate substantially (Armitage et al. 1969, JRSS A[7]). Confirmatory trials therefore require interim analyses and decision rules to be pre-specified in the protocol and statistical analysis plan; outside the pre-specified analyses, the sponsor should not see efficacy data unblinded by arm. This blockiness is institutional design, meant to protect the credibility of inference.
Fourth, capital and contracts come in blocks. Venture capital invests by round, licensing deals pay by milestone, boards meet quarterly. These arrangements reflect transaction costs and verifiability: an event such as “Phase II met its primary endpoint” is easy to write into a contract; a continuously changing posterior probability is not.
These four sources mean that what AI can directly ease is mainly the second; the first can only be improved through better experimental design and earlier-available data; the third and fourth need matching institutions and contracts. Lumping them together under “continuous decisions” makes it easy to overestimate what AI can do.
2. Decisions as investments: from net present value to real options
In valuation, the industry’s standard practice is risk-adjusted net present value (rNPV): each stage’s costs and the cash flows after success are discounted stage by stage by the stage success rates. The textbook rNPV has an often-overlooked mathematical structure: it implicitly treats a project as a Markov chain jumping between “preclinical, Phase I, Phase II, Phase III, filing, launch”, with each transition probability set to the historical average for similar projects, and assumes that a project always continues unless it fails technically. This is the “coarse-grained Markov process” of the introduction.
Real portfolio management is somewhat finer. Most large and mid-sized drug companies start from industry benchmarks and have cross-functional expert panels adjust individual projects’ success probabilities for how well validated the target is, whether a usable biomarker exists, the molecule type and the indication, updating them at annual or semi-annual portfolio reviews. But these adjustments are usually implicit: what evidence they rested on, how large they were and whether they were calibrated afterwards is seldom systematically recorded. Mathematically, it is still a jump chain built on stages and corrected by hand at a few points — and the valuation usually leaves out management’s options to stop, scale up or transfer along the way.
These limitations have long been discussed in real-options theory. The term “real options” is generally credited to Stewart Myers, who treated a firm’s future investment opportunities as options (Myers 1977, J Financial Econ[8]). McDonald and Siegel showed that when investment is irreversible and project value uncertain, waiting itself has value, and the optimal investment threshold lies above the point where net present value is zero (McDonald & Siegel 1986, QJE[9]). Dixit and Pindyck’s Investment under Uncertainty systematised these results and became the field’s standard reference (Dixit & Pindyck 1994, Princeton University Press).
The stage structure of R&D is naturally a compound option: completing Phase I buys the right, not the obligation, to run Phase II. Geske gave a method for pricing “options on options” (Geske 1979, J Financial Econ[10]), and a great deal of work since has applied compound-option ideas to drug development — for example, Kellogg and Charnes’s valuation of Agouron Pharmaceuticals with decision-tree and binomial methods (Kellogg & Charnes 2000, Financial Analysts Journal[11]).
Most important for this piece are the results on entry, exit and adjustment. Dixit showed that when both entry and exit have sunk costs, the optimal policy is not to switch back and forth at a single point; instead there are two boundaries with a “band of inaction” between them (Dixit 1989, JPE[12]). Abel and Eberly extended this to continuous investment intensity with adjustment costs and irreversibility (Abel & Eberly 1994, AER[13]). While the price stays inside the band, the optimal choice is to stay put even as information changes. This hysteresis will come up again and again; it shows that “continuous observation” and “continuous action” are two different things.
Real options are not widely used in practice in pharma; the usual criticisms are that parameters are hard to estimate and the models are opaque to management. That is exactly the change this piece discusses: only if the cost of maintaining a continuously updated probability model falls sharply do real-options ideas have a chance to move from annual valuation reports into day-to-day decisions.
3. The mathematical lineage of sequential decisions
“When to stop gathering information and decide” is a long line of research in statistics and operations research. A few of its milestones cover almost all the tools this piece needs (figure 1).
During the Second World War, Abraham Wald proposed the sequential probability ratio test: update the likelihood ratio with each new observation, stop and conclude once it crosses one of two boundaries, and otherwise keep sampling (Wald 1945, Ann Math Stat[14]). Wald and Wolfowitz then proved that, under certain conditions, it is the test with the smallest average sample size (Wald & Wolfowitz 1948, Ann Math Stat[15]). Wald’s 1950 Statistical Decision Functions brought estimation, testing and sampling into a single framework of “choosing actions under uncertainty to minimise loss”.
Bayesian decision theory took shape in the 1960s. Raiffa and Schlaifer’s Applied Statistical Decision Theory (1961) and DeGroot’s Optimal Statistical Decisions (1970) gave the standard account of how prior, likelihood, posterior and utility combine into decisions. Ronald Howard introduced the “value of information”: the value of a piece of information equals the expected payoff of the optimal decision with it minus the expected payoff without it (Howard 1966, IEEE Trans SSC[16]). Hidden in this definition is a corollary crucial for this piece: for a decision maker who cares only about outcomes, information that could not change any decision is worth nothing.
Richard Bellman’s dynamic programming gave a recursive method for solving multi-stage decisions. Karl Åström extended it to the case where the state cannot be observed directly, showing that one should then do dynamic programming with “the posterior distribution over the state” as the new state (Åström 1965, J Math Anal Appl[17]); Smallwood and Sondik gave a solution method for finite-horizon problems (Smallwood & Sondik 1973, Oper Res[18]). This is the partially observable Markov decision process (POMDP), in which the posterior distribution is called the belief state.
In continuous time, stopping problems become free-boundary problems. Shiryaev studied Bayesian optimal stopping for “detecting a hidden change point as quickly as possible”, where the optimal rule is to raise the alarm when the posterior probability crosses a threshold (Shiryaev 1963, Theory Probab Appl[19]). At the portfolio level, Gittins proved that a class of multi-project allocation problems can be solved by computing an index for each project separately and always investing in the project with the highest index (Gittins 1979, JRSS B[20]); but when switching between projects is costly, such index policies are generally no longer optimal (Banks & Sundaram 1994, Econometrica[21]).
These tools were applied to clinical trials themselves early on. The earliest thread even predates Wald: in 1933 Thompson proposed allocating patients according to the posterior probability that one treatment is better than another — the origin of today’s adaptive randomisation (Thompson 1933, Biometrika[22]). Berry and Ho derived sequential stopping boundaries for trials from decision theory (Berry & Ho 1988, Biometrics[23]); Chick and colleagues went on to handle delayed outcomes, where patients already enrolled but without outcomes change the optimal rule (Chick et al. 2017, JRSS B[24]). Research on “experimentation” in economics applied them to firms’ decisions: Moscarini and Smith studied how to choose experimentation intensity and stopping time jointly in continuous time (Moscarini & Smith 2001, Econometrica[25]), and Keller, Rady and Cripps studied experimentation when information arrives as Poisson “good news” (Keller et al. 2005, Econometrica[26]) — the latter closer to the jump-like evidence of clinical readouts. Bergemann and Hege analysed belief-driven stopping of R&D together with staged financing (Bergemann & Hege 2005, RAND J Econ[27]).
Taken together, the lineage makes the conclusion clear: “making decisions from the posterior on a nearly continuous scale” is not new in theory, and its optimal structure — boundaries, thresholds, inaction bands, indices — has mostly been characterised precisely. What is new is the conditions for implementation: maintaining a posterior used to need a dedicated statistics team and a great deal of manual work, so it could only be done for a few key trials; now the marginal cost is falling.
4. Restating the problem: a stage label is not a Markov state
Back to the intuition of “going from a coarse-grained Markov process to a continuous one”. Its direction is right, but two things need separating: whether the state is chosen correctly, and how finely time is divided.
The Markov property is not a property of a process alone, but of “the process together with the chosen state”. This has two layers. First, relative to information that includes all the trial data, the stage label is not a sufficient statistic. Two projects both in Phase II — one already showing a clear dose response and biomarker signal, the other with murky early data — obviously have different chances of future success. Second, even looking only at the sequence of stages, because projects in the same stage differ in underlying quality, the “lumpability” condition in the sense of Kemeny and Snell generally fails (Kemeny & Snell 1960, Finite Markov Chains): how long a project has stayed in a stage and whether things went smoothly before both carry predictive information, so the stage chain itself is not Markov. To be clear, rNPV gives the average prospect given the coarse information “only the stage is known”; its problem is that the information is too coarse (and, as section 2 noted, that it leaves out options), not that it gets the average wrong.
The state that does satisfy the Markov property is the posterior distribution over the project’s key unknowns, together with observable state such as stage, time and resources already committed.
- Belief state: the posterior distribution, given all current evidence, over a project’s key unknowns (effect size, safety, cost and time), from which measures such as probability of success are derived. Together with the observable state (stage, time, resources committed), it is sufficient for future decisions: if the model is correctly specified, once you know them, past data add no further information.
So the first improvement is to expand the state: add the belief state alongside the stage label. The stage still matters — it determines the remaining costs, the available actions and the regulatory requirements — but the stage alone doesn’t describe a project’s prospects. As the previous section noted, the industry does treat projects in the same stage differently; it’s just that the differentiation depends on experts’ implicit judgments. What expanding the state adds is making that differentiation explicit, reproducible and testable against outcomes. Only the second step is to make time finer.
“Continuous” also needs unpacking. This piece distinguishes three kinds of continuity (figure 2):
- Continuity of observation: how often data enter the system. Enrolment, safety events, biomarkers, competitor news and literature can come in almost in real time, provided data contracts and pipelines allow it.
- Continuity of belief updating: how often the posterior is recomputed. Once the model and data pipelines are in place, the marginal cost of recomputing can be very low; but the fixed cost of putting them in place is not — data integration, prior elicitation, model validation, and revalidation after standards of care and endpoint definitions change. What AI mainly lowers is the work of organising and aligning the evidence.
- Continuity of action: how often resource allocation is adjusted. This is constrained by switching costs, irreversible commitments, organisational rhythm and statistical integrity, and its optimal frequency is usually far lower than the first two.
AI mainly brings the first two close to continuous. Whether the third should be continuous is a question that has to be solved separately, and the answer is usually no: the optimal policy is mostly “observe continuously, update continuously, act discretely when a boundary is crossed”.
In continuous time, the evolution of the belief state has a clear mathematical form. Take the simplest case: a project’s efficacy has only two possibilities, “effective” and “ineffective”, and the observed signal is a Brownian motion with drift μ₁ if effective and μ₀ if not, with noise intensity σ. Then the posterior probability π that the project is effective is itself a driftless diffusion, with instantaneous volatility exactly π(1−π)·(μ₁−μ₀)/σ (Liptser & Shiryaev, Statistics of Random Processes): largest at π = 1/2 and smaller the closer it gets to certainty. If evidence is revealed in bulk at readouts, π jumps at those times, giving a mixture of jumps and diffusion. This is the precise meaning of “continuous Markov process” in this problem: what evolves continuously is belief, not stage.
This process has one more property that drives most of what follows: under the decision maker’s own prior and model (that is, averaging over a set of projects drawn from the prior), the posterior probability is a martingale — today’s best forecast of tomorrow’s belief is today’s belief. Two corollaries follow.
First, looking more often doesn’t change the expected belief. By the optional stopping theorem for martingales, however often you look and whenever you stop, the average belief when you stop equals the initial prior. Looking doesn’t make beliefs more optimistic or pessimistic on average; its effect shows only in the actions taken on it.
Second, information has a total budget. The expected quadratic variation of the posterior probability over the life of the project equals the variance of the final posterior, which cannot exceed the prior variance p(1−p) of the event “the project is effective”; the bound is reached only if the truth is fully revealed at the end. How much of this budget is spent, and how early, depends on how fast the data themselves are generated and how strong the signal is — not on how often we compute the posterior. Recomputing more often just shows the resolution of uncertainty in finer detail; it doesn’t speed it up. What makes the budget pay out earlier and more fully is experiments that make evidence available sooner and with a stronger signal: better biomarkers, more precise enrolment, endpoints that can be read earlier, shorter data lags.
5. How learning enters prices
Once a project’s value is written as a function of its belief state, valuation enters the territory of continuous-time asset pricing.
The basic tools of continuous-time finance were laid down by Black, Scholes and Merton in 1973 (Black & Scholes 1973, JPE[28]; Merton 1973, Bell J Econ Manag Sci[29]) and later systematised as dynamic asset pricing theory (Duffie, Dynamic Asset Pricing Theory). The branch most relevant here concerns how prices form when investors don’t know the parameters of an asset’s fundamentals and have to learn them from data. Detemple studied pricing in a production economy with incomplete information early on (Detemple 1986, J Finance[30]); Veronesi showed that learning makes prices react asymmetrically to bad and good news (Veronesi 1999, RFS[31]). Pástor and Veronesi found that investors’ uncertainty about a firm’s average profitability raises its market-to-book ratio, an effect that fades as the firm ages and investors learn (Pástor & Veronesi 2003, J Finance[32]); the mechanism is that, with compounding growth, firm value is convex in profitability.
The value of an R&D asset is also convex in belief, but for a different reason. For any fixed development strategy, expected value is linear in the probability of success π; it is precisely because the firm can stop when beliefs worsen, and scale up or transfer when they improve, that the value of continuing under the optimal strategy, V(π), is the upper envelope of these linear functions, and hence convex in π. The convexity comes from the options, not from the payoff distribution itself. If a project had to be carried through whatever the beliefs, V would be linear in π and fluctuations in belief would be worth nothing.
This gives a compact correspondence: a project’s value at time t, P(t), equals V(π(t), t), where V is the value of continuing under the optimal strategy given the current belief and stage. The textbook static rNPV amounts to computing this function only at stage gates, with the stage label in place of π, under the fixed strategy “continue if technically successful” — so it ignores both the state information and management’s options.
In this correspondence, the informativeness of a signal shows up as the volatility of belief. Because V is convex in π, if the model is calibrated, a more informative experiment spreads the distribution of beliefs at decision times more widely and so raises option value. Combined with the “information budget” of the previous section, “raising belief volatility” here means resolving uncertainty earlier and more fully, not adding risk out of nothing. Only volatility that comes from real information has this effect: overreaction caused by a misspecified model also makes beliefs more volatile, but it lowers value (see section 9).
To bring out the mechanisms, this piece’s illustrative models neither discount nor distinguish real-world from risk-neutral probabilities. To first order this is acceptable: clinical success and failure are mainly idiosyncratic risk and can be computed with real-world probabilities, leaving systematic risk to the discount rate on commercial cash flows. But rigorous pricing has to be done under a stochastic discount factor; Berk, Green and Naik showed that the systematic risk of a multi-stage R&D project changes as it moves through the stages (Berk et al. 2004, RFS[33]). Discounting would also raise the relative value of stopping early.
Another correspondence comes from option pricing itself. An option that can be exercised only on certain specified dates is a Bermudan option; one that can be exercised at any time is an American option. The former’s set of exercise dates is a subset of the latter’s, so it is worth no more; as the exercise dates get denser, its value converges to the American value. Under regularity conditions such as a one-dimensional diffusion and a smooth stopping payoff, the rate of convergence is known precisely: the value gap is of the same order as the (average) exercise interval h, and the shift in the optimal boundary of order √h (Dupuis & Wang 2005, Ann Appl Probab[34], in a setting where exercise opportunities arrive randomly); if the stopping payoff itself jumps (milestone payments, say), convergence can be as slow as order √h; and pricing American options by simulation works precisely by approximating them with sufficiently dense exercise dates (Longstaff & Schwartz 2001, RFS[35]). Stage gates are Bermudan: the rights to “stop”, “scale up” and “transfer” can only be exercised at the gates.
This corrects the earlier corollary that “looking more often doesn’t change the expectation”. Under a fixed strategy the expectation is unchanged; but when new decision times are added on top of the existing ones (the new grid contains the old), the firm’s set of available strategies grows, and expected value under the optimal strategy cannot fall. This rests on three premises: nested grids, no cost to adding decision times, and a single decision maker with a fixed objective. Merely shortening the interval without keeping the original times doesn’t guarantee higher value; and where several parties interact, more options can sometimes do harm (see the discussion after proposition 7 in section 7).
6. “Near-continuous” practices that already exist, and their lessons
In clinical trials and early development decisions, “looking repeatedly and deciding on accumulating data” has decades of methodological and regulatory history behind it. These are the most direct precedents for belief-flow decisions, and they also mark out its limits.
Group sequential designs responded to the repeated-testing problem that Armitage and colleagues identified. Pocock proposed using the same nominal significance level at each of several interim analyses (Pocock 1977, Biometrika[36]); O’Brien and Fleming proposed boundaries that are strict early and lenient later (O’Brien & Fleming 1979, Biometrics[37]); and Lan and DeMets’s α-spending function allows the timing and number of interim analyses not to be fully fixed in advance, so long as how the type I error is “spent” as information accrues is pre-specified, and when to look does not depend on the efficacy trend already seen (Lan & DeMets 1983, Biometrika[38]). The idea behind these methods is close to this piece’s framework: decision times can be flexible, but “what decision to take on what evidence” has to be written down beforehand.
Adaptive designs and platform trials go further. Donald Berry and others have long advocated updating the posterior continuously during a trial and adjusting randomisation ratios, doses and sample sizes accordingly (Berry 2006, Nat Rev Drug Discov[39]). I-SPY 2, a platform trial in neoadjuvant breast cancer treatment, adaptively randomises according to posterior probabilities in each biomarker subtype; when a drug’s Bayesian predictive probability of success in a confirmatory trial of about 300 patients reaches 85% in a pre-specified biomarker signature, it “graduates” to later confirmation (Barker et al. 2009, Clin Pharmacol Ther[40]; Park et al. 2016, NEJM[41]). REMAP-CAP is a Bayesian adaptive platform trial run in severe pneumonia and COVID-19 (REMAP-CAP Investigators 2021, NEJM[42]). The UK’s RECOVERY, an adaptive platform trial in a frequentist framework, established within months early in the pandemic that dexamethasone lowers 28-day mortality in hospitalised patients needing oxygen or invasive mechanical ventilation (RECOVERY Collaborative Group 2021, NEJM[43]).
Regulators are increasingly accepting such designs. The US FDA issued final guidance on adaptive designs in 2019 (FDA 2019[44]); in January 2026 it issued draft guidance on using Bayesian methods in clinical trials of drugs and biological products, recommending that simulations report a design’s operating characteristics across realistic scenarios (FDA 2026[45]). The FDA has run a paired-meeting programme for model-informed drug development (MIDD) since 2018, and the ICH’s draft adaptive-design guideline E20 reached Step 2 for public consultation in June 2025 (ICH E20 Step 2b[46]).
Outside trials, quantitative go/no-go criteria for development decisions are also widespread. In promoting model-based drug development, Lalonde and colleagues proposed writing decision criteria with a “target value” (TV) and a “lower reference value” (LRV), expressing go/no-go rules as the probability that the true effect exceeds these thresholds (Lalonde et al. 2007, Clin Pharmacol Ther[47]); AstraZeneca’s statisticians published a three-zone “go / consider / stop” framework for early development based on this idea (Frewer et al. 2016, Pharm Stat[48]); and in moving from Phase II to Phase III, using “assurance” — the predictive probability that a future trial will succeed — to estimate Phase III’s chance of success is now quite common (O’Hagan et al. 2005, Pharm Stat[49]). Pfizer’s “three pillars” (drug exposure, target binding, pharmacological effect) are another standard that moves Phase II decisions from stage labels towards mechanistic evidence (Morgan et al. 2012, Drug Discov Today[50]).
These practices leave four lessons.
First, the monitoring is continuous; the rules are pre-specified. Neither group sequential nor Bayesian adaptive designs allow “deciding how to look after looking at the data”; they put the flexibility inside the rules.
Second, confirmatory trials have an information firewall. Interim efficacy data are usually reviewed by an independent data monitoring committee, with the sponsor’s decision makers kept blinded; FDA guidance states explicitly that sponsor access to interim comparative data can pose a significant risk to trial integrity (FDA 2006, DMC Guidance[51]), a position the 2024 revised draft maintains (FDA 2024, DMC Draft Guidance[52]). So within Phase III, “deciding continuously on the efficacy posterior” is tightly restricted; what a company can use continuously is mainly blinded pooled data, operational data, safety data and external evidence. Even blinded pooled data can carry efficacy information indirectly — for example, a pooled event rate in an event-driven trial well below expectations; a sponsor that adjusts portfolio or deal strategy on that basis must likewise pre-specify and document it, so as not to compromise the trial’s integrity.
Third, quantitative decisions at the project level already have a foundation. Frameworks such as TV/LRV, assurance and the mechanistic pillars show that “writing boundaries in belief space” is not a new invention. What this piece calls belief-flow decisions extends such criteria — which used to live only in a single trial protocol or a single review — to the whole life of a project and to the portfolio, and checks them continuously between reviews.
Fourth, improvements at the portfolio level often come from better decision criteria and the culture that goes with them. Around 2011 AstraZeneca introduced its 5R framework, built around the right target, tissue, safety, patient and commercial potential (Cook et al. 2014, Nat Rev Drug Discov[53]); the company later reported that its success rate from candidate nomination to completion of Phase III rose from 4% in 2005–2010 to 19% in 2012–2016 (Morgan et al. 2018, Nat Rev Drug Discov[54]). The company also shrank and focused its pipeline sharply over the same period, so the improvement cannot all be credited to decision criteria; Morgan and colleagues particularly stressed the accompanying “right culture” — shifting evaluation from the quantity of progressions to their quality. Paul and colleagues’ “quick win, fast fail” likewise centres on bringing stop decisions forward (Paul et al. 2010, Nat Rev Drug Discov[55]).
Part two · 7. Belief-flow pricing: a framework and seven propositions
Building on part one, the framework can now be defined.
- Belief-flow pricing: using a project’s belief state, alongside its stage, as the state variable for valuation and decisions. Valuation is a function of the belief state, updated continuously as evidence flows in; decisions are discrete actions triggered when the belief trajectory hits pre-set boundaries, with the set of actions including continuing, scaling up, slowing down, stopping and transferring to another developer.
Relative to stage gates, the change in value from belief-flow decisions can be understood through four sources and three kinds of cost:
- State: adding the belief state alongside the stage, so projects in the same stage get explicit, testable differential treatment.
- Timing: denser decision times, moving from a Bermudan option towards an American one.
- Holder: adding “transfer” to the set of actions, so assets can move to better-suited developers at better times.
- Information: better experimental design and shorter data lags, so the information budget pays out earlier and more fully.
- Costs: the cost of maintaining models and holding reviews, the switching cost of actions, and the transaction frictions of transfers.
Note that these four sources do not simply add up. Denser timing only matters once the state has been expanded: if the state is still the stage label, adding reviews between two gates doesn’t change the state, and denser timing gains nothing. The stronger the signal, the larger the gain from denser timing. The value of the transfer option in turn depends on being able to recognise the right time and act on it. So this is an incremental decomposition, adding sources one at a time in order, in which the size of each depends on the order; the net value is the result of their combined effect minus the costs. The seven propositions below unfold the framework.
Proposition 1: adding the belief state alongside the stage cannot lower expected value. The reason: “stage plus all the evidence” contains strictly more information than “stage” alone, and the optimal decision conditioned on more information has an expected value no lower than the optimal decision conditioned on less (Blackwell’s result on comparing the informativeness of experiments is the general form of this idea; Blackwell 1953, Ann Math Stat[56]). This holds on condition that the belief model itself is calibrated; replacing expert judgment with a misspecified model guarantees nothing — which is why sections 9 and 12 stress calibration and backtesting.
Proposition 2: expectation is conserved; value comes from decisions. Belief is a martingale, so more frequent updating by itself doesn’t change the expected belief; for a risk-neutral decision maker with a fixed objective, the value of continuous observation is realised entirely through the decisions it changes — no change in decisions, no value. So a continuous monitoring system should be judged not by how many “insights” it produces but by how many decisions it changes, and whether it changes them correctly. (Under some preferences, resolving uncertainty earlier is valuable in itself even if it changes no decision, Kreps & Porteus 1978, Econometrica[57]; this piece doesn’t rely on that.)
Proposition 3: refining the grid on top of the existing decision times raises value weakly, with diminishing and converging gains. With no review cost and a new grid that contains the old, optimal value cannot fall, but the gains diminish. With diffusive information, the value lost to a coarse grid is roughly proportional to the review interval; in this piece’s model, the stronger the signal, the larger the constant of proportionality; once the review interval is much shorter than the time beliefs need to change significantly near a boundary, further refinement gains almost nothing. With a review cost, value first rises and then falls with the review interval, and there is an optimal cadence.
Figure 3 illustrates this with an illustrative model. Imagine a set of hypothetical projects, each with a 20% prior probability of being effective, a 6-year development period costing 1 unit a year, and a payoff of 40 units at the end if effective; the efficacy signal arrives continuously as a Brownian motion (the signal drifts of effective and ineffective projects differ by 1, with noise of 1.2 a year; the discrimination accumulated over the whole development period is roughly that of a moderately powered trial), and the firm can decide at each review, from the posterior, whether to stop. For each review cadence, we solved the time-varying optimal stopping boundary by backward induction. If the firm never stops, the portfolio’s expected value is 2.0; with nearly continuous reviews it is about 3.15, and weekly reviews are already almost the same. Reviewing every 2 years gives about 2.85, yearly about 3.01. With these parameters, a two-yearly cadence captures about three-quarters of the stopping-option value that continuous review can capture, and a yearly cadence nearly 90%; the loss from a coarse grid is roughly proportional to the review interval, about 0.14 per extra year. The proportion is sensitive to signal strength: with the signal noise halved, a two-yearly cadence captures about 80%; with it doubled, only just over half. Note that this compares proportions: the stronger the signal, the more the stopping option itself is worth, and the larger the absolute loss from a coarse grid; but the weaker the signal, the scarcer the useful information, and the larger the share of all the option value that a coarse grid misses.
If each review costs 0.02 units per project, the optimal cadence falls between six months and a year, and the value of weekly reviews drops to about 0.37. This setting is actually extreme: at weekly reviews, the annual review cost exceeds the project’s own annual development cost. If review costs fall to a tenth, the optimal cadence shortens to about a quarter (value about 3.08). The optimal cadence here is set mainly by the review cost — exactly where AI’s lowering of updating costs can make a difference.
Proposition 4: decision cadence should follow the arrival of information, not the calendar; looking more often only matters once information is available earlier. If evidence is revealed in bulk at readouts, beliefs don’t change between two readouts however often you review, and continuous review gains nothing. When reviews are misaligned with readouts, action on a readout must wait for the next review, and the spending in between is wasted. In the same model as figure 3, if information is revealed only at two readouts in years 2 and 4, every review cadence aligned with the readouts has the same value (about 2.85), while reviewing every 1.5 or 3 years drops it to about 2.4–2.5.
The same figure also shows a complementarity. Start from “revealed only at readouts, reviewed every 2 years” (about 2.85): making reviews weekly alone changes nothing; letting information arrive continuously between readouts while still reviewing every 2 years also changes nothing (again about 2.85); only doing both raises value to about 3.15. In other words, in this model, earlier-available information and denser decision times each produce no value on their own; they must go together — a numerical example of the earlier point that “the four sources do not simply add up”. This proposition shows that belief-flow decisions are more accurately described as “event- and belief-driven decisions”: reviews are triggered by information events and changes in belief, not by the clock. For most projects the first constraint is that information arrives too late and too lumpily, so shortening data lags and designing endpoints that can be read earlier are the precondition for denser decision times to matter.
Proposition 5: the timing of hand-offs can be chosen more finely, but the gains are net of frictions and limited by financing constraints. The same asset grows in expected value at different rates in different developers’ hands. An early-discovery company may be more efficient at target biology and chemistry optimisation; a large drug company has the advantage in large-scale clinical execution, registration and commercialisation, and a different cost of capital.
| Start with a minimal deterministic setting (figure 4): the asset’s rate of expected-value growth in the original developer’s hands falls as development progresses, while in a potential acquirer’s hands it is steadier; if transfer is costless, and the transfer price only decides how the two parties split the surplus, the transfer time τ* that maximises joint value lies where the two rate curves cross, and near its maximum the joint value V(τ) is approximately a downward-opening parabola. If decisions can only be made at intervals of h, the loss from missing the maximum is proportional to h²: if the optimum can be foreseen and the nearest time chosen, the expected loss is about | V″ | ·h²/24; if one can only act at the next time after seeing the signal, about | V″ | ·h²/6. This is exactly the problem of finding the maximum of a quadratic on a sparse grid: the coarser the grid, the more likely it misses the true maximum. In the setting of figure 4, stage gates allow decisions only at years 0, 2, 4 and 6, while the true optimal transfer time is year 3; transferring at year 2 loses about 0.16 units, at year 4 about 0.10 units, roughly 1.3% of the optimal joint value; monthly evaluation loses less than 0.001 units. |
This deterministic setting only shows the geometry, and three points need adding. First, in the deterministic setting τ* is known in advance and the loss comes only from misaligning decision times with it; in reality the positions of the two curves depend on evidence not yet revealed, τ* is a random stopping time that evolves with beliefs, and the problem becomes one of optimal stopping, in which the loss from a coarse grid is of order h rather than h² (proposition 3 and figure 3) — so when review intervals are not large, deterministic calculations like figure 4 probably understate the cost of a coarse grid. Second, transfers are hard to reverse and carry transaction costs; by McDonald and Siegel’s logic, the optimal rule is to “transfer when the acquirer’s value advantage exceeds the transaction cost plus the option value of waiting”, usually later than the point where the curves cross. Third, in reality many transfers don’t happen where the curves cross but are dictated by cash runway: a single-asset company is often forced to seek a partner when the financing window closes and little cash is left; buyers have their own clocks too, such as needing to restock the pipeline ahead of patent expiries. The original developer’s rate of value growth should take its cost of capital and financing risk into account.
Among transaction frictions, information asymmetry is the most studied. Akerlof’s “lemons” model shows that when sellers know more about an asset’s quality than buyers, trade can shrink (Akerlof 1970, QJE[58]). Whether R&D licensing markets suffer such adverse selection is empirically contested. Danzon and colleagues found that compounds developed through partnerships had higher success rates in Phases II and III than in-house projects, especially when a large firm took them over (Danzon et al. 2005, J Health Econ[59]); but this could mean either that adverse selection is not severe or that acquirers execute better, and the two are hard to tell apart. Other studies give the opposite signal: Pisano reported evidence supporting the lemons hypothesis in a working paper (Pisano 1997, Harvard Business School working paper), and Hermosilla found that drugs large companies licensed in hastily soon after their own Phase III failures had a markedly lower chance of later success, which he interpreted as buyers lowering their standards under pipeline pressure (Hermosilla 2021, Management Science[60]). The latter mechanism also shows that transfer timing is often set by the buyer’s and seller’s own clocks rather than by the asset itself.
Proposition 6: with switching costs, continuous observation needs a matching band of inaction. Following Dixit, and Abel and Eberly, when both scaling up and slowing down carry adjustment costs or irreversibility, the optimal policy in belief space is two boundaries with a band between them. The more continuous the observation, the more visible small jitters in belief near a boundary become; without an explicit band of inaction, the organisation will be pulled back and forth by noise. Continuous observation does not require continuous reaction; if anything, it requires spelling out more clearly “when not to move”.
Proposition 7: once information-processing capacity is commonplace, excess returns shift to generating information. Grossman and Stiglitz pointed out that if prices fully reflected all information, no one would have an incentive to pay to acquire it, so in equilibrium markets cannot be fully efficient and those who acquire information must be compensated (Grossman & Stiglitz 1980, AER[61]). Borrowing this equilibrium logic by analogy: when AI drives down the cost of organising public literature, competitor news and historical trial data, the rent to be earned from processing public information falls with it. Lasting advantage is more likely to come from the ability to generate new information: proprietary data, experimental designs with stronger signals, faster trial execution, and biomarkers that separate effective from ineffective earlier. There is a precedent to refer to: Ewens, Nanda and Rhodes-Kropf found that after cloud computing lowered start-ups’ cost of experimentation, venture capital shifted to a “small bets, many tries, quick cut-offs” model (Ewens et al. 2018, J Financial Econ[62]). Note that cloud computing lowered the cost of running experiments, not just of processing information; what the precedent shows is that once generating information becomes cheap, financing structures change with it. In biopharma, the cost of processing information may fall first, while the cost of the trials themselves need not fall in step. In belief-flow terms, the former corresponds to letting the information budget pay out more fully in less time and at lower cost — not merely looking at it more often.
A necessary qualification: more options are not always better. The “weakly better” of propositions 1 and 3 holds only with a single decision maker, time-consistent preferences and no strategic interaction. With several parties, fewer options are sometimes better. Being able to re-evaluate and refinance at any time weakens the credibility of an ex-ante threat to stop, creating a soft budget constraint (Dewatripont & Maskin 1995, Rev Econ Stud[63]); incentive contracts that encourage exploration need a commitment to tolerate early failure (Manso 2011, J Finance[64]); decision makers with time-inconsistent preferences exercise too early, eroding the option value of waiting, and so can benefit from commitments that restrict their own future choices (Grenadier & Wang 2007, J Financial Econ[65]). In transfer markets with information asymmetry, the timing of a transfer itself sends a signal: high-quality sellers are willing to wait for more news, and restricting later trading opportunities can sometimes improve efficiency (Daley & Green 2012, Econometrica[66]; Fuchs & Skrzypacz 2015, J Econ Theory[67]). Nor is the movement of assets between developers necessarily a social gain: research has found that some acquisitions aim to shelve projects that compete with the acquirer’s products (Cunningham et al. 2021, JPE[68]). Part of the value of stage gates comes precisely from their role as a commitment device; the pre-specified decision charter of belief-flow decisions (see section 10) can be seen as rebuilding that commitment while keeping flexibility.
8. Effects on investment and decision behaviour
Valuation moves from pricing by round to pricing by path. For listed companies, share prices already price beliefs continuously; the price jumps around readouts are the market version of the jump-and-diffusion process described earlier. The bigger change is for private companies and internal project valuations. Investment funds are required by accounting standards to measure holdings at fair value periodically, but in practice they often anchor on the last round’s price and seldom adjust much between rounds. Under a belief-flow system, valuation is a function of the belief state and can form a continuous time series. That lets investors see changes in a project’s prospects earlier, and also makes the valuation itself something that must be audited: whoever maintains the model largely determines the price.
This is not without precedent in accounting. In business combinations, contingent consideration classified as a liability (follow-on payments conditional on clinical or regulatory milestones, for example) must be remeasured at fair value every reporting period, usually estimated in practice with scenario-probability weighting or option-pricing methods; in-process R&D (IPR&D) acquired in a business combination is recognised as an intangible asset and must be tested for impairment at least annually and whenever a triggering event occurs. These valuations are all audited. (Under US GAAP, IPR&D acquired in an asset acquisition with no alternative future use is expensed immediately and falls outside this; most licensing deals are of this kind.) Once a company has a continuously updated internal belief trajectory, auditors and regulators will naturally ask: when the posterior probability drops sharply, is that an impairment trigger, and is it material information that must be disclosed?
Financing and deal contracts need redesigning. Milestones are popular because they can be verified. Milestones in licensing contracts are already getting finer — payments conditional on a particular endpoint reaching an agreed threshold, or on obtaining a particular regulatory designation. But as long as one party controls the generation of evidence, disputes shift to “whether the other side made reasonable development efforts”; many recent lawsuits over contingent consideration in M&A turn on exactly this. Examples include the milestone-payment case over Alexion’s acquisition of Syntimmune, in which the Delaware Court of Chancery found in 2024 that the buyer, by ending the programme for post-merger integration reasons, had failed to meet the contractual standard of “commercially reasonable efforts” (Cleary Gottlieb 2024[69]), and the CVR case over BMS’s acquisition of Celgene mentioned below. For a posterior probability to become a contractual trigger, the model, data and updating rules must be transparent and reproducible to both sides. Incomplete-contract theory starts from the premise that “variables that cannot be verified are hard to write directly into contracts”, and shows that the allocation of residual control rights then affects each party’s ex-ante investment (Grossman & Hart 1986, JPE[70]); in biotech alliances, how control rights such as termination and scale-up are allocated is already central to contract negotiations (Lerner & Merges 1998, J Ind Econ[71]). So “the verifiability of beliefs” and “who decides when beliefs cross a boundary” will be key to whether belief-flow pricing can enter dealmaking.
Portfolio management moves from annual budgets to continuous reallocation. The Gittins index offers an idealised reference: if projects are independent and switching is costless, it is optimal always to give resources to the project with the highest index. In reality switching has costs, and teams and outsourcing contracts can’t be moved at will, so continuity at the portfolio level also needs bands of inaction; “near-optimal” usually means optimising over the few resources that can be adjusted (budget slack, outsourced capacity). The change to expect is that reallocation no longer happens only in the annual budget cycle, but as local adjustments triggered when a project’s beliefs cross a boundary.
The timing of BD and licensing deals is re-chosen. Assets moving between developers is already the norm: by IQVIA’s count, of the 48 new active substances launched in the US in 2024, 41 (85%) originated with emerging biopharma companies, while those originated and launched by emerging companies themselves made up 63% (IQVIA 2025[72]). The difference implies that about one in five new drugs originated with an emerging company and was launched by another company after licensing or acquisition; transfers between emerging companies are not in this difference, so it is only a lower bound on the hand-off rate (this inference is ours). Proposition 5 implies that asset transfers need not wait for “the window after a readout”: when the original developer’s and a potential acquirer’s rates of value growth are about to cross, both sides have reason to make contact early. Then again, a seller actively seeking a deal before a readout may itself be read by the buyer as a bad signal (see section 7); a shared data room and pre-agreed analysis methods help with this. In real deals, buyers and sellers rarely share the model itself, because the seller’s prior and value function would reveal its reservation price, and sensitive information in M&A due diligence must be isolated through “clean teams”; the more realistic form is that both sides share one data room and pre-agreed analysis methods while each keeps its own value function. This will push towards a finer division of labour in the industry: the boundaries between discovery companies, clinical-development companies and commercialisation companies will be set by the relative size of transaction frictions and the gains from finer timing, in line with the logic of transaction-cost theory (Coase 1937, Economica[73]; Williamson 1985, The Economic Institutions of Capitalism).
Behavioural biases will take new forms. Continuous valuation may ease escalation of commitment: Staw’s classic experiment showed that decision makers personally responsible for a failing decision tend to keep investing in it (Staw 1976, OBHP[74]). If stop boundaries are written into the decision charter before the project starts, and reviews are triggered automatically by the belief trajectory, the project sponsor’s sunk-cost bias is held in check. But continuous valuation may also bring the opposite problem. Benartzi and Thaler proposed “myopic loss aversion” to explain the equity premium puzzle (Benartzi & Thaler 1995, QJE[75]); later experiments found that the more frequent the feedback and evaluation, the less risk subjects took (Thaler et al. 1997, QJE[76]; Gneezy & Potters 1997, QJE[77]). This evidence comes from individual decisions and should be extrapolated to organisations with care, but the direction deserves vigilance: if an organisation evaluates teams on continuously changing valuations, it may cut high-risk projects that were worth continuing too early. The sensible approach is to evaluate on decision quality and forecast calibration rather than short-term movements in the valuation itself — consistent with Manso’s conclusion about tolerating early failure.
Homogenisation risk may appear at the market level. If most institutions use similar data, similar priors and similar models, their beliefs will be highly correlated and their decisions synchronised: the same class of targets embraced, or abandoned, all at once. Goodhart’s law warns that once a statistical regularity is used as a target for control, it loses its former stability; when belief models influence financing and evaluation, teams have an incentive to choose endpoints and data that make the model “look good”. These problems are all amplified by continuity.
9. Risks and limits
Repeated looks and type I error. A Bayesian posterior does not in itself depend on when you stop looking: as long as the prior and model hold, posterior probabilities are calibrated across the portfolio. But that doesn’t mean frequentist type I error can’t inflate. Under a rule of “declare efficacy once the posterior exceeds a threshold”, looking repeatedly raises the probability of wrongly declaring an ineffective drug effective (Ryan et al. 2020, BMC Med Res Methodol[78]), which is why regulators recommend that Bayesian designs report operating characteristics by simulation, including the probability of a false positive under the null. In this piece’s two-point model, the probability has an upper bound: the chance that the posterior odds ever exceed k times the prior odds is at most 1/k, and the more continuous the monitoring and the longer the horizon, the closer it gets to that bound.
Chasing noise. Models are almost always approximations. Continuous updating amplifies the consequences of misspecification — underestimating noise, or ignoring batch or site effects, for example — and beliefs overreact to noise. Bayesian methods need frequentist calibration checks too: things the posterior says have a 30% chance should, in the long run, happen about 30% of the time.
Selection bias from scaling up at the boundary. Scaling up triggered by crossing the upper boundary naturally tends to pick overestimated projects: stopping exactly when the boundary is crossed tends to leave the estimate too high, one source of the drop in success rates from Phase II to Phase III. Shrinkage estimates under hierarchical models or empirical priors can partly correct this, provided the prior is appropriate. The scale-up boundary should be designed with this in mind.
Where priors come from. Building priors from historical trials can greatly improve the efficiency of early judgments, but the exchangeability between historical data and the current project is an assumption: indications, patient populations, endpoint definitions and standards of care all change. Dynamic borrowing — adjusting how much to borrow according to how consistent current and historical data are — can partly ease this risk but not remove it.
Data lags and patients in the pipeline. Clinical data have a natural lag between collection and cleaning, and uncleaned data may contain systematic errors; a “real-time” belief built on uncleaned data may be worse than one built on periodically cleaned data. The lag is not only a data-quality problem: patients enrolled but without outcomes yet change the optimal stopping rule (Chick et al. 2017).
Regulatory integrity and patient protection. As section 6 described, efficacy data in confirmatory trials sit behind a firewall. In clinical development, belief-flow decisions mainly apply to early exploratory studies, open-label single-arm studies, blinded and operational data, safety monitoring, external evidence and portfolio-level resource allocation — not to adjusting strategy midway through a confirmatory trial on unblinded efficacy data.
Organisational incentives. If project teams’ bonuses and promotions are still tied to “how many projects were advanced to the next stage”, even the clearest stop boundaries will be circumvented: teams will delay submitting unfavourable evidence, or argue about the model’s specification. AstraZeneca itself regards shifting evaluation from the quantity of progressions to their quality as one of the key conditions for its improved success rate. Without matching changes to incentives, belief-flow decisions can easily become one more report sitting beside the existing process.
Data ownership. Most clinical data are collected and cleaned by CROs, with transfer frequency written into the service contract; data from co-development projects are governed by alliance agreements, and one party may not have the right to access the other’s data at any time. For emerging companies that outsource heavily, “near-real-time observation” first requires rewriting the data terms with CROs and partners, before building any model.
Disclosure and legal risk. For listed companies, a versioned, continuously updated internal probability of success may constitute material non-public information, affecting insider-trading windows and the timing of announcements; in securities litigation, these records may also be subpoenaed to argue that the company “knew all along”. The better the audit layer, the sharper this problem. Implementing belief-flow decisions needs legal and disclosure teams involved from the start, to define which beliefs are internal management estimates and how they relate to what is disclosed externally.
The cost of frequent changes of direction. Every adjustment consumes the team’s attention and trust. An organisation that changes direction often may lose more in execution than it gains in decision-making.
10. A new decision mechanism: belief-flow decisions
If the analysis above holds, what organisations need is not “more frequent review meetings” but a mechanism that keeps beliefs updating continuously and lets decisions happen only when they are needed. Start with an example.
Imagine an antibody–drug conjugate (ADC) for solid tumours in Phase II dose expansion: two dose arms with randomised allocation (which also matches the FDA’s expectations for dose optimisation in oncology), open label, with confirmed objective response rate (ORR) as the main focus. At the start, the decision charter states for each dose arm: a target value of 35% ORR and a lower reference value of 20%; if the posterior probability that “the true ORR is at least 20%” falls below 10%, a stop review is triggered; if the posterior probability that “the true ORR is at least 35%” exceeds 80%, a scale-up review is triggered to discuss preparing Phase III process scale-up and regulatory discussions early; in between, the status quo holds. To avoid triggering too early, the charter says the boundaries are checked only after each arm has at least 15 evaluable patients. The charter also attaches simulation results: under scenarios such as “drug ineffective”, “moderately effective” and “highly effective”, how likely these boundaries are to wrongly stop or wrongly advance. A rough calculation with a uniform prior, at most 40 patients per arm and a look after each enrolment: if the true ORR is exactly 20%, the chance these boundaries trigger a scale-up review at some point is about 3%; if the true ORR is exactly 35%, about 44%; if looks start from the 5th patient, these become about 12% and 57%. This is the repeated-looks problem section 9 discusses, and the choice of starting sample size is itself a trade-off between wrongly advancing and acting late. Because the study is open label, the sponsor can legitimately see response data continuously as patients enrol — exactly the setting where belief-flow decisions fit best. All these numbers are illustrative.
Suppose that at the high-dose arm’s 18th evaluable patient (when crossing the boundary requires at least 8 confirmed responses, an observed ORR of about 44%), its posterior crosses the scale-up boundary while the low-dose arm stays inside the band; a review is triggered without waiting for the scheduled year-end portfolio meeting. The review has to answer not just whether to advance, but at which dose: the difference in efficacy between the arms must be weighed against differences in toxicities such as interstitial lung disease. Nor is the review’s core quantity the ORR posterior itself, but the assurance it implies for the Phase III primary endpoint (progression-free survival, say), which must account for the uncertainty in bridging from ORR to the survival endpoint; and the ORR estimate seen at the crossing carries the selection bias described in section 9 and needs discounting. If the study might also support accelerated approval, response assessment should rely on blinded independent central review, and the sponsor’s continuous looking should be documented as not affecting that review’s independence. Three months later, a competitor with the same target reports a higher response rate and a lower rate of interstitial lung disease at a scientific meeting: the project’s own efficacy evidence hasn’t changed (if the model borrows external data on the same target, the posterior may even rise slightly), but the commercial value function has, and the target value, set with reference to the competitive landscape, needs raising. This is another kind of trigger: what changes is not the belief but the boundary. The review might conclude to continue but narrow the indication, or to start preparing for out-licensing. Once the same project enters a randomised controlled Phase III, efficacy data go to an independent data monitoring committee, the belief model can use only blinded operational data, safety data and external evidence, and updating naturally slows down.
Five components of belief-flow decisions can be drawn from this example.
First, the decision charter. At the start of a project, write down boundaries in belief space: below what posterior probability of success or expected efficacy a stop review is triggered, above what a scale-up review, and that the status quo holds in between; and under what conditions an evaluation of out-transfer is started, including conditions unrelated to belief, such as no longer fitting company strategy. The charter should attach operating characteristics from simulation, be pre-specified, and leave a record of any changes. This extends the ideas of TV/LRV and group sequential designs to projects and portfolios, and is also the commitment device described in section 7.
Second, ownership and independence of the model. The team maintaining the belief model should be independent of the project team. The closer reference is the separation in financial institutions between model development, independent validation and business use (the 2011 guidance on model risk management from the Federal Reserve and the OCC, SR 11-7, is the canonical statement of this practice, Federal Reserve 2011[79]; it was superseded in April 2026 by the revised SR 26-2, Federal Reserve 2026[80]) — not a data monitoring committee, whose independence comes from seeing unblinded data the sponsor can’t, whereas the belief-model team still sees only data the sponsor may see. The project team can submit evidence, provide priors through structured expert elicitation (the SHELF method, for example) and challenge the model, but cannot edit the posterior directly; when the two sides disagree, the charter escalates the matter to the portfolio committee, which decides and records its decision. The model itself needs independent validation, stating clearly which decisions it applies to and which it doesn’t; if its outputs are to support regulatory submissions, it should also follow the “context of use”-based credibility assessment framework in the FDA’s draft guidance on using artificial intelligence to support regulatory decision-making (FDA 2025[81]). That draft doesn’t cover uses that are purely internal and operational and don’t affect patient safety or the reliability of study results, so a pure portfolio-decision model falls outside it; but once a model enters GCP-regulated processes such as safety monitoring, it must also meet requirements for electronic records and computerised-system validation (21 CFR Part 11, for example).
Third, the division of labour between people and machines. The model integrates evidence, updates beliefs, computes the value of information and makes recommendations; people define the value function (what the company is willing to pay for what risk and return), set the boundaries, approve actions once a boundary is crossed, and judge factors outside the model, such as strategic fit, ethics and reputation. Note that the commercial assumptions in the value function need calibrating too: in rNPV, peak sales, pricing and market share often affect valuation as much as the probability of technical success, and pre-launch sales forecasts have historically been badly wrong (Cha et al. 2013, Nat Rev Drug Discov[82]). If only the technical probabilities are calibrated while commercial forecasts are still self-reported by the project team, the valuation is only half independent; commercial assumptions should also be versioned and reviewed against actual sales after launch.
Fourth, triggered by events, not the calendar. Reviews are triggered by four kinds of event: beliefs crossing a charter boundary; new evidence of high information value arriving (a competitor readout, key biomarker data); a change in the value function itself (a competitor’s approval changing the standard of care, reimbursement or pricing policy changing expected returns, a shift in company strategy); and a marked deviation in the model’s calibration. Routine periodic reviews can be kept, but their function shifts from “making decisions” to “checking whether the model and charter are still sound”.
Fifth, coexistence with existing stage gates. Stage gates won’t disappear, because regulatory, contractual and capital milestones remain discrete. Belief-flow decisions run between the gates, turning a gate review from “the first serious look at the data” into “confirming a trajectory that has been tracked all along”.
11. AI infrastructure
The infrastructure supporting belief-flow decisions can be divided into five layers (figure 5).
Data layer: a unified evidence graph. Bring in in-house experimental data, clinical operations data, safety data, real-world data, and competitor and literature information in a structured, traceable way. Clinical data already have established standards here — CDISC’s SDTM and ADaM, MedDRA for coding adverse events, the OMOP common data model often used for real-world data; the hard part is aligning literature, competitor disclosures and internal early-research data to the same set of concepts. Large models can take on much of the extraction and alignment, but they make plausible-looking extraction errors, so a proportion should be sampled for human review with error rates recorded, and every piece of evidence entering the model must be traceable to its original source. Two engineering problems are often underestimated. One is keeping data in-house: whether unpublished clinical data and material non-public information can be sent to an external model service, and whether the provider retains them, needs both contractual and deployment safeguards. The other is version drift: when the underlying model is upgraded, extraction and alignment behaviour changes and previously measured error rates no longer hold, so versions should be pinned and revalidated against a fixed test set before upgrading.
Model layer: hierarchical Bayesian and mechanistic models. At the project level, hierarchical Bayesian models use historical data on the same target, indication and molecule type as sources of prior information, estimating the degree of borrowing explicitly. At the mechanistic level, pharmacokinetic and pharmacodynamic models, disease-progression models and quantitative systems pharmacology models can turn early data into predictions of later outcomes; predictive probabilities such as assurance are one standard output of this layer. Machine-learning models can be used to predict stage success probabilities from historical project features, but their output should serve as one source of prior information rather than directly replace the posterior.
Decision layer: a value-of-information engine and policy optimisation. For each project, compute: the expected value of each action under the current belief; the expected value of information from continuing to collect each kind of information; and how far the project is from its charter boundaries. At the portfolio level, solve for a near-optimal allocation under resource constraints and switching costs. This layer’s output is recommendations and trigger signals, not instructions that execute automatically.
Audit and access layer: versioning beliefs, calibrating after the fact, controlling access. Every belief update has a version, its input evidence and a model version; every decision records the belief at the time, the available actions and the final choice. After a project ends, actual outcomes are used to check whether the predictions were calibrated — by computing Brier scores and calibration curves, for example. This layer also manages access: who can see which project’s posterior, how blinded and unblinded data are separated, how partners’ data are restricted to their contractual uses, and who has access to valuations that are material non-public information. This layer is the precondition for belief-flow pricing to enter contracts and external valuation.
Interaction layer: an interface for decision makers. What decision makers need to see is not a single number but the distribution of belief, why it recently changed, how far it is from the boundaries, and which information is most worth getting. The interface’s design goal is to lower the cost of attention — corresponding directly to the second source of blockiness in part one.
12. Recommendations for implementation
Belief-flow decisions don’t need to replace the existing system all at once; they can be introduced in three steps.
Step one: shadow running. Choose one therapeutic area and run the belief model alongside the existing stage-gate process, recording its beliefs and recommendations at each point without changing actual decisions. A full decision cycle in the clinical phase often takes years, and a therapeutic area has only a few key readouts a year, so relying on one’s own projects alone gives too small a sample for calibration checks. A faster way: record the model’s prediction in advance for every upcoming readout (including competitors’ public trials), compute Brier scores after each readout, and accumulate dozens to a hundred or more checkpoints within a year or two.
Step two: calibration and backtesting. Backtest the model on historical projects: with the information available at the time, could it have identified the projects that eventually failed earlier, without wrongly killing those that eventually succeeded? Focus on calibration rather than discrimination, because decisions depend on the probabilities themselves. Backtesting requires point-in-time snapshots of the data: only data and literature actually available at the time, in the state they were then — not later cleaned and completed versions — or there will be look-ahead bias. Most companies’ historical project archives don’t meet this standard, and rebuilding the snapshots is itself a sizeable project.
Step three: limited delegation. First hand procedural actions to automatic triggering by the charter boundaries — automatically convening a stop review, or automatically notifying authorised BD staff to prepare a data package (with the scope of notification governed by the access layer, because for a listed company, preparing to out-license may itself be material non-public information); then gradually open up low-risk resource decisions, such as reprioritising non-critical experiments. High-risk and irreversible decisions (stopping a key project, transferring it out) are still made by the existing portfolio or investment committee; when a decision differs from the model’s recommendation, the reasons must be recorded in writing and reviewed at the next calibration.
Metrics for assessing the effect include: how much earlier failures are caught (how much sooner failing projects are identified than under the stage-gate system); proxies for the rate of wrongful kills (stopped projects have no outcome, so wrongful kills can only be estimated indirectly — whether competitors with the same target or mechanism later succeeded, whether assets out-licensed after being stopped were approved elsewhere, or how projects turned out that the model recommended stopping during shadow running but were in fact continued); the Phase III success rate of projects entering Phase III; the time from data cut-off to the corresponding decision; and the calibration of readout predictions. “How much uncertainty was resolved per unit of capital” is also a useful metric, but it should be measured on a basis verifiable outside the model (the uncertainty actually resolved by readout results, for example), so the model isn’t grading itself.
Conclusion, and questions still open
Back to the opening intuition. R&D decisions today really are like a coarse-grained jump process, but to make them “continuous”, what must change first is the state, and only then the time. AI lets beliefs about R&D projects update almost continuously, which may raise the value of assets — but the sources of that value need spelling out. Looking more often creates no value by itself, because under the decision maker’s own model belief is a martingale; value comes from several interdependent things: adding an explicit belief state alongside the stage, extending decision times from a few gates to near-continuous, bringing “transfer to a better-suited developer” into the options that can be evaluated at any time, and making information available sooner and with a stronger signal. These gains must be netted against review costs, switching costs and transaction frictions; they are constrained by the pace at which biological information arrives and by regulatory integrity; and with several parties involved, commitment has to be maintained through pre-specified rules.
So belief-flow pricing describes not “reaching conclusions faster” but making a project’s confidence visible and auditable at any moment, and letting decisions happen when they are really needed.
What should be continuous is belief; actions remain discrete. Decision cadence should follow information, not the calendar.
This framework has some limitations that should be stated.
First, most of the propositions combine and extend existing theory in an R&D setting, and the numbers in figures 3 and 4 come from illustrative models. They show mechanisms and directions, but the actual size of the gains depends on specific signal strengths, cost structures and transaction frictions, and needs to be estimated with real portfolio data.
Second, this piece takes the value function as given, and the illustrative models don’t discount. In reality, a company’s risk preferences, strategic considerations and external financing constraints all change the value function, and these are precisely what is hard to model.
Third, this piece mainly discusses decisions within a single company. When many institutions adopt similar belief models at once, asset prices and dealmaking interact, which calls for equilibrium analysis; this piece only discusses it qualitatively.
Key points
- A stage label is not a Markov state; the first step towards continuity is adding an explicit, calibratable belief state alongside the stage, and only the second is making time finer.
- Observation and belief updating can come close to continuous; action usually should not. Switching costs call for an explicit band of inaction.
- Under the decision maker’s own model belief is a martingale, so looking more often doesn’t change the expectation; value comes from a fuller state, finer decision times, better-suited holders and earlier information, net of review costs and transaction frictions.
- In the illustrative model, earlier information and denser decision times are complements: neither adds value on its own, only together; review cadence should be triggered by information arriving and beliefs changing, not by the calendar.
- With several parties, more options are not always better; a pre-specified decision charter is both a boundary and a commitment.
- For belief-flow pricing to enter contracts, accounting and external valuation, the model, data and updating rules must be verifiable and auditable, and disclosure and ownership must be handled.
At least three questions remain open. First, can beliefs become something that is traded? Contingent value rights (CVRs) issued in M&A are already instruments priced continuously by the market on specific clinical or regulatory events — for example, the CVR BMS issued when it acquired Celgene in 2019: listed on the New York Stock Exchange, it required three pipeline drugs to be approved by deadlines or it would lapse. Liso-cel was not approved by the end of 2020, the CVRs terminated in January 2021, and litigation followed over the obligation of “diligent efforts”; the suit was dismissed at first instance and revived by an appeals court in August 2026 (BMS 2021, SEC 8-K[83]; Fierce Pharma 2026[84]); Fernandez, Stein and Lo have also proposed financing structures built on large R&D portfolios (Fernandez et al. 2012, Nat Biotechnol[85]). If an R&D asset’s valuation is a continuous, auditable belief trajectory, will instruments written on it become more common? Second, as the ability to process public information becomes commonplace, will competition in R&D shift more decisively towards data and experimental capability? Third, can regulatory systems accept decision rules pre-specified at the level of a project rather than a single trial, letting belief-flow decisions cover more of the development chain?
References
- https://doi.org/10.1016/0007-6813(90)90040-I
- https://go.bio.org/rs/490-EHZ-999/images/ClinicalDevelopmentSuccessRates2011_2020.pdf
- https://doi.org/10.1093/biostatistics/kxx069
- https://doi.org/10.1016/j.jhealeco.2016.01.012
- https://doi.org/10.2307/1884852
- https://doi.org/10.1016/S0304-3932(03)00029-1
- https://doi.org/10.2307/2343787
- https://doi.org/10.1016/0304-405X(77)90015-0
- https://doi.org/10.2307/1884175
- https://doi.org/10.1016/0304-405X(79)90022-9
- https://doi.org/10.2469/faj.v56.n3.2362
- https://doi.org/10.1086/261619
- https://ideas.repec.org/a/aea/aecrev/v84y1994i5p1369-84.html
- https://doi.org/10.1214/aoms/1177731118
- https://doi.org/10.1214/aoms/1177730197
- https://doi.org/10.1109/TSSC.1966.300074
- https://doi.org/10.1016/0022-247X(65)90154-X
- https://doi.org/10.1287/opre.21.5.1071
- https://doi.org/10.1137/1108002
- https://doi.org/10.1111/j.2517-6161.1979.tb01068.x
- https://econometricsociety.org/publications/econometrica/1994/05/01/notes-and-comments-switching-costs-and-gittins-index
- https://doi.org/10.1093/biomet/25.3-4.285
- https://doi.org/10.2307/2531909
- https://doi.org/10.1111/rssb.12225
- https://doi.org/10.1111/1468-0262.00259
- https://doi.org/10.1111/j.1468-0262.2005.00564.x
- https://www.tse-fr.eu/pub/30051
- https://doi.org/10.1086/260062
- https://doi.org/10.2307/3003143
- https://doi.org/10.1111/j.1540-6261.1986.tb05043.x
- https://doi.org/10.1093/rfs/12.5.975
- https://doi.org/10.1111/1540-6261.00587
- https://doi.org/10.1093/rfs/hhg021
- https://doi.org/10.1214/105051605000000034
- https://doi.org/10.1093/rfs/14.1.113
- https://doi.org/10.1093/biomet/64.2.191
- https://doi.org/10.2307/2530245
- https://doi.org/10.1093/biomet/70.3.659
- https://doi.org/10.1038/nrd1927
- https://doi.org/10.1038/clpt.2009.68
- https://doi.org/10.1056/NEJMoa1513750
- https://doi.org/10.1056/NEJMoa2100433
- https://doi.org/10.1056/NEJMoa2021436
- https://www.fda.gov/regulatory-information/search-fda-guidance-documents/adaptive-design-clinical-trials-drugs-and-biologics-guidance-industry
- https://www.fda.gov/regulatory-information/search-fda-guidance-documents/use-bayesian-methodology-clinical-trials-drug-and-biological-products
- https://www.fda.gov/media/188961/download
- https://doi.org/10.1038/sj.clpt.6100235
- https://doi.org/10.1002/pst.1746
- https://doi.org/10.1002/pst.175
- https://doi.org/10.1016/j.drudis.2011.12.020
- https://www.fda.gov/media/75398/download
- https://www.fda.gov/media/176107/download
- https://doi.org/10.1038/nrd4309
- https://doi.org/10.1038/nrd.2017.244
- https://doi.org/10.1038/nrd3078
- https://doi.org/10.1214/aoms/1177729032
- https://doi.org/10.2307/1913656
- https://doi.org/10.2307/1879431
- https://doi.org/10.1016/j.jhealeco.2004.09.006
- https://doi.org/10.1287/mnsc.2019.3530
- https://www.jstor.org/stable/1805228
- https://doi.org/10.1016/j.jfineco.2018.03.001
- https://doi.org/10.2307/2298076
- https://doi.org/10.1111/j.1540-6261.2011.01688.x
- https://doi.org/10.1016/j.jfineco.2006.01.002
- https://doi.org/10.3982/ECTA9278
- https://doi.org/10.1016/j.jet.2015.04.009
- https://doi.org/10.1086/712506
- https://www.clearygottlieb.com/news-and-insights/publication-listing/delaware-court-of-chancery-finds-buyer-failed-to-use-commercially-reasonable-efforts-in-pharma-milestone-payment-case
- https://doi.org/10.1086/261404
- https://doi.org/10.1111/1467-6451.00066
- https://www.iqvia.com/insights/the-iqvia-institute/reports-and-publications/reports/global-trends-in-r-and-d
- https://doi.org/10.1111/j.1468-0335.1937.tb00002.x
- https://doi.org/10.1016/0030-5073(76)90005-2
- https://doi.org/10.2307/2118511
- https://doi.org/10.1162/003355397555226
- https://doi.org/10.1162/003355397555217
- https://doi.org/10.1186/s12874-020-01042-7
- https://www.federalreserve.gov/supervisionreg/srletters/sr1107.htm
- https://www.federalreserve.gov/supervisionreg/srletters/sr2602.htm
- https://www.fda.gov/regulatory-information/search-fda-guidance-documents/considerations-use-artificial-intelligence-support-regulatory-decision-making-drug-and-biological
- https://doi.org/10.1038/nrd4127
- https://www.sec.gov/Archives/edgar/data/14272/000114036121000023/brhc10018575_ex99-1.htm
- https://www.fiercepharma.com/pharma/its-bms-vs-celgene-investors-again-after-us-appeals-court-ruling
- https://doi.org/10.1038/nbt.2374