Redefining Regulatory Proof: The End of the Two‐Trial Paradigm Authors: Jean‐Christophe Lega, Steeve Provencher, Bruno Falissard, Clara Locher, Matthieu Roustit, Arthur Gougeon, Rémy Boussageon, Sylvie Chevret Published: 2026 ABSTRACT ABSTRACTThe recent decision to establish a single pivotal trial as the default evidentiary standard for Food and Drug Administration (FDA) approval marks a substantive shift in the architecture of regulatory proof. This marks a shift from replication‐based validation to a coherence‐based model of sufficiency, beyond a procedural adjustment. Because FDA standards likely influence global development strategies, this change will likely extend beyond the United States. Although accelerating access and reducing development costs are legitimate objectives, the reform raises methodological and epistemological concerns. The statistical rationale does not fully incorporate the contextual factors that influence false‐positive risk, including prior plausibility and statistical power. Prior plausibility remains informal, and post‐marketing evidence seldom restores the epistemic security provided by independent replication. Replication is not merely a statistical redundancy but rather a structural safeguard of robustness and external validity. Without explicit safeguards, increased regulatory flexibility may generate heterogeneity in evidentiary thresholds and progressively weaken the discriminative function of health technology assessment. In February 2026, the Food and Drug Administration (FDA) formally announced that “one adequate and well‐controlled trial, combined with confirmatory evidence,” accepting the measurement of effects on biomarkers or intermediate endpoints while collecting robust data on all drugs and devices, would constitute its new default standard for marketing authorization [1]. The reform of the 1998 guidance requiring two trials, each with a two‐sided alpha level of 0.05, was justified by advancements in biological understanding, the multifactorial nature of clinical credibility, and the financial and temporal costs of duplicative trials [1]. However, given the increased uncertainty surrounding the estimated effects of the treatment under study that could result from such an evaluation process, the methodological and epistemological foundations of this position require examination. In light of the 1998 criteria, the aim of this paper is to review the arguments justifying it. As a preliminary point, it should be noted that replication does not imply internal validity. As Prasad and Makary state [1], the replication of trials with substandard controls, dubious endpoints, inadequate power, and post hoc statistical analysis plans is useless. Therefore, we only advocate the replication of well‐planned and well‐conducted trials. First, it is argued that two clinical trials are no longer necessary to minimize the risk of false‐positive conclusions because modern drug candidates generally have greater biological plausibility than those considered under the historical framework [1]. However, this can be somewhat misleading. Even if the significance level is the only available guide for decision‐making, it should only lead to a false‐positive result once in 20 trials with no true effect. Therefore, even if a drug shows promising results on preclinical data, it may have no true effect on clinical endpoints. This was exemplified by the reported 90% failure rate during clinical development [2]. Moreover, it is impossible to control the risk of false indications because this depends on the proportion of trials for which a true effect exists, and this cannot be quantified directly. In this framework, the probability that a statistically significant finding is erroneous depends on prior plausibility that the intervention is effective and on statistical thresholds for decision rules (type I error rate and power) [3, 4, 5]. Even if a decision‐making process with a strict cut‐off is desired for drug approval, which is aligned with the Fisher “weight of evidence” approach, the aim is also to control the power of the trial by defining an alternative of interest, as set out by Neyman and Pearson's formal hypothesis test [6]. The implications of statistical power for the posterior reliability of positive findings are not explicitly incorporated into the FDA's current position. They are also not translated into structured decision rules governing evidentiary sufficiency [1]. In addition, although the new framework invokes Bayesian language, the integration of prior beliefs remains implicit [7]. When no past information is incorporated (i.e., when “vague” or “non‐informative” priors are used), the uncertainty in the trial effect, as measured by the credible interval, coincides with that reported in the frequentist confidence interval. However, the Bayesian framework provides more intuitive results by offering a direct estimate of uncertainty in the treatment effect based on a probability distribution, facilitating the interpretation of results, unlike p‐values. Nevertheless, greater contextual flexibility, without explicit definition of priors, may increase interpretive heterogeneity regarding what constitutes “sufficient coherence.” This shift moves health technology assessment (HTA) toward implicit Bayesian reasoning in which prior plausibility influences decisions without explicit modeling [8]. Second, the capacity of “confirmatory evidence,” such as mechanistic data or observational studies, to provide additional evidence of the treatment's benefits is uncertain. There appears to have been an increase in post‐marketing surveillance obligations or commitments at the time of approval by both the FDA and the European Medicines Agency (EMA). Whereas these obligations are often designed to address specific quality, efficacy, or safety concerns that were not addressed during the regulatory evaluation, most appear to only address safety concerns [9]. More recently, a review of 181 post‐marketing requirements for 161 oncology indications following accelerated approvals between January 2011 and July 2023 showed that less rigorous studies may stem from insufficient statements negotiated between the FDA and sponsors [10]. This is in line with our experience at the Commission de la Transparence, where post‐approval studies often fail to address the initial questions about efficacy and safety due to incomplete, heterogeneous data and inadequate methodologies. Moreover, once a product is widely adopted, conducting rigorous comparative studies becomes difficult due to loss of equipoise and recruitment constraints. This emphasizes the need for either the FDA or the EMA to ensure that post‐market studies are well designed and completed quickly, while also ensuring that expedited withdrawal is enforced when necessary [11]. Therefore, assuming that post‐marketing evidence will systematically compensate for reduced pre‐approval replication appears optimistic. Third, design quality and replication address distinct dimensions of uncertainty [1]. As stated in the previous 1998 guidance that allowed some flexibility, data from a single trial could be sufficient for approval [12]. This was only applicable to a adequate and well‐controlled multicenter trial with “very strong results” on several distinct, important outcomes. Although the FDA document no longer considers replication to be the structural default safeguard, it simultaneously acknowledges that additional adequate and well‐controlled studies may be necessary when endpoints are intermediate. However, the use of an intermediate outcome as a substitute for the effects of a treatment on the target outcome(s) of ultimate interest must be demonstrated. This is illustrated by the examples of drugs, authorized on the basis of surrogate endpoints, that were later withdrawn due to an absence of overall survival benefit [13]. This issue cannot be resolved through replication. When endpoints are poorly validated [14, 15, 16], repeating methodological weaknesses does not restore validity. Even though there is a lack of consensus among stakeholders regarding the acceptance and interpretation of intermediate outcomes in trials as surrogate endpoints or target outcomes [17], this should have been clarified in the revised guidance. Whereas design safeguards internal validity, replication tests robustness across contexts by revealing context‐specific biases, analytical fragilities, and idiosyncratic findings. Fourth, it seems that replication fulfills an epistemic function by institutionalizing falsifiability through independent attempts at refutation [18]. Findings reproduced across settings acquire credibility that cannot be replaced by mechanistic plausibility alone. A parallel exists in prediction modeling, where external validation is required to establish reproducibility in new populations. This dimension is particularly salient for Europe. Healthcare systems, standards of care, and patient characteristics differ from those in the United States. When approval relies on a single pivotal study conducted within one environment, extrapolation to another becomes inherently more conjectural [19]. National HTA bodies assessing relative effectiveness within specific health systems consequently bear a greater burden of residual uncertainty. Fifth, early regulatory scientific advice may mitigate these concerns by improving the evidentiary quality of pivotal trials [1]. Both the FDA and the EMA provide protocol guidance before pivotal trials begin. However, such advice is not fully coercive. Sponsors retain substantial autonomy in defining development strategies. In the case of tofersen, EMA's prior advice to extend the study duration to capture meaningful clinical change was not followed [20]. This was probably not due to regulatory passivity but rather to the limited enforceability of recommendations. Lowering the default evidentiary threshold does not enhance leverage. Moreover, recent experience shows that sustained social and political pressure may maintain regulatory decisions despite limited confirmatory evidence [21]. Therefore, redefining the evidentiary threshold would require an explicit and informed public debate because these standards structure collective trust, access to care, and the level of public acceptance of healthcare [21]. Sixth, in an era increasingly shaped by personalized medicine, in which treatments target biomarker‐defined subgroups, such as antibody levels in lupus [22], the evaluation of predictive biomarkers raises specific inferential challenges [23]. Subgroup effects, that is, statistical interactions, are typically estimated with reduced precision and a higher risk of spurious findings. In this context, reliance on a single pivotal trial is generally insufficient to establish the credibility of biomarker‐defined treatment effects. Two independent trials not only replicate the overall treatment effect but also provide a more robust basis for assessing subgroup‐ and biomarker‐defined effects. This may enable the consistency of interaction signals to be evaluated across distinct datasets rather than relying on underpowered and unplanned analyses from a single study. Importantly, this FDA decision cannot be interpreted as a purely domestic adjustment. The United States remains the largest pharmaceutical market worldwide [24], and FDA standards exert a powerful normative and strategic influence on global development programs. This was exemplified by the fast‐track and accelerated approval processes introduced by the FDA in 1988 and 1992, respectively, following the AIDS crisis. This approach was only later adopted in Europe, with France pioneering the field by introducing its “Temporary Authorisation Programme” in 1994 [25]. Major differences exist between the regulatory systems, likely due to structural differences. Lythgoe et al. found that, between 2010 and 2019, 72% of marketing applications for cancer treatments were submitted to the FDA before the EMA [26]. The study attributes this to the larger pharmaceutical market and higher pricing levels for new oncology therapeutics in the United States, among other factors. There are indeed significant differences between the two agencies, likely due to structural differences in their regulatory systems. Both agencies grant most approvals to companies based in their respective regions, and the fewest to those based elsewhere. However, the FDA's approvals demonstrate a stronger regional focus compared to the EMA's [27]. We also wonder whether sponsors frequently design pivotal trials with FDA expectations in mind, and US regulatory precedents often shape negotiations elsewhere. Although European domestic markets are governed by the EMA and national HTA bodies, they still operate within this shared ecosystem [28]. Global development pathways and investor expectations create alignment pressures. Therefore, we hypothesize that a redefinition of evidentiary defaults in the United States will likely reverberate across European countries, even without formal harmonization. Taken together, the FDA's recent decision represents more than just a procedural simplification. It shifts the proof architecture from replication‐based refutation to coherence‐based sufficiency. Due to the significant influence of the US market, the implications may extend to other regions through indirect alignment mechanisms. Flexibility and accelerated access are legitimate policy goals. However, regulatory science must also govern uncertainty in a transparent and institutionally stable manner. Therefore, the central issue is not whether flexibility should exist, but whether reproducibility remains institutionally protected within that flexibility. If replication ceases to function as the structural basis for approval, the center of gravity of regulatory proof appears to shift from empirical refutation to negotiated coherence. In a system influenced by industrial incentives, political urgency, and patient expectations, maintaining explicit safeguards against epistemic drift is not merely a conservative reflex. Rather, it is a prerequisite for maintaining the long‐term credibility and accountability of regulatory institutions. The “one‐trial” approach should be reserved for rare diseases where an adequately controlled, multicenter trial has been conducted and there is evidence of a significant effect on several important outcomes. Author Contributions J.‐C.L., S.C., and S.P. drafted the first version of the manuscript. All authors critically revised the manuscript for important intellectual content and approved the final version for submission. Conflicts of Interest The authors declare no conflicts of interest. Artificial Intelligence Assistance GPT‐5.2 (OpenAI) was used to improve the English language of the first draft version. The authors retained full responsibility for the content of the manuscript. Acknowledgments The authors thank Léonie Belot for editorial assistance. Data Availability Statement Data sharing is not applicable to this article as no datasets were generated or analyzed during the current study. References Associated Data Data Availability Statement Data sharing is not applicable to this article as no datasets were generated or analyzed during the current study.