Retrospective decision-governance evaluation across heterogeneous clinical programs
Year: 2026
Type: preprint
Abstract: Separating Program Failure from Target Failure A Temporally Restricted Backtest of the HISTAMOS Failure Attribution Standard and a Preregistered Validation Protocol Retrospective decision-governance evaluation across heterogeneous clinical programs El Amin Bouchelit HISTAMOS — Independent Research Project, GermanyCorrespondence: [email protected] Website: https://www.histamos.com/ Public evidence cutoff for this manuscript: 31 August 2026 Abstract Background. A negative clinical trial does not automatically establish target failure, molecule-wide failure, or a valid rescue strategy. Development decisions can therefore be wrong even when the trial result itself is clear. We evaluated a public failure-attribution standard designed to separate the observed miss from the scope of what the evidence actually falsified and from the action that can reasonably follow. Methods. Eight heterogeneous historical programs were evaluated retrospectively using evidence restricted to information that was available at a defined earlier time point (T0), before later resolving evidence (T1). The analysts were not prospectively blinded to T1, so this is a temporally restricted backtest rather than a prospective validation. Case-level decisions were judged after T1 as fully aligned, partly/mixed aligned, or hard-contradicted. Three high-cost error classes were tracked separately: false target death, false global molecule death, and false rescue promotion. Xevinapant was retained as an adversarial confirmatory-reversal control. Results. Across the eight primary cases, five were judged fully aligned with later resolution and three partly/mixed aligned; no hard action contradiction was observed. No case was scored as false target death, false global molecule death, or false rescue promotion. The benchmark spanned context-specific failure, repeated confirmatory nulls, biomarker activity without verified clinical benefit, divergent replication, post-hoc subgroup rescue tested prospectively, and interim futility followed by a larger completed dataset. These results support bounded retrospective utility as decision governance, but do not establish mechanistic accuracy, prospective calibration, or superiority to expert review. Prospective protocol. We preregister a larger blinded benchmark of 20–30 historical programs using T0 evidence packets, independent expert reviewers, HFAS review, and a simple scope-only comparator, with all decisions frozen before T1 reveal. The primary outcome will be action fidelity; key safety outcomes will include false target death, false molecule death, false rescue, over-attribution, abstention appropriateness, and inter-rater agreement. Conclusion. The central proposition is that decision sufficiency does not require mechanistic completeness. A development team may have enough evidence to avoid an unjustified target-death, molecule-death, or rescue decision even when the final causal mechanism remains unresolved. The present backtest is hypothesis-generating; the preregistered blinded benchmark is designed to provide the stronger test. Keywords: clinical trial failure; drug development; causal attribution; decision governance; target validation; biomarker rescue; replication; prospective validation. 1. Introduction Clinical-development failures are often compressed into labels such as “the target failed,” “the molecule failed,” or “the subgroup should be rescued.” Those labels can exceed what a trial actually established. A study may falsify a particular therapeutic role, dose, population, endpoint, combination or development claim while leaving other levels of the causal chain unresolved. Conversely, a biomarker change or a favorable subgroup may survive a primary miss without demonstrating verified clinical benefit. This distinction matters because the consequences of over-attribution are asymmetric. Declaring target death too early can terminate a viable biological program; declaring molecule-wide failure can erase context-specific value; and promoting a post-hoc rescue can consume additional capital without a prospectively validated treatment interaction. A useful failure-attribution framework should therefore be judged not only by whether it can narrate a mechanism, but also by whether it avoids high-cost development errors. The HISTAMOS Failure Attribution Standard (HFAS) is presented here at a public, non-proprietary level. It asks four simple questions: What event occurred? What scope of claim did that event directly test? What remains unresolved? What development action is justified by the available evidence? The public standard does not require a mechanistic explanation when the evidence cannot identify one. Figure 1. Public decision problem addressed by HFAS. The purpose is not to force a mechanism after every miss, but to keep the scope of falsification and the justified action aligned with the evidence. 2. Public definition of the standard For this manuscript, HFAS is defined by the following public principles. These principles describe the scientific output of the standard and do not disclose the internal computational implementation. · Program failure, target failure and rescue opportunity are separate claims and require separate evidence. · A primary endpoint miss establishes failure of the tested claim before it establishes a broader biological conclusion. · Later evidence can resolve whether an earlier miss was context-specific, confirmatory, transient, or incorrectly over-interpreted. · A positive biomarker or pharmacodynamic result can preserve evidence of biological activity without establishing clinical benefit. · A subgroup or rescue hypothesis requires prospective treatment-effect evidence before it is treated as a validated development path. · Abstention is appropriate when the evidence can support an action boundary but cannot localize the final mechanism. The practical objective is therefore decision governance: prevent the evidence from being made to say more than it actually says. 3. Retrospective temporally restricted backtest 3.1 Design The R1 benchmark used eight primary historical cases selected to represent qualitatively different ways in which development evidence later resolved. For each case, an earlier evidence boundary (T0) was reconstructed from public material that existed before the later resolving event (T1). The central question was whether the action and scope that would have been justified from T0 remained defensible after T1 became known. The term temporally restricted is deliberate. The evidence packets were restricted by historical date, but the analysts performed the reconstruction after later outcomes existed and therefore could not be prospectively blinded to T1. This design can identify gross contradictions and stress-test scope discipline; it cannot by itself establish prospective predictive validity. 3.2 Case set Program Resolution pattern Benchmark question Asundexian AF replacement failure followed by efficacy in a different therapeutic role Tests whether failure is scoped to comparator/role rather than generalized to target or molecule. [1,2] Navacaprant Initial Phase 3 miss followed by additional Phase 3 nulls Tests whether repeated confirmatory nulls appropriately collapse generic rescue space. [3,4] Simufilam Two concordant Phase 3 efficacy failures Tests indication-scoped molecule failure without automatically inferring target-level failure. [5,6] Sacituzumab govitecan Broad NSCLC primary miss with a favorable non-alpha-controlled subgroup Tests whether subgroup signal is preserved as hypothesis but not promoted as validated rescue. [7] Tofersen Strong target/biomarker effects with a negative randomized 28-week clinical primary endpoint Tests separation of biological activity from verified clinical benefit. [8] Pirfenidone Divergent Phase 3 replication followed by a successful additional confirmatory trial Tests whether one negative replicate is distinguished from completed confirmatory reversal. [9] Dalcetrapib Post-hoc pharmacogenetic rescue tested in a prospective genotype-selected trial Tests whether retrospective subgroup rescue is withheld pending prospective confirmation. [10] Aducanumab Interim futility stop followed by a larger completed dataset with discordant Phase 3 outcomes Tests whether interim futility is distinguished from a completed confirmatory dataset. [11,12] 3.3 Adversarial control Xevinapant was retained as an adversarial control because an earlier randomized Phase 2 study reported substantial long-term benefit, whereas the larger Phase 3 TrilynX study later reversed direction, with shorter event-free survival, worse overall survival and a less favorable safety profile. A standard that reflexively preserves every prior positive signal would fail this case. The appropriate response to this pattern is to stop the broad tested claim and avoid rescue unless new independent evidence identifies a narrower, prospectively testable hypothesis. [13,14] 3.4 Outcome definitions Case-level action fidelity was adjudicated in three categories: · Full alignment: later evidence supported the earlier scope/action without requiring a material change in decision direction. · Partial or mixed alignment: the central action boundary remained defensible, but later evidence required narrowing, qualification, or left important components unresolved. · Hard contradiction: later evidence directly showed that the earlier action or scope would have been materially wrong. Three asymmetric high-cost errors were tracked separately: false target death, false global molecule death, and false rescue promotion. These are development-safety metrics rather than claims of mechanistic accuracy. 4.
Author: Bouchelit, Mohammed El Amin
Language: English
Subfield: Statistics and Probability
Field: Mathematics
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W7204807157