Can HHS cut clinical trials below four years without weakening the evidence?
ARPA-H wants predictive models, common controls, continuous statistics and agentic operations to replace fragmented trial phases. Three linked projects carry awards of up to $100.03 million, but SURPASS is still an open solicitation and no shortened trial or patient benefit has been demonstrated.
The short version
- What happened
- SURPASS seeks trials lasting under four years through simulation, common control groups, continuous analysis and automated operations, but it is an open research solicitation rather than a validated trial method.
- Who it affects
- Patients could reach promising trials sooner and face fewer repetitive visits, but need clear consent and assurance that speed does not weaken safety evidence.
- What remains unproven
- SURPASS is an open solicitation and has not yet produced a funded platform, shortened trial, approved treatment or measured patient benefit.
Keep in mind
SURPASS is an open solicitation and has not yet produced a funded platform, shortened trial, approved treatment or measured patient benefit.
The programme's decade, $1–2 billion and 90% failure figures are broad agency summaries; they vary by disease, product and definition.
The three linked award amounts are ceilings contingent on project terms and milestones, not evidence that the full $100.03 million has been paid.
COMMONS' 40-million-person and three-million-record targets do not establish current participation, usable data or representativeness.
Multi-source analysis · 5 linked sources. Primary analysed source date: 30 September 2026. Earliest linked supporting source: 21 September 2026.
Read the full report · 6 min The evidence, methods, implications and limitations
The programme aims to replace stop-start trials with a continuous platform
The US Department of Health and Human Services announced SURPASS on 30 September through ARPA-H. The programme asks whether drug and biologic trials that often take more than a decade can be redesigned to take less than four years. Its proposed platform would combine predictive computational models, shared infrastructure, common control groups and statistical analysis that remains valid as data accumulate, allowing treatment arms and decisions to change without rebuilding a trial from the beginning.
A third technical layer would use agents and automation for activities such as site start-up, onboarding new treatment arms, data collection, cleaning and dataset construction. This is a research agenda, not a national rule replacing conventional trial phases. The solicitation remains open: required solution summaries are due on 30 November 2026 and full proposals on 22 January 2027. No SURPASS performer, trial population, drug or demonstrated time saving is reported yet.[1][2]
Three linked projects address sites, data and patient navigation
ARPA-H announced three complementary projects alongside the open SURPASS call. STACK carries an award of up to $41.18 million to create a governance framework, trial-activation consortium and navigator network. COMMONS carries up to $49.9 million for a consent-governed national data resource, and CINCH up to $8.95 million for an oncologist-reviewed service that helps adults with cancer assemble records, understand gaps and connect with suitable care or trials. The three published award ceilings total $100.03 million.
The pieces address real bottlenecks: a statistically elegant design cannot help if a hospital takes months to open a site, records cannot be linked lawfully, or a patient cannot identify an appropriate study. Yet combining them also expands the system's risk surface. Governance has to cover automated administrative decisions, record matching, consent changes, site quality, commercial partners, cyber incidents and the possibility that a navigation system directs a person toward an unsuitable trial.[1][3][4][5]
The largest denominator is a data ambition, not an enrolled cohort
COMMONS says it targets more than 40 million individuals and three million linked records, including data from rural, community and safety-net providers. Those figures describe the intended reach of a future resource, not the number of people enrolled in a trial or a sample that has already been validated. The public award page does not show how many organisations have signed, which data fields will be complete, how duplicates will be resolved or how many people will consent to each type of research access.
Scale can improve analysis of uncommon outcomes and under-represented groups, but millions of records do not automatically remove bias. Health records reflect who reaches care, what clinicians document, which tests are affordable and how coding differs among institutions. A regulatory-grade resource needs traceable provenance, data-quality measures, transparent inclusion rules and a way for consent choices and corrections to propagate. Researchers also need to report the usable denominator for every analysis rather than citing the larger network total.[3]
Fewer participants can be ethical only if uncertainty stays visible
Common control groups and predictive simulations may reduce the number of people assigned to standard care, especially when several treatments share the same disease platform. Continuous inference may stop a harmful arm sooner or expand a promising one. But historical and modelled controls can differ from current participants because care, diagnostics, population mix and data collection change. A digital twin that predicts an untreated outcome must be tested for calibration and transportability before it substitutes for observed evidence.
Always-valid analysis also requires pre-specified statistical safeguards. Repeatedly examining results and adapting allocation can inflate false-positive findings if the design and decision rules are not controlled. Regulators, ethics boards and participants will need to understand which changes can occur automatically, who approves them and how safety signals are escalated. The goal should be fewer unnecessary participants, not a lower evidential threshold for approving treatments.[1][2]
What would show that faster still means trustworthy
The strongest evidence would be pre-registered demonstrations comparing SURPASS-supported trials with relevant conventional programmes. Measures should include calendar time by stage, site-activation time, enrolment and retention, cost, participant burden, protocol amendments, data-query rates, statistical error control, representation, safety detection, regulatory review and whether later studies reproduce the conclusion. Agentic operations should also report exception, human-override and error rates rather than only tasks automated.
Confidence would rise if protocols, simulation code, statistical plans, consent governance and regulatory feedback become public, and if trial results remain robust after follow-up. It would fall if speed comes mainly from narrower populations, opaque synthetic controls, reduced safety observation or data access that patients cannot understand or contest. The development is material because the US government is funding infrastructure around the whole trial pathway. Its success must be judged by evidence quality and patient outcomes, not the number of AI components or the size of the announced awards.[1][2][3][4][5]