The questions
statisticians ask.

A futility signal can only remove opportunities to reject the null, so it cannot inflate Type I error and no alpha is partitioned — this is a mathematical property. For detection of success that does not initiate early stopping, the trial runs to its pre-specified completion with sample size and final analysis unchanged, so the null distribution of the final statistic is untouched and alpha stays nominal. All of our packages release non-binding signals, so the decision to act on them is up to the sponsor.

The cost of any futility rule is stopping a trial that would have succeeded. On our dataset derived from the outcome distributions from phase 2 trials on CT.gov, the balanced configuration has a false futility rate of 1.0% and the conservative configuration 0.5%, against 69.8% of null trials stopped at a mean information fraction of 50%. Conditional power at the conventional 0.2 threshold, taken once at midpoint, gives 4.8%. Matched to our futility package on sensitivity, conditional power at 0.07 still gives 2.2%. Continuous monitoring costs less power than the single look most protocols already contain.

You provide us with a sample size, endpoint, assumed effect size and alpha, and we calibrate the configurations against a synthetic dataset emulating ClinicalTrials.gov distributions. Our model controls the false signal rate to 1%, and conservative configurations to 0.5%. Because a rate that is calibrated at one effect size is worth little if your trial differs, each configuration is also evaluated across a grid of fixed clinically meaningful effect sizes including the null and your SAP assumption, with the false positive rate required to stay under 2.5% throughout.

No, for a structural reason. Adjustment is required when repeated looks create repeated opportunities to declare success. A futility signal only ends trials early, removing opportunities to reject the null hypothesis rather than increasing it, and a success signal changes nothing at all. The final test happens once, on the pre-specified sample, at the pre-specified alpha. Therefore, look frequency is not something you pay for as the bottleneck in conventional practice is the cost of unblinding people, which the enclave removes.

Our analysis packages run parallel to existing designs that you may already have. Your efficacy boundaries and alpha spending function are untouched, since our packages neither stops for efficacy nor alters the final test. What we provide is continuous futility and success detection between your scheduled looks, reporting to the committee that governs them.

Pre.sentient only analyses completed patients, so a patient who has not reached the primary endpoint assessment contributes nothing rather than an imputed value, and the information fraction is defined on completions. Our analysis packages run on preliminary data below submission grade, missing data handling follows a plan pre-specified before the trial opens, and every signal carries a documented completeness score.

Two-arm trials with continuous outcomes are the base case that we support. Time-to-event endpoints use a test statistic adjusted for right-censoring, compared against a boundary defined on accumulating events rather than completed patients. Extensions to further designs, such as single-arm, are in development and will substitute in the appropriate statistical test while the core prediction mechanism stays unchanged.

Our model is a statistical formula that runs inside a sealed environment that no human can access. The algorithm is reproducible and fully documented in our whitepaper. Our model does not incorporate machine learning, and every enclave computation is logged cryptographically to a tamper-resistant log from which an independent third party certifies that blinding is held. Our ISO/IEC 27001:2022 and ISO 9001:2015 certifications from UKAS-registered auditors are publicly verifiable on CertCheck. If you wish to evaluate our model's performance, our biostatisticians can provide a simulation code that can be calibrated independently.

On an empirically derived Phase 2 dataset, our model has been validated at scale. Built from distributions estimated from completed trials on ClinicalTrials.gov, the futility configuration stopped 69.8% of null trials at a mean information fraction of 50%, with a false futility rate of 1.0%. Conditional power at the conventional 0.2 threshold, taken once at the midpoint, gives 4.8% on the same data, and 2.2% even when matched to the futility package on sensitivity. The algorithm has also been replayed against endpoints from completed trials, including ADAPT-2.