Selected historical trial audit

16 September 2026. GPT-Astra (medium), assisting GPT-Astra (xhigh). AI editorial/source review, not clinical validation. Scope: five additional parent-assigned records only. Registry records were read from the current ClinicalTrials.gov API cache; primary abstracts were retrieved through Europe PMC. The bihormonal iLet and two-year ATG papers were available as full-text XML, including tables. No claim of comprehensive full-text review is made for the others.

ilet-bionic-pancreas-pivotal

Correction: Main randomized comparison is 326 participants (219 iLet with aspart/lispro, 107 standard care), distinct from the registry's 440 across the broader protocol, including faster aspart. Corrected summary denominator and pediatric companion-analysis wording; removed an eligibility maximum derived from observed age rather than the registry. HbA1c adjusted difference −0.5 percentage points, CGM low-glucose noninferiority and severe-event rates agree with the primary abstract. No DKA applies to that main comparison, not automatically to every protocol cohort.

Sources: Primary NEJM report, registry.

Limit: Secondary subgroup equality and every faster-aspart outcome were not independently re-reviewed. Nonsignificant severe-hypoglycemia comparison does not establish identical risks.

ilet-bihormonal-feasibility

Correction: Published report concerns ten participants completing two seven-day periods; registry lists twelve enrolled. Stated both without inventing reasons for the difference. Full Table 1 confirms engineering targets, descriptive glucose metrics, dasiglucagon dose and five insulin leaks. Removed Control-IQ relationship and qualified broad efficacy conclusions. Engineering feasibility and descriptive short-term glucose differences do not establish sustained clinical superiority.

Sources: Primary full report, registry.

Limit: Small, short open-label crossover; not powered to establish rare-event safety. Publication-versus-registry enrollment discrepancy remains source-specific.

imatinib-newonset

Correction: 67 randomized, 64 contributing to 12-month analysis. Added adjusted C-peptide estimate, the one-tailed p=0.048 and 90% interval crossing zero; benefit not sustained at month 24 after randomization, rather than 24 months after treatment ended. Added adverse events and temporary/permanent treatment modifications, which materially constrain benefit interpretation.

Sources: Primary Lancet Diabetes & Endocrinology abstract, registry.

Limit: Primary abstract reviewed; supplementary analyses not independently reproduced. Meeting the prespecified one-tailed test is not equivalent to a conventional two-sided result.

tn-07-oral-insulin

Correction: Retained negative primary outcome and prespecified secondary-stratum identity, but labeled secondary analyses exploratory because of absent multiplicity adjustment. Removed teplizumab relationship and categorical safety wording. Replaced inconsistent subgroup percentages with published event counts (13 oral-insulin, 19 placebo), n=55 and reported HR 0.45.

Sources: Primary JAMA paper, 2024 post-hoc primary analysis, registry.

Source inconsistency: Original abstract itself reports 48.1% and 70.3%; publisher Table 1 gives oral n=28 and placebo n=27. The oral percentage does not reconcile with 13/28. Do not silently substitute a recalculated rate or pretend the original abstract never reported those percentages. Main-group and 2024 post-hoc hazard ratios were checked against abstracts; the latter remains hypothesis-generating.

low-dose-atg-gcsf-haller

Correction: Primary 89-participant study and 29/29/31 allocation verified. ATG+GCSF's p=0.031 at one year misses the prespecified one-sided p<0.025 threshold; the two-year p=0.032 also misses it. Added threshold explicitly, removed teplizumab relationship and overstated winner wording. Added serum sickness in most ATG recipients from the full follow-up report. C-peptide preservation is partial and does not establish prevention before diagnosis.

Sources: 2018 primary trial, 2019 full follow-up, registry.

Limit: Mechanistic cell ratios are correlates, not independently demonstrated mediators of benefit. No independent patient-level reanalysis.

Validation

npm run validate passed after these edits: 186 items, 227 trials, 1,239 citations; no warnings. This checks content structure and references, not medical truth.