17.10.1 Treatment status as a missing context (GLP1–PILOT–0)

17.10.1 Treatment status as a missing context (GLP1–PILOT–0)

Active-treatment efficacy and off-treatment maintenance answer different questions. STEP 1 and STEP 5 report weight outcomes while semaglutide is being administered, whereas STEP 4 explicitly randomizes participants after an active-drug lead-in to continued semaglutide or placebo [ Wilding et al. , 2021 , Rubino et al. , 2021 , Garvey et al. , 2022 ] . SURMOUNT–1 supplies active-treatment evidence for the distinct agent tirzepatide [ Jastreboff et al. , 2022 ] . A flat list can summarize these results, but it need not represent treatment status as an argument of the outcome map.

The pilot therefore supplied a human-curated semantic interface: semaglutide and tirzepatide remained different drugs but were grouped under the broad role incretin-based anti-obesity therapy for this particular analogy; body-weight change and cardiovascular events remained different endpoints; and populations with and without diabetes remained different contexts. The constructor was required either to abstain or to add TreatmentStatusAtEvaluation and produce

This extension compiles an external probe: administer tirzepatide during a lead-in, randomize continued treatment against withdrawal to placebo, and compare subsequent weight trajectories. The registered prediction was only directional—withdrawal would produce greater regain than continuation—and contained no effect-size estimate.

The four construction records were STEP 1, STEP 4, STEP 5, and SURMOUNT–1. The resulting proposal and its SHA-256 hash were written before the program opened a separate SURMOUNT–4 confirmation ledger. SURMOUNT–4 in fact used a tirzepatide lead-in followed by randomized continuation or switch to placebo and reported greater regain after withdrawal [ Aronne et al. , 2024 ] . SELECT was registered as a counter-control: although it studies semaglutide in people with overweight or obesity and no diabetes, its endpoint is major adverse cardiovascular events rather than weight maintenance [ Lincoff et al. , 2023 ] .

Every registered necessity and type-safety intervention behaved as required:

• removing STEP 4 yielded ABSTAIN_NO_ANALOG;

• removing SURMOUNT–1 yielded ABSTAIN_NO_TARGET_SUPPORT;

• SELECT yielded ABSTAIN_ENDPOINT;

• extrapolation to type 2 diabetes yielded ABSTAIN_POPULATION; and

• a request for an exact magnitude yielded ABSTAIN_QUANTITATIVE.

All 24 permutations of the construction ledger produced the same proposal hash. On the frozen nine-field artifact rubric, typed construction scored \(1.000\), flat aggregation \(0.333\), and untyped analogy \(0.444\). These are completeness scores over a declared artifact, not sampling estimates.

Experiment: GLP1–PILOT–0: retrospective withdrawal-context transport. Input: four human-curated, source-bearing trial records and a supplied typed interface.
Construction: add treatment status, transport a withdrawal-test pattern across two distinct agents, and compile a qualitative external probe.
Confirmation: open the separate SURMOUNT–4 ledger only after the proposal is frozen.
Controls: remove the analog or target support; substitute the SELECT endpoint; change the population; request an unlicensed magnitude; permute all construction records.
Boundary: one retrospective prediction-and-confirmation artifact, not automated full-paper understanding, prospective discovery, a newly identified biological mechanism, clinical guidance, or evidence that the two drugs have identical pharmacology.