17.10.2 Discovering the missing context (GLP1–PILOT–1)

17.10.2 Discovering the missing context (GLP1–PILOT–1)

PILOT–0 tests a supplied context. PILOT–1 removes the treatment-status field, its values, the withdrawal-design name, and the declared context object. The constructor receives human-curated arm-level records containing only an investigational agent, population and endpoint types, evaluation time, qualitative trajectory direction, and a sequence of neutral set_exposure events. The comparison role joining the two agents is still supplied. Thus the experiment removes one central semantic aid without conflating context discovery with full-paper parsing or pharmacology.

The frozen grammar contains five possible factorizations: no context, evaluation time, current agent identity, current investigational exposure, and the pair

For each partition, a group is contradictory when it contains both increasing and decreasing weight trajectories. The registered score is ten times the number of contradictory groups plus the number of nonempty groups. This penalty and the five candidates were fixed before running the constructor.

The obstruction is concrete. A current-exposure representation places ordinary control arms and the STEP 4 switched arm in the same “agent absent” class. The ordinary controls have decreasing mean trajectories in the curated records, whereas the previously exposed switched arm has an increasing trajectory. Current state therefore loses relevant history. Refining it by whether the agent occurred before the final assignment removes the conflict and yields four anonymous classes:

Only after the proposal was frozen did the evaluator identify these classes, up to isomorphism, with never exposed, initially exposed and present, previously exposed then replaced by control, and previously exposed and preserved. The resulting presentation is

The constructor compiled the same two-stage experimental shape without using that name: set tirzepatide exposure for every participant, randomly preserve it or replace it with control, and compare subsequent trajectories. The directional prediction was greater regain after replacement. After the proposal hash was written, the neutralized SURMOUNT–4 ledger matched all six registered agent, population, endpoint, assignment, class, and direction fields [ Aronne et al. , 2024 ] .

PILOT–1 passed all 13 gates. The four classes were recovered exactly up to the registered isomorphism; all 24 source permutations produced one proposal hash; and all six controls returned their frozen abstentions. In particular, removing STEP 4 removed support for the history object, removing SURMOUNT–1 removed target support, and adding an explicitly synthetic opposite-direction copy of the switched arm produced ABSTAIN_CONTEXT_CONTRADICTION. SELECT, population transport, and exact-magnitude requests were also rejected. The typed artifact-completeness score was \(1.000\), compared with \(0.455\) for the current-exposure representation. These are field-coverage scores, not independent experimental samples.

Experiment: GLP1–PILOT–1: hidden-context recovery. Input: curated arm timelines with status labels, named design, and context object removed.
Obstruction: current exposure conflates never-exposed controls with a previously exposed arm later assigned control.
Construction: a minimal four-class exposure-history object and a two-stage tirzepatide probe.
Result: \(13/13\) gates, \(6/6\) controls, \(24/24\) invariant order replays, and complete qualitative agreement with the sealed confirmation.
Boundary: finite-grammar context recovery from human-curated events, not open-ended variable invention, prose understanding, prospective discovery, clinical inference, or evidence of identical drug mechanisms.