An unattended research loop · and a blind test of what it claimed
Sixteen unattended moves left behind a graph of testable claims.
PLSI is a personal health assistant. Given a share of one weekly quota it researches on its own, move by move, and records its working as a graph: the questions it opened, the rival hypotheses it wrote, the finding that would refute each one, and a verdict its own code computes from the records. Five of those hypotheses are claims about whole classes of intervention — which means they can be tested against evidence gathered independently, by a procedure that could not see them.
01 — What the loop left behind
Five questions, seventeen rival explanations, and what each one rests on
Every node below is one PLSI wrote; every link is one it drew. Click anything for its text, its sources and its verdict.
02 — Testing a graph against evidence it never saw
Five of those hypotheses are claims about whole classes of intervention
So they can be checked on records their author never gathered. Independently, 139 free-text proposals from all seven systems were resolved into 88 ideas, authorship was sealed, and an agent searched each one from zero — returning records only. That corpus speaks to the same classes.
The numbers behind each class, and why “holds” is weak evidence
The criteria were written after the corpus existed. They are
the plain reading of each hypothesis's own wording, and they are recorded as descriptive
rather than confirmatory, so a reader who reads the wording differently can substitute their
own and recompute. No pooled verdict is reported: pooling dozens of ideas
always yields both a record for and a record against, so the pooled verdict is
contested for every class however strong the evidence — a property of pooling,
not of the evidence.
03 — The corpus the test rests on
88 ideas, placed by the records rather than by who proposed them
Two axes that cannot contaminate each other: independent human origins behind an idea, and how far the world has already committed to it. Every proposal is placed by the same procedure, and each facet below holds one call — about twenty ideas — so no facet rests on a larger sample than another.
04 — The audit trail is the product
Every idea, every origin, every quote
05 — How it was done, and how hard we tried to break it
Method, checks and limits
How a proposal becomes two coordinates
Independent human evidence is counted from origins — the group that produced a result, not the paper it appeared in — so an aggregate weighs two and is still one source, and an unattributed record counts against a claim, never for it. Occupancy is a ladder from nothing anywhere to an approved product, read from trial registries. The two pipelines never see each other's output. Thresholds were fixed before the data: evidence is high at two independent human origins, occupancy is crowded at rung 2 — both absolute, neither a median split. PLSI is one of the seven arms and computes no coordinate, including its own; the scoring code is the instrument and it ran on sealed rows.