Modern AI scientific discovery

Instruments, not oracles.

LLMs accelerate literature triage and formalization drafts. They do not replace versioned experiments, baseline comparisons, or the inventor's verification step before anything goes public.

Observers on manifolds

ML is a dynamical observer on structured carriers — not an authority. Outputs ship with claim ceilings, comparator passes, and explicit abstention when baselines still win.

Artifacts before adjectives

Structured results, human report, checksum. No upgrade from summary to validated claim without bound evidence at reproducible artifacts.

Human-verified outward claims

Tools assist; the inventor owns public language. AI disclosure is plain; overclaim is treated as a failure mode, not a marketing option.

Discovery pipeline

Literature → hypothesis → experiment → gate → preprint

Same electron-flow discipline as the System diagram — directed edges, dwell at gates, emit only on pass.

Climate / EWS

Interpretable early-warning features vs deep-learning baselines on published benchmark families — scoped, publication-dependent claims only.

Methods-ready · not universal predictor

Materials

Composition-to-phase signals on public elasticity datasets — classical methods demonstration without quantum hold-up for filing narrative.

Methods-demonstrated on cited public N

Medical dynamics

Hypothesis-generating phenotyping and cohort de-mixing — research-stage, not diagnostic clearance.

Hypothesis-generating only

Formal mathematics

Statements in library or finite certificates cited externally — not unsourced "solved" language.

COMPILED gate for proof claims