Do bibliometric profile cards change an AI forecast of the 2026 Nobel Prizes?

How the forecaster reasons (both arms). The question resolves on the discovery named in the official motivation, so every run ranks discoveries rather than people. It weighs the strength and recency of outside recognition (Lasker, Breakthrough Prize, Clarivate Citation Laureates and similar awards), how well a discovery fits a Nobel motivation (fundamental physiology that has led to benefit for patients), and discounts options whose credit is spread over many contributors. Awards are explicitly treated as correlated assessments of one achievement, not as independent signals to be multiplied, and the runs repeat that public awards do not reveal the committee's confidential shortlist.

GLP-1 leads in every run (17–23 %). Its hormone-to-treatment pathway is described as the strongest cumulative case on the list: the 2024 Lasker Clinical Award, the 2025 Breakthrough Prize and Clarivate's 2026 selection, joining endocrine physiology to large clinical benefit (including the cardiovascular indication of semaglutide). The main discount is attribution: several groups contributed, although a different trio of laureates would still resolve to this option.

Orexin is second (11–14 %), on fresh physiological and therapeutic validation (a new basic-science award and direct therapeutic validation of the sleep–wake pathway).

“Other” is the largest single outcome (29–35 %). The runs argue that the list omits serious contenders, naming DNA sensing (cGAS–STING), optical coherence tomography and the targeted treatment of acute promyelocytic leukaemia, and that even the strongest named option still loses about four times in five.

What the cards changed. In the treat runs the subforecasts add a passage on “the supplied bibliometrics”: they are used as a secondary cross-check and deliberately not converted into likelihood ratios, because percentiles among past laureates are not calibrated selection probabilities, author disambiguation is imperfect and there are no unsuccessful candidates for comparison. The ranking logic is otherwise unchanged. The small shifts point the way the cards' laureate comparisons do (our reading, exploratory): optogenetics, whose lead person Karl Deisseroth is above 98 % of past laureates on impact, gains 1.4 points and rises three places; GLP-1, whose first-named person Svetlana Mojsov is above 27 %, loses 1.4 points, within the range of the control runs.

The run weighs profiles and awards side by side. It states that the profiles get "influence comparable to awards and other external evidence", then runs three checks: whether orexin's 2026 milestones are substantive, whether the profiles change a headline-led ranking, and whether omitted discoveries justify a large "Other". A base-rate audit frames the awards: 18 of 26 Medicine prizes since 2000 went to a team with a prior Lasker, but only 1 of 25 Lasker Basic awards matched the same year's Nobel, so a fresh Lasker is read as readiness, not as an imminent prize.

Orexin leads at 11.7 % (control 12.1 %). The case combines its defining papers (de Lecea's hypocretin paper: disruption percentile 0.94, foundation share 0.86) with the 2026 Lasker and the FDA approval of an orexin-targeting drug in August 2026.

GLP-1 halves but survives: 20.3 % → 11.0 % (main-evidence run: 3.0 %). Holst's inventions and textbook reach are called exceptional, but the team's career-wide impact comparisons are "less uniformly exceptional than those of several competing teams"; the dense award record (Lasker 2024, Breakthrough 2025, Clarivate 2026) keeps it second.

Profile-strong options gain modestly: CAR-T 6.0 → 8.6 %, Wnt/organoids 2.3 → 5.4 %, leptin 1.9 → 3.5 %, PCSK9 2.9 → 4.1 %; the unfolded protein response falls 5.8 → 3.5 %. "Other" is 34 %.

The run turns the cards into a scoring model. The main-evidence run scores every named person from the card figures (impact, technological translation, textbook reach and collaboration weighted 40/30/25/5), blends the strongest person with the group mean (65:35), adds the option's defining works (20 %), and averages 81 weightings with a 10 % uniform component. It says it accepts the premises "without a vintage or attribution penalty". Prizes and news no longer rank the named options; they are used to size "Other" (40 %), which now lists optical coherence tomography, integrins, cGAS, TOR and low-complexity domains as unlisted contenders.

The award favourite collapses. GLP-1 falls from 20.3 % (control mean) to 3.0 %: the write-up cites Holst's translation figures but finds the group's impact comparisons and Mojsov's and Knudsen's book reach "below the leaders". Its 2024 Lasker and the 2026 Clarivate selection are mentioned but no longer move it. Orexin, the other award-backed option, falls from 12.1 % to 4.8 %.

Profile-strong options rise. Optogenetics (7.9 %; Deisseroth with 37 % of works in the cohort top 1 %, 103 patents, 3,009 citing books) and Wnt/organoids (7.7 %; Clevers with 2,724 citing inventions) lead the named options, followed by inherited breast-cancer genes, leptin and PCSK9 (5.7–6.4 % each). The named options now lie between 2 % and 8 % instead of led by one or two favourites.

The same-day control run is unchanged: GLP-1 18 %, orexin 12 %, Other 32.9 %, as on 1 October.

How the forecaster reasons (both arms). As in Medicine, every run ranks discoveries by how compact and well recognised the discovery is, treats public prizes as evidence of scientific stature rather than of the committee's confidential choice, and spreads probability widely: no named discovery is allowed a dominant share.

“Other” is the largest single outcome in every run (29–35 %), because the menu omits several recognised families of discoveries; the runs stress that “Other” stands for many omitted candidates, not for one stronger rival.

Optical lattice clocks lead the named discoveries in every run (14–16 %). The write-ups call them the strongest compact discovery case: the Breakthrough Prize for their invention and development, and a Wolf Prize announced on 1 October 2026 that recognises Jun Ye (clocks) and Immanuel Bloch (optical-lattice quantum simulation); the runs split that update between the two options rather than crediting it to clocks alone, and note that the redefinition of the SI second has not happened yet.

Topological insulators follow (8–11 %), then negative-index metamaterials (7–9 %). A Nobel Symposium on metamaterials is read as strengthening that field's case without identifying the winner. Quantum information and quantum simulation receive a small timing discount for adjacent recent prizes (entanglement in 2022, macroscopic quantum tunnelling in 2025), while the runs warn against treating adjacent science as already rewarded.

What the cards changed. The treat subforecasts again discuss the supplied bibliometrics as secondary evidence (about eight references per run, against 0.5 in control) without turning them into odds. The changes are small: “Other” falls by 1.6 points (within the control range), aberration-corrected electron optics gains 0.9 points and is the one clear move outside the control range (cosmic inflation, −0.2 points, sits exactly at the lower edge of its control range), and optical lattice clocks and quantum simulators gain about 0.8 points each, within their control ranges. Unlike Medicine, the shifts are not related to how the options' people compare with past laureates on their cards (exploratory correlation 0.07).

The run writes the balance down as a formula. Each discovery gets a profile-support score B (impact, defining works, textbooks, translation, collaboration) and an outside-evidence weight R (validation, maturity, honours, current recognition, matching risk). Equal-weight logarithmic pooling gives w = exp(3B)·√R; the named options share 70 % in proportion to w and "Other" keeps 30 %.

It reports what the profiles changed: outside evidence alone would give optical lattice clocks 17 % and quantum simulation 10 %; with the profiles they get 9 % and 13 %.

Quantum simulation with ultracold atoms leads at 12.8 % (control 5.4 %, main-evidence run 13.3 %): Bloch's 181 works have median impact percentile 0.99 with 49 % in the cohort top 1 %, Cirac and Zoller add impact and textbook reach, and the 2026 Wolf Prize citation credits Bloch's lattice simulation.

Optical lattice clocks fall to 9.3 % (control 14.7 %, main 4.5 %). They keep the strongest outside weight (Breakthrough, Wolf), but Katori's textbook profile is weaker than the simulation group's.

Elsewhere: magic-angle graphene rises 6.5 → 9.2 %, topological insulators stay at 9.2 % (control 10.0 %), aberration-corrected electron optics drops 4.3 → 1.9 %. "Other" is 30 %, lower than control (33.5 %).

The run builds a profile ensemble. Four weightings of the card figures (career pathways, person-level balance, discovery-specific scores, an impact-heavy variant; pooled 15/35/35/15) give a probability among the named options; "Other" is set first (38 %) and the named options share the rest. People are averaged or blended with the strongest contributor, so an option does not gain merely by naming more people. Outside information enters as small multipliers: the 2026 Wolf Prize raises quantum simulation by 10 % and clocks by 20 % in relative weight.

The award favourite falls. Optical lattice clocks drop from 14.7 % to 4.5 %. Ye's impact counts for them, but Katori's textbook comparisons exceed only 27 % and 18 % of past laureates.

New order. Quantum simulation with ultracold atoms leads at 13.4 % (Bloch with 49 % of 181 works in the cohort top 1 %; Cirac and Zoller's textbook reach), ahead of negative-index metamaterials (8 %), magic-angle graphene (7 %) and topological insulators (6 %). Cosmic inflation rises from 0.9 % to 3.3 % on Linde's impact.

The same-day control run is unchanged: optical lattice clocks 15 %, topological insulators 10 %, Other 36.5 %.

How the forecaster reasons (both arms). The runs rank discoveries by the breadth of past Chemistry awards and the recognition record of each discovery, treat awards as correlated evidence of one achievement, and keep the forecast dispersed: no named discovery approaches an odds-on favourite.

“Other” is the largest single outcome in every run (32–35 %). The write-ups name concrete omitted contenders rather than generic uncertainty: Clarivate's 2026 Chemistry selections (self-assembled monolayers, long-range electron transfer in proteins, base and prime editing), the 2022 Kavli Prize for self-assembled monolayers, Buchwald–Hartwig carbon–nitrogen coupling, and recent Gairdner awards (systems proteomics, cryo-electron tomography). Several of these were just outside the committee's top twelve (self-assembled monolayers 13th, electron transfer in proteins 18th, Buchwald–Hartwig coupling 19th after it was split from C–H functionalization); base and prime editing were not nominated at all.

Sequencing-by-synthesis leads the named discoveries in every run (18–20 %). It combines mature practical impact with repeated recognition of the same discovery, most recently a new Wolf Prize in Chemistry. The runs calibrate that signal with a small reference class: of 18 Wolf Chemistry cohorts, only two (2001 and 2022) contained a same-year Chemistry laureate, so the Wolf Prize is read as a standing signal, not a timing detector. A stress test in one run moves the sequencing forecast between 13 % and 23 % depending on how much the new award is worth. Perovskite solar cells (about 8 %) and controlled radical polymerization (about 7 %) follow.

What the cards changed. The treat subforecasts refer to the supplied bibliometrics about eleven times per run (control: none) and again use them as secondary evidence. The ranking is stable (Spearman 0.96); five small options move just outside their control ranges (molecular chaperones +1.0 point, targeted protein degradation +0.5, nanopore sequencing +0.4, C–H functionalization −0.6, DNA origami −0.3), while the largest single change, perovskite solar cells −1.1 points, stays within its control range. The shifts are not related to how the options' people compare with past laureates on impact (exploratory correlation 0.01).

The profiles get half the weight among the named discoveries. The run states that the supplied profiles "receive half the weight in the comparison among listed discoveries", emphasises defining work, technological translation and textbook influence, keeps career impact as substantial evidence and gives collaboration less weight. It judges a discovery through its strongest contributors rather than its weakest named person.

"Other" leads at 37.5 % (control 33.6 %): the run calls the breadth of the omitted field "the central forecasting problem".

Sequencing-by-synthesis stays the leading named discovery but halves: 19.0 % → 10.0 % (main-evidence run 5.5 %). It has the strongest external package (2026 Wolf Prize, 2026 Princess of Asturias award, a named specialist prediction) and a strong invention record (1,532 citing inventions for the reversible-terminator paper), which keeps it first.

Profile-strong options rise: nanoparticle delivery 6.0 → 8.4 % ("the strongest overall profile": Langer's 20,350 citing inventions), perovskite solar cells 5.9 → 6.8 % (Snaith's median impact percentile 0.98; Kyoto Prize 2026), OLEDs 3.7 → 5.7 %, nanopore sequencing 2.4 → 3.9 %. C–H functionalization (5.6 → 4.3 %) and controlled radical polymerization (6.5 → 5.5 %) fall.

The run scores the profiles and calls the result spread out. The write-up allocates 62 % across the named options from the card figures and says the profile evidence "supports a spread-out forecast, not a dominant named favorite". "Other" is set at 38 % as a coverage judgment, with a stress test from 28 % to 48 %.

The award favourite falls. Sequencing-by-synthesis drops from 19.1 % to 5.5 %. Its shared 2008 paper has 1,532 citing inventions, but "Mayer's smaller scholarly record moderates the group score" (Mayer's card is one of the thin profiles). Its fresh Wolf and Princess of Asturias awards, which lead the same-day control run, no longer rank it.

Translation-heavy options rise. Efficient OLEDs lead at 9.4 % (Van Slyke's 1987 paper: disruption 0.95, 2,409 citing inventions; Thompson's 4,962 citing inventions and 219 patents, above 98 % of past laureates), followed by perovskite solar cells (7.6 %; Snaith's impact), nanoparticle delivery (6.9 %; Langer's 20,350 citing inventions) and nanopores (5.9 %). Controlled radical polymerization and C–H functionalization fall.

The same-day control run is unchanged: sequencing-by-synthesis 20 %, Other 34.6 %.