How novel was the AI discovery?
You’re reading the published research snapshot. Interactive filters become available when the index loads. Claim titles open their primary sources; each mark on a timeline opens the work it stands for.
| DISCOVERY / CLAIM | CLOSEST EARLIER WORK by others | NOVELTY ANALYSIS |
|---|---|---|
| new resultArray-associated reverse transcriptases in jumbo phages Claude identified a family of jumbo-phage reverse transcriptases associated with repeat arrays and a partner gene; the arrays produce abundant short RNAs, but the system's function is not yet known. Evidence & sources
The exact ART architecture appears new as a documented biological system, but its components have close precedents. Retrons already combine RT, noncoding RNA and accessory proteins; DRT2 already uses an RT and neighboring ncRNA to generate tandem repeats; and genome-context mining of antiphage systems is established. The strongest novelty is the specific jumbo-phage lineage plus genomic repeat-array organization and expression phenotype. No new enzymatic function, defense mechanism, or programmable CRISPR-like activity has yet been demonstrated. The expanded screen does not move the level: the patent space anticipates every ingredient — RT on structured noncoding nucleic acid, RT plus an accessory protein, RT-generated tandem repeats, and defense discovery by genome-context mining — without disclosing the combination. Anthropic itself states that the underlying reverse transcriptase had been identified before; what it reports as unrecognised is the adjacent partner gene and the repeat array. A proximity ranking across both literature and patents puts Wilkinson's DRT2 paper first and WO2025101943A1 second, and neither discloses the combination. None of the closest comparators shares an author with the technical report: the author overlap that exists is with the same team's adjacent RNA-guided DNA engineering work, which is background rather than prior art against this claim. A lineage audit narrows what the level covers. The inventors behind the retron patents — Rodríguez Mestre above all — published RT-plus-partner-gene defense systems repeatedly from 2018 onward, so the general claim of discovering an RT with a partner gene and a noncoding RNA is not new. What that lineage does not contain is an ART RT family, a jumbo-phage ART locus, a long genomic repeat array or multiple discrete array-derived RNAs, and it is not confirmed that anyone in it reached the ART family first. The level therefore holds for the exact architecture and expression phenotype, and for nothing broader. Why this level: A previously uncharacterized jumbo-phage RT clade with a dedicated partner gene, a genomic array of short repeated DNA units separated by long spacers, and abundant discrete array-derived RNAs. The exact combination was not found in the pre-claim comparison corpus. A subsequent patent screen of six close families found no pre-cutoff disclosure of the same combination either, which leaves the level where the literature put it rather than raising it. A lineage audit of the retron patent inventors reached the same place: they repeatedly discovered related RT defense systems, and none of that work discloses the ART family or locus. Prior-art coverage: B · 72 of 100 Quantitative distance: Claimant-reported counts only: 95 RT clusters, 28 array-associated loci, arrays of 3-21 repeat copies of 15-49 nt separated by 120-220 nt spacers, and array-derived transcripts reaching up to 8% of reads during infection. No pre-claim like-for-like count exists, because the ART family was not previously defined, so there is no before/after delta to state. Validity: reported — Released with a technical report and computational evidence, but not independently re-verified here. The level holds if the reported ART locus architecture and expression results reproduce; function and mechanism remain unknown, and no wet-lab characterisation of the array or the partner gene is reported. Result kind: study — A computational survey reporting a locus architecture and an expression phenotype, not a made object or a theorem. AI contribution: authored — Claimant-reported: the model ran the genome mining, clustering and expression analysis and wrote the report. Genome-context mining of defense loci is an established human method (Doron et al., 2018), so the model supplied the process, not the strategy. First-year window: 5 of 365 days observed · window closes 2027-09-23. Five of 365 days observed as of 28 September 2026. Nothing independent has appeared, and the window is far too young for anything to have. The lineage audit found continued retron discovery by the same groups, but no work that reaches the ART family. Author overlap: None of the 5 closest comparators shares an author with the claim. 2 earlier papers by the same team are listed above and marked SAME TEAM. The author overlap is with adjacent background and engineering work, not with the closest mechanistic prior art. For attribution, the two overlapping papers should be marked same-team earlier work rather than independent prior art; their existence does not by itself anticipate the ART claim. Lineage: Traced from WO2024044723A1 (priority 25 Aug 2022) through 6 earlier papers by its inventors. Not confirmed that the patent authors discovered the same ART protein family before Anthropic. Confirmed that the lineage repeatedly discovered and characterized related bacterial RT/retron defense systems. The broad system concept is prior art; exact ART family/locus architecture remains unanticipated in the reviewed lineage. Patents: No exact ART patent anticipation was identified in the targeted title/abstract search. The pre-claim patent space is crowded with adjacent retron, DRT2/RT-DNA, CRISPR, and genome-mining claims. Infobase patent search returned records, but batched add_patents timed out and a smaller retry hit an existing duplicate citation edge. Patent conclusions therefore use the cloud search output plus public Google Patents records, not a fully persisted patent corpus. This is a prior-art screen, not a patentability or freedom-to-operate opinion. Ranked pre-cutoff patent families:
Review limits: The exact ART locus accession-level comparison, exhaustive jumbo-phage survey, biochemical mechanism, antiphage function, and post-publication one-year follow-up remain open. The claim was public on 2026-09-23, so the one-year post-publication window has observed 5 days. This is a literature prior-art assessment, not a patentability opinion, and the patent screen is a title/abstract search of six close families, not a claim-by-claim reading, a patentability opinion or a freedom-to-operate opinion. Neither the literature expansion nor the patent corpus was persisted, so both rest on searches that were not stored for re-audit. The lineage audit is a named-inventor trace, not an exhaustive one, and Q222 never completed. Still open: Establish whether the arrays are transcribed as a unit or as discrete RNAs, determine whether the RT acts on them, test whether the partner gene is required, and check the loci accession by accession against a complete jumbo-phage survey. Then run the biochemistry: is there a reverse-transcription product, and is there any antiphage phenotype? On the paper side, persist the patent corpus and read WO2025101943A1's claims rather than its abstract, since it is the closest mechanistic art. Two of the six report authors could not be placed in this field at all, so the search was by name and may miss work published under another. On the lineage, read the sequence listing of WO2024044723A1: it claims 7,257 newly identified retron sequences, and whether any ART reverse transcriptase sits among them is the one check that could still move this record. What this result is not:
Ranked pre-cutoff comparisons:
| 5 yr1 yrCLAIM 23 Sep 202614 Oct 2024Wilkinson et al. — Phage-triggered reverse transcription assembles a toxic repetitive gene from a noncoding RNART, neighboring noncoding RNA, reverse-transcription-linked tandem repeats, phage defense210 Dec 2020Millman et al. — Bacterial Retrons Function In Anti-Phage DefenseRT + noncoding RNA + accessory effector/partner; antiphage context320 Jul 2022Bobonis et al. — Bacterial retrons encode phage-defending tripartite toxin–antitoxin systemsRT-based defense systems with accessory proteins and nucleic-acid-dependent regulation428 Aug 2020Gao et al. — Diverse enzymatic activities mediate antiviral immunity in prokaryotesDefense-associated reverse transcriptases found by genomic context, in a survey of antiviral enzyme families51 Jun 2022Mestre et al. — UG/Abi: a highly diverse family of prokaryotic reverse transcriptases associated with defense functionsA large defense-associated RT family assembled by RT-family mining and accessory genomic context614 Jul 2026Rodríguez Mestre et al. — A DNA-scaffolded retron ring mediates antiphage immunityRetron IX: an RT, an ncRNA, a tandem repeated subunit and an effector assembled into an 8–10-module ring scaffolded by msDNA76 Jun 2024Rodríguez Mestre et al. — Retron-Eco1 assembles NAD+-hydrolyzing filaments that provide immunity against bacteriophagesStructural and mechanistic account of an RT–ncRNA/msDNA–effector system defending against phage823 Oct 2025Rodríguez Mestre et al. — New retron systems from environmental bacteria identify triggers of anti-phage defense and expand tools for genome editingSeven newly described environmental retron systems with two characterised defense triggers — continued discovery of new RT defense systems by the same lineage, eleven months before the claim94 Dec 2020Rodríguez Mestre et al. — Systematic prediction of genes functionally associated with bacterial retrons and classification of the encoded tripartite systemsRT plus contiguous msr/msd ncRNA plus a partner protein, predicted systematically across 13 retron types — the partner-gene logic ART reports1030 Mar 2022Palka et al. — Retron reverse transcriptase termination and phage defense are dependent on host RNase H1RT, structured ncRNA and phage-defense mechanism111 Mar 2018Doron et al. — Systematic discovery of antiphage defense systems in the microbial pangenomeGenome-context mining of uncharacterized defense loci1226 Jun 2024Durrant, Perry, Athukoralage et al. — Bridge RNAs direct programmable recombination of target and donor DNAThree of the six technical-report authors; RNA-guided DNA engineering cited as background1317 Sep 2026Yoon et al. — A noncontiguous code for RNA-guided DNA recognition at the origin of CRISPR-CasOne technical-report author; RNA-guided DNA recognition cited as background | Novelty: new result.1 new New scoped biological system
|
| new resultA proposed resolution of the Navier–Stokes problem A claimed finite-time blowup for Navier–Stokes, with forcing and ordinary viscosity. Evidence & sources
A proposed blowup theorem for ordinary 3D Navier–Stokes with smooth forcing. Unforced singularity formation for smooth 3D Euler was established by 2025, so the separation from prior work is ordinary positive viscosity rather than smoothness or forcing — a real separation that nothing located closes. Validity is unverified and no proof review has happened; this is not an established resolution of the Millennium Prize problem. Why this level: Unforced smooth blowup for 3D Euler was established by 2025, so the separation from prior work is ordinary positive viscosity — a real separation that nothing located closes. Correctness is carried on the validity axis as unverified. Prior-art coverage: A · 87 of 100 Quantitative distance: Qualitative hypothesis gap, not a percentage: claimed smooth finite-energy 3D solution from zero velocity, positive ordinary viscosity, globally smooth compactly supported forcing, and finite-time blowup. The nearest inviscid results already have smooth data and no force, so viscosity is the variable that matters; it is not a demonstrated unforced viscous blowup theorem. Validity: unverified — A claimant report whose formal artifact was not pinned, built or audited. No proof review has happened. Result kind: proof — A claimed theorem about a precisely scoped forced equation, awaiting specialist review. AI contribution: proved — The model produced the argument; its separation from established smooth Euler blowup is viscosity. First-year window: 9 of 365 days observed · window closes 2027-09-08. Only a short post-announcement interval is observable; the first-year window remains open, and the follow-up query returned no relevant mathematics. No independent proof validation was completed here. Patents: Navier–Stokes patents chiefly concern numerical simulation/engineering. No matching mathematical theorem disclosure was verified; numerical fluid solvers do not anticipate a smooth-blowup proof. Review limits: No proof review was performed. Theorem 1.1 has not been specialist-checked against the exact Clay formulation or against the smooth Euler construction, and the formal artifact was not pinned, built or audited. Concurrent results for different equations are not treated as identical, and no claim of copying is made. Still open: Specialist-check theorem 1.1 against the exact Clay formulation, and specifically against the Euler C^∞ construction — whether the viscous case is a modification of that argument or a different one decides how large the step is. Pin, build and audit the formal artifact. What this result is not:
Ranked pre-cutoff comparisons:
| 50 yr20 yr5 yr1 yrCLAIM 8 Sep 2026111 May 2026Córdoba, Martínez-Zoroa & Zheng — Finite time blow-up for the hypodissipative Navier Stokes equations with a force in L¹_t C^(1,ε)_x ∩ L∞_t L²_xFinite-time blowup with forcing in a dissipative fluid model.26 Dec 2021Albritton, Brué & Colombo — Non-uniqueness of Leray solutions of the forced Navier–Stokes equationsForced Navier–Stokes from zero initial velocity; finite-energy weak solutions.39 Apr 2019Elgindi — Finite-time singularity formation for C^(1,α) solutions to the incompressible Euler equationsFinite-time singularity for C^(1,α) solutions of 3D incompressible Euler — a genuine unforced blowup for the true nonlinearity.47 Jul 2025Córdoba, Martínez-Zoroa & Zheng — Finite time singularities to the 3D incompressible Euler equations for solutions in C^∞C^∞ finite-time singularities for 3D incompressible Euler, from the group that then produced the hypodissipative Navier–Stokes blowup.53 Feb 2014Tao — Finite time blowup for an averaged three-dimensional Navier–Stokes equationFinite-time blowup for an averaged 3D Navier–Stokes system.615 Sep 2023Córdoba & Martínez-Zoroa — Blow-up for the incompressible 3D-Euler equations with uniform C^(1,1/2−ε) ∩ L² forceForced 3D Euler blowup.729 Oct 2019Elgindi, Ghoul & Masmoudi — Finite-time singularity formation for strong solutions to the axisymmetric 3D Euler equationsFinite-time singularity for strong axisymmetric 3D Euler solutions.828 Sep 2017Buckmaster & Vicol — Nonuniqueness of weak solutions to the Navier–Stokes equationNonuniqueness constructions for weak Navier–Stokes solutions.91982Caffarelli, Kohn & Nirenberg — Partial regularity of suitable weak solutions of the Navier–Stokes equationsControls possible singular sets for suitable weak solutions.101934Leray — Sur le mouvement d’un liquide visqueux emplissant l’espaceGlobal finite-energy weak solutions. | Novelty: new result.nothing new · 1 unverified A new viscous case, correctness unverified
|
| new resultA claimed improvement in the zeta-zero proportion A claimed jump in how many zeta zeros sit on the critical line. Evidence & sources
A new unconditional zeta-zero proportion bound assembled from earlier analytic and spectral tools. Nothing located reaches the claimed level, and both halves of the claim — simple zeros, and zeros on the critical line — have separate weaker earlier baselines. The score records that the result was not already public. Whether the proof holds is carried separately on the validity axis, where it is unverified; the result does not establish the Riemann hypothesis. Why this level: Scored on what novelty means: nothing located reaches the claimed proportion, and both halves of the claim — simple, and on the critical line — have separate weaker earlier baselines. Whether the proof is correct is a different question, carried on the validity axis as unverified. Withholding the novelty score for an unverified proof double-counts the same doubt. Prior-art coverage: B · 74 of 100 Quantitative distance: The later formal paper claims about 67.25% simple zeros on the critical line. The pre-claim unconditional record is the peer-reviewed bound strictly above 5/12 (41.666…%), so the rounded website 41.6→67.2 comparison is not an exact verified delta. Roughly +25.6 percentage points is descriptive only; simple/on-line denominators must be aligned. Validity: unverified — A claimant report with proof artifacts. The Lean development was not pinned or compiled here, and no independent mathematical review has taken place. Result kind: proof — A claimed theorem awaiting review. AI contribution: proved — The model assembled the argument from named earlier analytic and spectral inputs. First-year window: 38 of 365 days observed · window closes 2027-08-10. Alpöge–Furman arXiv v1 appeared +3 days and identifies Claude’s contribution: same-result publication, not independent corroboration. The one-year window is still open. September revisions of older papers must not be backdated to their 2025 v1. Patents: No substantively overlapping theorem patent identified. The Riemann keyword return is not evidence of prior mathematical proof. Review limits: The Lean artifact was not pinned or compiled, zero-counting conventions were not aligned against both baselines — the on-line bound above 5/12 and the simple-zero lineage — and no independent mathematical review has taken place. This is not a claimed proof of the Riemann hypothesis. Still open: Check exact zero-counting conventions against both baselines — the >5/12 on-line bound and the simple-zero proportion lineage — and align denominators. Pin the Lean commit and compile it, then seek independent mathematical review. What this result is not:
Ranked pre-cutoff comparisons:
| 50 yr20 yr5 yr1 yrCLAIM 10 Aug 202611973Montgomery — The pair correlation of zeros of the zeta functionThe pair-correlation framework the claimed proof’s main input generalises unconditionally.224 Jan 2025Baluyot et al. — Pair Correlation of Zeros of the Riemann Zeta Function I: Proportions of Simple Zeros and Critical ZerosPair correlation used to investigate simple/critical zeros before the claim.37 Jun 2023Baluyot, Goldston, Suriajaya & Turnage-Butterbaugh — An unconditional Montgomery theorem for pair correlation of zeros of the Riemann zeta functionUnconditional Montgomery-type pair-correlation input.414 Feb 2019Aryan — On an extension of the Landau–Gonek formulaUnconditional extension of the Landau–Gonek formula.52000Bombieri — Remarks on Weil’s quadratic functional in the theory of prime numbers, IWeil quadratic form, finite truncations and eigenvalue counts linked to off-line zeros.628 Feb 2018Pratt, Robles, Zaharescu & Zeindler — Combinatorial applications of autocorrelation ratios (later titled More than five-twelfths…)Unconditional lower bound slightly above 5/12.76 Dec 2019Pratt, Robles, Zaharescu & Zeindler — More than five-twelfths of the zeros of ζ are on the critical line (journal version)The peer-reviewed version of the >5/12 unconditional bound, the actual pre-claim record.8May 1998Conrey, Ghosh & Gonek — Simple Zeros of the Riemann Zeta-FunctionProportion of zeros that are simple — the claim is about simple *and* on-line zeros, and this lineage is the simple-zero baseline.92011Bui, Conrey & Young — More than 41% of the zeros of the zeta function are on the critical lineUnconditional improvement past 41%.101989Conrey — More than two fifths of the zeros of the Riemann zeta function are on the critical lineUnconditional improvement past 2/5.13 Aug 2026Alpöge & Furman — More than two thirds of the zeta zeros are simple and on the critical linePublished three days after announcement; cannot count as independent confirmation or prior art. | Novelty: new result.nothing new · 1 unverified A new proportion, correctness unverified
|
| new resultA claimed disproof of the unit-distance conjecture A claimed construction that would break a long-standing conjecture about unit distances. Evidence & sources
A new polynomial lower-bound construction for planar unit distances, using older class-field-tower, class-group-torsion and norm-one ingredients in a configuration none of them supplies. External expert commentary supports the distinction from prior work, and one third-party artifact has recomputed a single finite inequality. That is not a proof review, and validity remains unverified. Why this level: Class-field towers, class-group torsion and norm-one elements are older ingredients, and none of them supplies a planar configuration with a fixed positive exponent. Correctness is carried on the validity axis, where one third-party artifact has recomputed a single finite inequality and nothing more. Prior-art coverage: B · 75 of 100 Quantitative distance: Classical lower-bound scale n^(1+Ω(1/log log n)) → n^(1+ε) for fixed ε>0 along infinitely many sizes, as claimed. This changes the asymptotic regime. No numeric ε or percent runtime/point improvement is assigned. The O(n^(4/3)) upper bound is not the improved baseline. Validity: unverified — A claimant report. One third-party artifact recomputed a single finite inequality and states that it contains no proof; the construction has not been audited. Result kind: proof — A claimed construction and theorem awaiting review. AI contribution: proved — The model produced the construction from older class-field-tower and algebraic-number-theory ingredients. First-year window: 120 of 365 days observed · window closes 2027-05-20. A third-party executable artifact reproduces equation (2.2) of the companion remarks to within 0.014% at 200-bit precision, stating plainly that it contains no new proof, does not construct the infinite Golod–Shafarevich tower, and is not peer-reviewed. A corpus built for the question returned a large, coherent body of class-group-torsion and number-field work — the ingredient layer — and no earlier planar unit-distance construction with a fixed positive exponent. The first-year window remains open. Patents: No relevant planar construction patent was identified. Distance-measurement and unit/geometry keyword patents were excluded. Review limits: The construction, its growing-degree estimates and the dependence of ε were not audited, and the exact Euclidean plane was not separated from higher-dimensional and other-norm variants. The one third-party artifact recomputes a single finite inequality and states that it contains no proof and does not build the infinite tower. Still open: Audit the full construction, growing-degree estimates and dependence of ε; distinguish exact Euclidean plane from higher-dimensional/norm variants. What this result is not:
Ranked pre-cutoff comparisons:
| 50 yr20 yr5 yr1 yrCLAIM 20 May 202612007Ellenberg & Venkatesh — Reflection principles and bounds for class group torsionSmall split primes and class-group pigeonhole methods produce controlled algebraic elements.214 Jan 2019Hajir, Maire & Ramakrishna — Cutting towers of number fieldsClass-field towers with prescribed/split-prime behavior.314 Jul 2025Akhtari, Vaaler & Widmer — Effective equidistribution of norm one elements in CM-fieldsNorm-one elements in CM fields.4May 1946Erdős — On sets of distances of n pointsClassical planar grid lower bound n^(1+Ω(1/log log n)).521 Jul 2025Erdős’s unit distance problem and rigidityUnit-distance problem with rigidity constraints.62015Guth & Katz — On the Erdős distinct distances problem in the planeResolves the companion Erdős distinct-distances problem in the plane.71984Spencer, Szemerédi & Trotter — Unit distances in the Euclidean planeBest classical O(n^(4/3)) upper bound for planar unit distances.822 Jun 2021Voronov, Neopryatnaya & Dergachev — Constructing 5-chromatic unit distance graphs embedded in the Euclidean planeExplicit unit-distance graph constructions in the plane.91986Lenstra — Codes from algebraic number fieldsUses number-field towers/lattices to construct codes.20 May 2026Alon et al. — Remarks on the disproof of the unit distance conjectureIndependent authors’ mathematical checking, but analysis of the same result rather than a separate discovery.2026Flamehaven Labs — erdos-ant-verification: An Executable Verification Artifact for the Sawin et al. Lower Bound on the Erdős Unit-Distance ExponentNecessarily later than the claim but its deposit date was not resolved; explicitly contains no proof and does not build the infinite tower. | Novelty: new result.nothing new · 1 unverified A new construction, correctness unverified
|
| new resultA smaller algorithm for complex matrix multiplication A 4×4 complex matrix product in 48 multiplications instead of 49. Evidence & sources
A new, narrowly scoped bilinear construction: 48 rather than 49 scalar multiplications for two 4×4 complex matrices. Earlier finite-field and commutative algorithms use different assumptions. Subsequent rational constructions build on this result. Novelty does not imply a universal runtime advantage. AlphaTensor is the same laboratory’s earlier system, so the independent comparison for the search idea is the flip-graph work. Why this level: A rank-48 scheme for 4×4 complex matrices where 49 was the task-matched baseline. Nothing located discloses it. Prior-art coverage: A · 97 of 100 Quantitative distance: 49 → 48 scalar products: 1 fewer, 2.04% reduction. This is not a 2.04% runtime guarantee, a new asymptotic exponent, or the first commutative algorithm below 49. Validity: reported — The algorithm and its technical report are public and the identity is checkable in principle; tensor-identity verification was not rerun in this review. Result kind: construction — A rank-48 bilinear scheme: an exact algebraic identity, not a measurement, an artifact or a resource. AI contribution: searched — AlphaEvolve searched a space mathematicians had already formalised, against a known 49-multiplication target. First-year window: First year observed in full to 2026-05-14. Rational/noncommutative rank-48 follow-up at +33 days explicitly builds on AlphaEvolve; a numerical-accuracy improvement and, at +201 days, an analysis of what the Moosbauer–Poole schemes imply, follow within the year. These extend the construction; none predates it. Patents: Strassen/Winograd precision patent is implementation context only. No matching pre-cutoff rank-48 disclosure verified. A retrieved 2026 patent claiming O(n²) was not treated as a validated algorithm. Review limits: Tensor-identity verification has not been rerun here, and coefficient-equivalent decompositions and older algorithm catalogues were not exhausted. Novelty of the construction says nothing about runtime on real hardware. Still open: Run tensor-identity verification; search coefficient-equivalent decompositions and older algorithm catalogues. What this result is not:
Ranked pre-cutoff comparisons:
| 50 yr20 yr5 yr1 yrCLAIM 14 May 202511969Strassen — Gaussian elimination is not optimalExact bilinear matrix multiplication, valid over complex numbers; two recursion levels give 49.25 Oct 2022Fawzi et al. — Discovering faster matrix multiplication algorithms with reinforcement learningExact 4×4 tensor decomposition and a learned algorithm search.32 Dec 2022Kauers & Moosbauer — Flip Graphs for Matrix MultiplicationNon-neural search of exact matrix multiplication schemes.46 Feb 2025Moosbauer & Poole — Flip Graphs with Symmetry and New Matrix Multiplication SchemesSymmetry-constrained search before the announcement.52013Smirnov — The bilinear complexity and practical algorithms for matrix multiplicationSmall-format bilinear algorithm construction.615 Mar 2018Chokaev & Shumkin — Two Bilinear (3 × 3)-Matrix Multiplication Algorithms of Complexity 25Explicit small-format bilinear algorithms found by targeted search.718 Jul 2021Sedoglavic & Smirnov — The tensor rank of 5 × 5 matrices multiplication is bounded by 98 and its border rank by 89Exact noncommutative small-matrix multiplication.816 Apr 2019Rosowski — Fast Commutative Matrix AlgorithmCommutative 4×4 multiplication with a lower operation count.915 Jan 2004US20040010533A1 — Precision improvement method for the Strassen/Winograd matrix multiplication methodStrassen/Winograd implementation and numerical precision.16 Jun 2025Dumas, Pernet & Sedoglavic — A non-commutative algorithm for multiplying 4×4 matrices using 48 non-complex multiplicationsDerives a rational noncommutative version from AlphaEvolve, 33 days later; not independent prior art.19 Mar 2026Dumas, Pernet & Sedoglavic — A more accurate rational non-commutative algorithm for multiplying 4×4 matrices using 48 multiplicationsNumerical-accuracy improvement within one year, not pre-announcement priority.1 Dec 2025Kauers & Wood — Consequences of the Moosbauer–Poole AlgorithmsLater analysis built on the post-announcement schemes; not pre-cutoff disclosure. | Novelty: new result.1 new New scoped construction
|
| new resultLarger cap-set constructions A larger cap set in dimension eight, and a slightly better asymptotic base. Evidence & sources
New cap-set constructions improve established bounds within a known product/admissible-set framework: 512 versus 496 points in dimension eight, and an asymptotic base of 2.2202 versus 2.2180. The advance is a construction, not a new general principle. Why this level: New admissible sets inside known product-cap machinery, and no superseding lower-bound construction in a corpus built to find one. Prior-art coverage: A · 92 of 100 Quantitative distance: Dimension 8: 496 → 512, +16 points (+3.23%). Asymptotic base: 2.2180 → 2.2202, +0.0022 (about 0.0992% in the base, not in the eventual set size). Validity: reported — Peer-reviewed in Nature; the explicit cap and its admissibility certificates were not checked here against the AG(5,3) and AG(6,3) record tables. Result kind: construction — New admissible sets inside existing product-cap machinery. AI contribution: searched — A language model searched programs that generate candidate sets; the framework and the bound to beat came from the existing literature. First-year window: First year observed in full to 2024-12-14. Within the year, the located movement is on the upper-bound side: an exposition of the Kelley–Meka bounds appeared 17 days after the announcement. No independent stronger dimension-eight construction was found in a corpus built for the question — a stronger negative than the first two passes could offer, though still not proof of absence. Patents: Full-table and computing-class searches yielded no substantively matching cap-set construction patent. Keyword collisions such as physical caps and unrelated progression terminology were excluded. Review limits: The explicit cap and its admissibility certificates were not checked against the AG(5,3) and AG(6,3) record tables, and neither the programs nor the asymptotic construction were reproduced. The open item is certificate checking, not further retrieval. Still open: Verify the explicit cap and admissibility certificates against the AG(5,3)/AG(6,3) record table. The seeded crawl is complete; the open item is certificate checking, not more retrieval. What this result is not:
Ranked pre-cutoff comparisons:
| 20 yr5 yr1 yrCLAIM 14 Dec 2023120 Sep 2022Tyrrell — New lower bounds for cap setsAdmissible-set/product-cap construction; asymptotic base 2.2180.22004Edel — Extensions of generalized product capsGeneralized product-cap machinery and prior low-dimensional constructions.3Jul 2001Edel & Bierbrauer — Large caps in small spacesLarge low-dimensional ternary caps.410 Jan 2008Potechin — Maximal caps in AG(6, 3)Determines the maximal cap in AG(6,3); the same finite-dimensional record series the claim extends.5Jul 2002Edel, Ferret, Landjev & Storme — The Classification of the Largest Caps in AG(5, 3)Classifies the largest caps in AG(5,3) — the established low-dimensional record table.61994Calderbank & Fishburn — Maximal three-independent subsets of {0,1,2}^nThree-independent subsets and asymptotic lower bounds.730 May 2016Ellenberg & Gijswijt — On large subsets of F_q^n with no three-term arithmetic progressionThe same cap-set problem over finite fields.85 May 2016Croot, Lev & Pach — Progression-free sets in Z_4^n are exponentially smallPolynomial-method bound for progression-free sets.Jun 2024Pelupessy — Constructions for the cap set problemExpository/experimental thesis, not a verified independent superseding record.2 May 2024Führer & Solymosi — Caps and WicketsA wicket hypergraph bound is not a new size-512 construction or an identical discovery.31 Dec 2023Bloom & Sisask — The Kelley–Meka bounds for sets free of three-term arithmetic progressionsAn upper bound cannot supersede a lower-bound construction; it narrows the corridor from the other side. | Novelty: new result.2 new New constructions in known framework
|
| new resultA 47-multiplication construction over the binary field A 4×4 matrix product in 47 multiplications — in binary arithmetic. Evidence & sources
The 47-product 4×4 construction over F2 improves the prior 49-product baseline and remains a new scoped result. Non-ML searches matched the record soon afterwards, showing that AI was not uniquely necessary to reach it; they are later work, not earlier disclosure. Why this level: A rank-47 scheme over F2 against a 49 baseline. A non-neural search matched it 58 days later, which bears on whether learning was necessary, not on priority. Prior-art coverage: B · 76 of 100 Quantitative distance: 49 → 47 scalar products: 2 fewer, 4.08% reduction in this arithmetic model. No general real/complex rank-47 conclusion. Validity: reported — Peer-reviewed in Nature with the scheme released; this review did not symbolically verify it. Result kind: construction — A rank-47 decomposition over F₂: an exact identity in one arithmetic model. AI contribution: searched — Reinforcement-learning search over tensor decompositions. A non-neural flip-graph search reached the same bound 58 days later, so the method was not uniquely necessary. First-year window: First year observed in full to 2023-10-05. Kauers–Moosbauer responded at +3 days with a 5×5/F2 improvement. Their +58-day Flip Graphs paper explicitly supplies further 4×4/F2 rank-47 schemes. Different discovery method, same later bound. Patents: Located multiplication patents concern implementation/precision; none inspected establishes the exact earlier F2 record. Review limits: The released scheme was not symbolically verified here, and the equivalence classes of the later flip-graph decompositions were not compared. Later constructions reaching the same bound do not change the assessment at the 2022 cutoff. Still open: Symbolically verify the released scheme and compare equivalence classes of the later decompositions. What this result is not:
Ranked pre-cutoff comparisons:
| 50 yr20 yr5 yr1 yrCLAIM 5 Oct 202211969Strassen — Gaussian elimination is not optimalSame exact bilinear task; valid in F2; recursive 4×4 cost 49.22013Smirnov — The bilinear complexity and practical algorithms for matrix multiplicationExact small-format bilinear search.31971Hopcroft & Kerr — On minimizing the number of multiplications necessary for matrix multiplicationBilinear multiplication-count optimization.41986Johnson & McLoughlin — Noncommutative bilinear algorithms for 3 × 3 matrix multiplicationFamilies of exact noncommutative algorithms.515 Mar 2018Chokaev & Shumkin — Two Bilinear (3 × 3)-Matrix Multiplication Algorithms of Complexity 25Explicit small-format bilinear algorithms of reduced complexity.618 Jul 2021Sedoglavic & Smirnov — The tensor rank of 5 × 5 matrices multiplication is bounded by 98 and its border rank by 89New small-format exact construction.716 Apr 2019Rosowski — Fast Commutative Matrix AlgorithmFewer arithmetic multiplications in commutative settings.815 Jan 2004US20040010533A1 — Precision improvement method for the Strassen/Winograd matrix multiplication methodImplementation of known fast matrix algorithms.2 Dec 2022Kauers & Moosbauer — Flip Graphs for Matrix Multiplication58 days later: a different search method reaches the record, not evidence of pre-announcement priority.8 Oct 2022Kauers & Moosbauer — The FBHHRBNRSSSHK-Algorithm for Multiplication in Z₂^(5×5) is still not the end of the story 2Three days later, but 5×5 improvement is not supersession of the benchmark’s 4×4 claim. | Novelty: new result.1 new New finite-field construction
|
| new resultLearning to control tokamak plasma shapes One learned controller holds plasma shapes that hand-built controllers already held. Evidence & sources
An engineering and control-method advance, not discovery of a new plasma phenomenon, and not the first time TCV held multiple plasma configurations: conventional real-time controllers did that, including the snowflake configuration in 2019. A learned closed-loop magnetic controller spanning those configurations is a new scoped result inside the method track. It is scored there and excluded from the discovery ranking. The conventional controllers it is measured against were built by the Swiss Plasma Center team that co-authored the claim, so the comparison is with the collaboration’s own earlier machine capability rather than with a rival result. Why this level: Scored inside the method track, which is what this record is. TCV already held varied configurations and had conventional real-time control of advanced ones, so a learned closed-loop magnetic controller spanning them is a new scoped engineering result — not a new plasma phenomenon, and not eligible for the discovery ranking. Prior-art coverage: B · 71 of 100 Quantitative distance: No single compatible before/after scalar improvement was established, and the located TCV controllers make a like-for-like comparison mandatory rather than optional: same machine, same actuators, same configurations, different synthesis method. Validity: reported — A peer-reviewed physical experiment on TCV. The learned controller was not compared like-for-like against the machine’s conventional controllers on matched discharges. Result kind: controller — An engineering demonstration on hardware, not a scientific result about plasma. AI contribution: learned — Deep reinforcement learning replaced a model-based controller. Conventional controllers held the same configurations before it, including the snowflake configuration in 2019. First-year window: First year observed in full to 2023-02-16. No exact, independently verified replication or supersession within the next year was established by the retrieved set; the follow-up search returned mostly unrelated reinforcement-learning work. Do not read this as an absence of later publications. Patents: RU2702137C1 documents earlier feedback magnetic modelling of plasma shape and current. A US field-shaping patent from 1982 was displaced from the ten closest by the TCV controllers located in this pass; no patent supplies the same deep-RL TCV experiment, and a lexical "plasma shape control" query returned semiconductor etch chambers rather than fusion devices. Review limits: Method-context record, excluded from discovery rankings. Controller novelty is not established here: a like-for-like comparison against the 2017 and 2019 TCV shape controllers on matched discharges, shapes and disturbances, reporting latency and actuator authority, has not been run. Still open: Compare directly against the 2017 and 2019 TCV shape controllers on matched discharges, shapes and disturbances, and report controller latency and actuator authority alongside. What this result is not:
Ranked pre-cutoff comparisons:
| 20 yr5 yr1 yrCLAIM 16 Feb 2022114 Oct 2019Anand et al. — Real time magnetic control of the snowflake plasma configuration in the TCV tokamakReal-time magnetic control of an advanced (snowflake) configuration on TCV itself, 2019.2Dec 1994Hofmann et al. — Creation and control of variably shaped plasmas in TCVCreation and control of variably shaped plasmas on TCV — the machine was built for multiple configurations.3Jan 1995Bishop et al. — Real-Time Control of a Tokamak Plasma Using Neural NetworksNeural-network control of real tokamak plasma existed in 1995.42 Sep 2021Seo et al. — Feedforward beta control in the KSTAR tokamak by deep reinforcement learningDeep reinforcement learning for tokamak beta control.522 Sep 2017Anand et al. — A novel plasma position and shape controller for advanced configuration development on the TCV tokamakA position and shape controller for advanced configuration development on TCV, 2017.62021Pesamosca — Model-based optimization of magnetic control in the TCV tokamak: design and experimentsModel-based magnetic control on TCV with experiments.72019Seo et al. — Safety factor profile control with reduced central solenoid flux consumption during plasma current ramp-up phase using a reinforcement learning techniqueRL for tokamak safety-factor/current-profile control.82005Ambrosino & Albanese — Magnetic control of plasma current, position, and shape in tokamaks: a survey of modeling and control approachesEstablished current, position and shape feedback control.92008Ariola & Pironti — Magnetic Control of Tokamak PlasmasMathematical/controller framework for magnetic tokamak control.104 Oct 2019RU2702137C1 — Method of forming magnetic model of shape and current of plasma with feedback in tokamakFeedback magnetic modelling of plasma shape and current. | Novelty: new result.1 new A new controller, not a new phenomenon — scored in the method track
|
| new resultPredicted structures across the human proteome Predicted structures for nearly every human protein, released at once. Evidence & sources
Four of the closest comparisons are the claimants’ own — AlphaFold2 seven days earlier, AlphaFold1, and two DeepMind filings — so they are part of this claim rather than prior art against it. Measured against what other groups had published, the result is a new scoped one: proteome-scale prediction was an established activity and an independent accurate method appeared the same week, but neither supplied accurate predicted structures for nearly the whole human proteome with calibrated per-residue confidence. Whether individual structures are correct is a separate axis, and deduplication against pre-release structures has not been run. Why this level: The method was published seven days earlier by the same team, and proteome-scale prediction was an established activity. Releasing it across the proteome is an increment in scale within that family, not a second discovery. Prior-art coverage: B · 68 of 100 Quantitative distance: 58% confident predicted residues; 36% very high confidence; full-chain predictions for about 98.5% of human proteins. The paper’s 17% experimentally determined-residue baseline is not the same measurement as 58% predicted confidence. Do not label their difference verified experimental coverage gain. Validity: reported — Peer-reviewed predictions with published confidence measures. No individual structure was experimentally validated here and calibration was not assessed. Result kind: catalogue — A proteome-wide prediction resource, not a set of solved structures. AI contribution: generated — AlphaFold2 produced the predictions. The method and this release are one body of work from one team, published a week apart. First-year window: First year observed in full to 2022-07-22. An independent preprint at +12 days measures human-proteome structural coverage before and after AlphaFold — the first outside quantification of the claim’s own headline comparison. AFDB documentation (+118 days) and ColabFold (+24 days) extend access. None validates individual predicted structures. Patents: Earlier DeepMind protein-structure patent families and a self-attention patent supply method context, not proof that the complete July 2021 catalogue was already public. Review limits: Predictions were not deduplicated against pre-release PDB and modelling resources, confidence calibration was not evaluated, and the independent coverage measurement was not read against the paper’s own 58% and 36% figures. This record aggregates a catalogue and receives no blanket novelty score. Still open: Deduplicate the released structures against pre-release PDB and modelling resources, evaluate confidence calibration independently, and read the independent coverage measurement against the paper’s own 58% and 36% figures. What this result is not:
Ranked pre-cutoff comparisons:
| 5 yr1 yrCLAIM 22 Jul 2021115 Jul 2021Jumper et al. — Highly accurate protein structure prediction with AlphaFoldThe actual AlphaFold2 predictive method was published a week earlier.215 Jul 2021Baek et al. — Accurate prediction of protein structures and interactions using a three-track neural networkAccurate neural structure prediction including interactions.34 Sep 2019Greener et al. — Deep learning extends de novo protein modelling coverage of genomes using iteratively predicted structural constraintsExtending de novo prediction coverage at genome scale.48 Aug 2011Drew et al. — The Proteome Folding Project: proteome-scale prediction of structure and functionProteome-scale structure/function prediction.521 May 2018Waterhouse et al. — SWISS-MODEL: homology modelling of protein structures and complexesLarge-scale homology models and structural coverage.615 Jan 2020Senior et al. — Improved protein structure prediction using potentials from deep learningPrior DeepMind deep-learning protein structure prediction.726 Mar 2020WO2020058174A1 — Machine learning for determining protein structuresEarlier published DeepMind structure-prediction family.83 Jun 2021US20210166779A1 — Protein structure prediction from amino acid sequences using self-attention neural networksSelf-attention for protein prediction.17 Nov 2021Varadi et al. — AlphaFold Protein Structure Database: massively expanding the structural coverage of protein-sequence space with high-accuracy modelsResource documentation, not independent discovery of each structure.15 Aug 2021Mirdita et al. — ColabFold: making protein folding accessible to allUsability/speed contribution, not reversal of catalogue novelty.3 Aug 2021Porta-Pardo et al. — The structural coverage of the human proteome before and after AlphaFoldMeasures the resource rather than validating individual structures; a preprint twelve days later, not prior art. | Novelty: new result.1 new · 1 unverified An accurate method and the proteome released with it
|
| incremental improvementA mixed sodium–lithium solid electrolyte A sodium–lithium chloride electrolyte, screened by AI and then actually made. Evidence & sources
Incremental exploration of an established halide-electrolyte family, with experimental validation of mixed sodium–lithium compositions. Exact composition novelty remains to be mapped against older structures and patent examples. Report measured conductivity and temperature; avoid turning screening scale or reduced lithium stoichiometry into demonstrated battery superiority. Why this level: A specific mixed Na/Li composition, synthesised and measured, inside a halide family whose computational screening was published in 2021. Prior-art coverage: B · 72 of 100 Quantitative distance: For the reported Na2LiYCl6 example at 100°C, 2.2×10⁻⁶ / 6×10⁻⁸ S/cm ≈36.7× versus Na3YCl6, not >100×. Check this arithmetic before repeating the manuscript’s order-of-magnitude wording. Replacing two of three Li sites is 66.7% of those sites, not a measured whole-battery lithium saving. Validity: reported — A preprint with experimental validation at PNNL; conductivity was not normalised across temperature and measurement method. Result kind: substance — A synthesised material with a measured property. AI contribution: searched — AI screening narrowed a candidate space; synthesis and measurement were done by the laboratory partner. First-year window: First year observed in full to 2025-01-08. Within the year, amorphous and nanocrystalline sodium halide electrolytes improved sodium-ion conduction by a different structural route, and the LaCl3 sodium-halide work is another chemistry family. Neither replicates the announced material. The claimant’s WO2025064256A1 appeared 2025-03-27, after the window. Patents: WO2021183944A1 (family of US20230115956A1) and WO2020070955A1 are relevant sodium/halide families. A direct "sodium yttrium chloride" patent query returned only glasses and ceramic powders — lexical retrieval, not evidence of absence. No exact mixed Na/Li embodiment or legal claim coverage was established. Review limits: Polymorph and stoichiometry matches against patents and older crystallography are unresolved, and conductivity was not normalized across temperatures and measurement methods. The conductivity ratio shown here is arithmetic on the reported example, not an independent measurement. Still open: Resolve polymorph and stoichiometry matches in patents and older crystallography; normalize conductivity by temperature and measurement method. What this result is not:
Ranked pre-cutoff comparisons:
| 20 yr5 yr1 yrCLAIM 8 Jan 2024114 Sep 2018Asano et al. — Solid halide electrolytes with high lithium-ion conductivity for application in 4 V class bulk-type all-solid-state batteriesLi3YCl6/Li3YBr6 conducting halide family and battery applications.213 Apr 2020Qie et al. — Yttrium–sodium halides as promising solid-state electrolytes with high ionic conductivity and stability for Na-ion batteriesNa–Y halide end-member chemistry and transport.32021Park et al. — Materials design of sodium chloride solid electrolytes Na3MCl6 for all-solid-state sodium-ion batteriesComputational materials design of exactly the Na3MCl6 sodium chloride electrolyte family for solid-state sodium batteries.423 Feb 2021Wu et al. — A stable cathode-solid electrolyte composite for high-voltage, long-cycle-life solid-state sodium-ion batteriesSodium chloride electrolyte/cathode integration and long cycling.516 Sep 2021WO2021183944A1 — Chlorine-based sodium solid electrolyteNa–Y chloride electrolyte family and substitution.630 Dec 2020Liu et al. — High Ionic Conductivity Achieved in Li3Y(Br3Cl3) Mixed Halide Solid Electrolyte via Promoted Diffusion PathwaysMixed-halide Li3Y(Br3Cl3) with promoted diffusion pathways; mixing on the halide sublattice of the same framework.71992Steiner & Lutz — Neue schnelle Ionenleiter vom Typ M(I)3M(III)Cl6 (M(I)=Li, Na, Ag; M(III)=In, Y)Li/Na/Y chloride ion conductors long predate AI screening.81993Stenzel & Meyer — Ag3−xNaxYCl6: synthesis, structures, ionic conductivityMixed monovalent-cation YCl6 framework and conductivity.926 Jun 2023Tham et al. — Computational design of a mixed A-site cation halide solid electrolyte for all-solid-state lithium batteriesComputational mixing in Li3MX6 halide electrolytes.109 Apr 2020WO2020070955A1 — Halide solid electrolyte material and battery using sameHalide solid-electrolyte materials and battery embodiments.21 May 2024LaCl3-based sodium halide solid electrolytes with high ionic conductivity for all-solid-state batteriesDifferent LaCl3 family, not an independent replication of the announced material.Feb 2024Ridley et al. — Amorphous and nanocrystalline halide solid electrolytes with enhanced sodium-ion conductivityA different structural strategy in the same family; not a replication of the announced composition.27 Mar 2025WO2025064256A1 — Compound for a solid-state battery electrolyte2023 priority is not public disclosure; 2025-03-27 publication is after 2025-01-08. | Novelty: incremental improvement.1 new · 1 unverified Incremental composition/validation
|
| incremental improvementA large catalogue of predicted stable crystals 381,000 crystal structures predicted stable by computation — not made in a lab. Evidence & sources
GNoME expands a computational materials resource using a workflow with substantial earlier paper and patent precedent, including a 2019 patent covering learned stability screening, ab initio validation and iterative retraining. The increment is scale. The reported 381,000 hull entries are computed convex-hull entries, not 381,000 experimentally established discoveries, and whether individual entries duplicate existing databases is still unmatched — that sits on the parts breakdown, not on the score. Why this level: The workflow is disclosed in a 2019 patent and the learned potentials are 2018–2022. What is new is a released catalogue at this scale — an increment in an established family of computational materials resources. This scores the claim as stated. Whether individual entries are new is a separate question that no literature search answers, and it stays on the parts breakdown as unverified. Prior-art coverage: D · 49 of 100 Quantitative distance: 381,000 claimed new computational-hull entries is a catalogue count, not a measured count of novel synthesized materials. No defensible new-material percentage can be computed without structure-level deduplication. Validity: contested — Peer-reviewed, but a within-year perspective challenges the prototype and disorder interpretation, and a separate paper criticises the same-day synthesis companion. Individual entries were never deduplicated against pre-existing databases. Result kind: catalogue — 381,000 computed convex-hull entries: a resource whose claim is scale, which no single result can be ranked against. AI contribution: generated — The model produced candidate structures inside an iterated screening workflow that a 2019 patent already describes. First-year window: First year observed in full to 2024-11-29. The autonomous-synthesis companion appeared the same day and is itself contested. Cheetham–Seshadri’s within-year perspective challenges aspects of novelty and chemical interpretation; Leeman et al. criticize A-Lab. That synthesis audit cannot invalidate every GNoME prediction. Patents: US20190171793A1 (2019-06-06), claims 1–2, is a strong workflow precedent: learned stability screening, ab initio evaluation, iterative database enlargement/retraining, and prototype decoration. US20120330632A1 is older evolutionary crystal-search context. Review limits: No structure-by-structure deduplication against pre-cutoff Materials Project, OQMD, ICSD or COD entries was performed, and symmetry, disorder and prototype handling remain unaddressed. Computational stability is distinct from experimental synthesizability. Still open: Match each released composition/structure against pre-cutoff MP/OQMD/ICSD/COD entries, handling symmetry, disorder and prototypes; separately check synthesis evidence. What this result is not:
Ranked pre-cutoff comparisons:
| 5 yr1 yrCLAIM 29 Nov 202316 Jun 2019US20190171793A1 — Artificial intelligence based stable materials discovery processIterated learned stability screening, ab initio validation, database expansion; claim 2 decorates prototypes.228 Nov 2022Chen & Ong — A universal graph deep learning interatomic potential for the periodic tableLearned graph interatomic potential across the periodic table.310 Apr 2019Chen et al. — Graph Networks as a Universal Machine Learning Framework for Molecules and CrystalsGraph networks for molecules/crystals and energies.46 Apr 2018Xie & Grossman — Crystal Graph Convolutional Neural Networks for an Accurate and Interpretable Prediction of Material PropertiesCrystal graph learning for material properties.518 Jul 2013Jain et al. — The Materials Project: a materials genome approach to accelerating materials innovationDFT structures, energies and convex-hull reference data.611 Dec 2015Kirklin et al. — The Open Quantum Materials Database (OQMD): assessing the accuracy of DFT formation energiesLarge computational formation-energy database.725 Aug 2023Antoniuk et al. — Predicting the synthesizability of crystalline inorganic materials from the data of known materialsLearns synthesizability of inorganic crystals from known-materials data, months before the release.828 Jun 2006Oganov & Glass — Crystal structure prediction using ab initio evolutionary techniques: Principles and applicationsSearch for stable crystal structures.927 Dec 2012US20120330632A1 — Method for predicting optimized crystal structuresEvolutionary optimized crystal-structure search.8 Apr 2024Cheetham & Seshadri — Artificial Intelligence Driving Materials Discovery? Perspective on the Article: Scaling Deep Learning for Materials DiscoveryRaises prototype/disorder concerns; does not prove all 381,000 entries old.7 Mar 2024Leeman et al. — Challenges in High-Throughput Inorganic Materials Prediction and Autonomous SynthesisA-Lab synthesis audit must not be conflated with the entire GNoME prediction catalogue.29 Nov 2023Szymanski et al. — An autonomous laboratory for the accelerated synthesis of inorganic materialsSame-day companion from the collaboration, not independent prior disclosure; its synthesis audit is itself contested. | Novelty: incremental improvement.1 new · 1 unverified A catalogue at new scale, on a precedented workflow
|
| incremental improvementAn AI-designed antifibrotic drug candidate An AI-designed TNIK inhibitor for fibrosis — on a target that was already public. Evidence & sources
The TNIK target and antifibrotic rationale were already public in February 2021, independently of the claimant: TNIK as a Wnt activator dates to 2009, target validation to 2010, an inhibitor series to 2013, and independent work tied TNIK inhibition to fibrosis in 2020. A specific new candidate inside that established family is an increment, which is what the level records. Insilico’s own August 2020 patent is not counted against the claim — it is the same body of work, and it means this chemistry was public six months before the announcement it is dated to. The exact molecule has still not been matched against the Markush claims. Why this level: Target, mechanism, indication rationale and inhibitor chemistry were all public before the announcement, some of it by the claimant six months earlier. A specific new molecule inside that established family is an increment. The molecule's exact identity is still unmatched against the Markush claims, and that stays on the parts breakdown. Prior-art coverage: A · 82 of 100 Quantitative distance: WO2020170202A1 was public 181 days before the benchmark cutoff. No comparable compound-level potency/clinical improvement or percent chemical novelty can be inferred from target overlap alone. Validity: established — The compound exists and has reached a randomized phase 2a trial. What is unresolved is priority, not existence or activity. Result kind: substance — A clinical-stage small molecule. AI contribution: generated — AI selected the target and designed candidates. The TNIK rationale and the chemistry were already public, including in the claimant’s own 2020 patent. First-year window: First year observed in full to 2022-02-24. Buchl et al. appeared online 2021-10-22 (+240 days), within one year, expanding the November 2020 hepatic-fibrosis abstract. The named TNIK-candidate paper (2024) and 2025 trial are outside this window. Patents: Insilico’s WO2020170202A1 discloses TNIK-inhibitor chemistry, fibrosis uses and assays before the announcement. Older third-party TNIK patents also overlap. No verified atom-level mapping of INS018_055 to a patent example is claimed. Review limits: The disclosed rentosertib structure has not been mapped atom-by-atom onto the Markush claims of the claimant’s own 2020 patent or onto the 2013 chemical series, the earliest exact public disclosure is unresolved, and pulmonary and hepatic disease models were not separated. Clinical efficacy is not rated here, and later clinical data cannot establish novelty at the 2021 cutoff. Still open: Map the disclosed rentosertib structure to the Markush claims and to the 2013 chemical series, establish earliest exact public disclosure, and distinguish pulmonary from hepatic disease models. What this result is not:
Ranked pre-cutoff comparisons:
| 5 yr1 yrCLAIM 24 Feb 2021127 Aug 2020WO2020170202A1 — Kinase inhibitors (Insilico Medicine)Public TNIK-inhibitor chemistry, fibrosis uses and collagen/TGF-β assays before the announcement.2Nov 2020Buchl/Maiers et al. — TNIK, procollagen-I trafficking and hepatic fibrogenesis (AASLD poster in 2020 abstracts supplement)TNIK inhibition with NCB-0846 and reduced hepatic fibrogenesis in vivo.325 Jan 2021Torres-Ayuso et al. — TNIK Is a Therapeutic Target in Lung Squamous Cell Carcinoma and Regulates FAK Activation through MerlinTNIK as a therapeutic target in lung, with mechanism, published one month before the announcement.422 Oct 2014Kim et al. — KY-05009 attenuates TGF-β1-mediated epithelial-to-mesenchymal transition in A549 cellsTNIK-directed inhibition of TGF-β-induced EMT in lung-derived cells.58 Oct 2009Mahmoudi et al. — The kinase TNIK is an essential activator of Wnt target genesEstablishes TNIK as an essential activator of Wnt target genes — the mechanistic basis of the whole programme, in 2009.626 Aug 2016Masuda et al. — TNIK inhibition abrogates colorectal cancer stemnessPotent TNIK inhibition and NCB-0846 biology.714 Jun 2010Shitashige et al. — Traf2- and Nck-interacting kinase is essential for Wnt signaling and colorectal cancer growthTNIK essential for Wnt signaling and colorectal cancer growth.8Jan 2013Ho et al. — Discovery of 4-phenyl-2-phenylaminopyridine based TNIK inhibitorsAn earlier published TNIK inhibitor chemical series, eight years before the announcement.925 Apr 2019WO2019079357A1 — TNIK modulators, conjugates, and uses thereofTNIK modulators and therapeutic embodiments.103 Jun 2010US20100137386A1 — TNIK inhibitor and the useTNIK as a druggable target and inhibitor patents.22 Oct 2021Buchl et al. — Traf2 and NCK Interacting Kinase Is a Critical Regulator of Procollagen I Trafficking and Hepatic Fibrogenesis in MiceMarch 2022 issue date would wrongly place it outside the first year; not a rentosertib trial.8 Mar 2024Ren et al. — A small-molecule TNIK inhibitor targets fibrosis in preclinical and clinical modelsCannot retroactively make an undisclosed 2021 target public in 2021.18 Oct 2024Discovery of bis-imidazolecarboxamide derivatives as novel, potent, and selective TNIK inhibitors for the treatment of IPFUseful for exact structure audit; not pre-announcement evidence. | Novelty: incremental improvement.1 new · 1 unverified A new molecule on a target that was already public
|
| incremental improvementGenerated inhibitors of the known DDR1 target Generated DDR1 inhibitors, synthesised and tested — against a well-worked target. Evidence & sources
A generative-design and experimental demonstration in a mature DDR1 inhibitor landscape. Earlier papers and patent families strongly overlap in target, mechanism and chemistry neighborhood. Modest novelty credit is appropriate, but exact molecular novelty and patent coverage require structure-level comparison. Why this level: New generated structures, made and assayed, on a target with potent selective inhibitors from 2013 and co-crystal structures from 2014. Prior-art coverage: C · 64 of 100 Quantitative distance: Earlier Gao inhibitors include reported low-nanomolar activity (7rh/7rj). Different assay protocols prevent a reliable cross-paper potency delta. Synthesis speed and training-corpus size are not molecular novelty measures. Validity: established — Peer-reviewed: compounds were synthesised and assayed, including in vivo. Result kind: substance — Synthesised, assayed molecules. AI contribution: generated — GENTRL proposed structures against a mature target; a contemporaneous analysis found close chemical neighbours in ChEMBL. First-year window: First year observed in full to 2020-09-02. The website cites Bender’s next-day chemical-similarity analysis; that is expert commentary, not a new discovery paper or a legal structure match. An independent deep generative model designing DDR1 inhibitors to a specified pharmacophore appeared online 2021-01-15 — past the one-year window, so it is later context, not within-year corroboration. GENTRL’s reviewed patent A1 is 2021-02-25, also outside the year. Patents: WO2017005583A1 and WO2017137334A1 disclose older DDR1 inhibitor series. Searching the full receptor name rather than the "DDR1" token surfaced two further families the first pass missed, including DDR1-directed inhibitors of fibroproliferative disorders — target and antifibrotic use in one document. The self patent’s 2019-08-23 priority must not be mistaken for a public pre-announcement document. Review limits: Generated structures, training-set nearest neighbours and pre-cutoff patent examples — including two DDR1 families located only on a full-receptor-name search — were not compared for exact identity, scaffold similarity or bioassay equivalence. Cross-paper potency values are not a normalized comparison, and the 75% similarity threshold in the contemporaneous external analysis is not a universal novelty score. Still open: Recover all generated structures, training-set nearest neighbors and pre-cutoff patent examples, including the two newly located families; compare exact identity, scaffold similarity and bioassays. What this result is not:
Ranked pre-cutoff comparisons:
| 5 yr1 yrCLAIM 2 Sep 2019110 Apr 2013Gao et al. — Selective and orally bioavailable DDR1 inhibitors (7rh/7rj)Selective orally bioavailable DDR1 inhibitors; 7rh/7rj reported at 6.8/7.0 nM.213 Aug 2013Kim et al. — Discovery of a potent and selective DDR1 receptor tyrosine kinase inhibitorSelective DDR1 receptor tyrosine kinase inhibitor.3Jun 2014Canning et al. — Structural mechanisms determining inhibition of DDR1 by selective and multi-targeted type II kinase inhibitorsStructures and type-II inhibition of DDR1.424 May 2016Structure-based design of tetrahydroisoquinoline-7-carboxamides as selective DDR1 inhibitorsDesigned selective small-molecule DDR1 chemotypes.58 Jun 2015Fragment-based discovery of potent and selective DDR1/2 inhibitorsFragment-derived potent DDR1/2 inhibitors.619 Nov 2018DNA-encoded library-derived DDR1 inhibitor prevents fibrosis and renal function loss in a genetic mouse model of Alport syndromeDDR1 inhibition with antifibrotic efficacy in a renal disease model.72015WO2015183943A2 — Inhibitors of fibroproliferative disorders and cancerPatented DDR1-directed inhibitors of fibroproliferative disorders and cancer — target and antifibrotic use together.812 Jan 2017WO2017005583A1 — Triaza-spirodecanones as DDR1 inhibitorsPatented small-molecule DDR1 inhibitors.917 Aug 2017WO2017137334A1 — Spiroindolinones as DDR1 inhibitorsIndependent spiroindolinone DDR1 patent family.102016WO2016064970A1 — Small-molecule inhibitors targeting discoidin domain receptor 1 and uses thereofSmall-molecule inhibitors targeting discoidin domain receptor 1 and their uses.25 Feb 2021US20210057050A1 — Workflow for generating compounds with biological activity against a specific biological targetFirst reviewed A1 is 2021-02-25, not its 2019-08-23 priority date; outside the 2019-09-02 to 2020-09-02 window.15 Jan 2021Yoshimori et al. — Design and Synthesis of DDR1 Inhibitors with a Desired Pharmacophore Using Deep Generative ModelsOnline 2021-01-15, past the one-year window; shows the approach was reproduced, not that it was anticipated. | Novelty: incremental improvement.1 new Molecules/workflow around a known target
|
| known resultA formalization of Fermat's Last Theorem Fermat's Last Theorem, formalized in Lean. The theorem is 30 years old. Evidence & sources
No new mathematical theorem: FLT was proved decades earlier. A complete checked formalization can still be a significant software and verification achievement. Theorem novelty stays at the bottom of the scale; the artifact is assessed separately in the method track. Why this level: The theorem is 30 years old and the regular-prime cases were formalized in 2023 and 2024. The artifact may be significant; the theorem is not new. Prior-art coverage: A · 83 of 100 Quantitative distance: Zero change to the theorem’s truth or exponent coverage relative to the established full mathematical proof. Potential formal coverage expands beyond earlier regular-prime Lean results; that is a different metric. Validity: unverified — The theorem is not in doubt — it was proved in 1995. The formal artifact is what is unchecked: not pinned, compiled, or inspected for axioms and admitted statements. Result kind: formalization — A machine-checked artifact of a known proof: a software and verification object, not a mathematical result. AI contribution: formalized — The model translated an existing published proof, following the Darmon–Diamond–Taylor exposition. First-year window: 13 of 365 days observed · window closes 2027-09-04. The first-year window is still open. No independently checked later formalization artifact or replication was established in the short elapsed period. Patents: Theorem-proving patents retrieved are generic software machinery, not earlier public full Lean FLT artifacts or new mathematical proofs. Review limits: Method-context record, excluded from discovery rankings. The Lean artifact was not pinned, compiled or inspected for axioms and admitted statements, and its dependencies were not compared with the earlier regular-prime formalizations. This is acknowledged reuse of a known proof, not a plagiarism allegation. Still open: Pin and compile the claimed full Lean artifact, inspect axioms/admitted statements, and compare dependencies with earlier regular-prime formalizations. What this result is not:
Ranked pre-cutoff comparisons:
| 20 yr5 yr1 yrCLAIM 4 Sep 20261May 1995Wiles — Modular elliptic curves and Fermat’s Last TheoremEstablishes the mathematical theorem through modularity.2May 1995Taylor & Wiles — Ring-theoretic properties of certain Hecke algebrasCompletes the crucial Hecke-algebra/modularity argument.31995Darmon, Diamond & Taylor — Fermat’s Last TheoremExpository proof lineage acknowledged by the formalization.42 Oct 2024Best et al. — A complete formalization of Fermat’s Last Theorem for regular primes in LeanComplete Lean treatment for regular prime exponents.515 May 2023Best et al. — Fermat’s Last Theorem for regular primes (first case, Lean)Lean first case for regular primes.615 Jul 2025Best et al. — A complete formalization of Fermat’s Last Theorem for regular primes in Lean (journal version)Peer-reviewed publication of the complete regular-prime Lean formalization.72021Baanen et al. — A Formalization of Dedekind Domains and Class Groups of Global FieldsFormalised Dedekind domains and class groups of global fields — a load-bearing prerequisite the FLT work stands on.821 Oct 2019The mathlib Community — The Lean mathematical libraryReusable Lean mathematical library. | Novelty: known result.nothing new · 1 unverified A known theorem; the formalization is scored in the method track
|
| known resultA generated study of compositional regularization A model wrote the paper. The finding inside it does not hold up. Evidence & sources
An automated research-process demonstration with no surviving scientific contribution. The regulariser form dates to 2017, and the train/test overlap its own code review flags is a measurement error named and documented in 2022. That is a scored zero, not an absent score: every component was already public and the study establishes no new result. The process demonstration is real, and it is carried on the AI-role axis rather than as novelty. Why this level: Nothing was added to knowledge. The penalty form is 2017, the train/test leakage failure mode was characterised in 2022, and the negative result is confounded by that very leakage. This is a scored zero on novelty, not an unassessed record: the components were all public, and the study establishes no new result to be novel about. The process demonstration is real and is carried on the AI-role axis. Prior-art coverage: A · 80 of 100 Quantitative distance: The claimant's code review reports roughly 57% train/test overlap in a sanity check, and the code penalizes input embeddings while the manuscript describes hidden states. This is a reported audit finding, not an independently rerun leakage measurement — but it is the same failure a 2022 paper had already characterised and told practitioners how to control. Validity: contested — The claimant’s own code review reports roughly 57% train/test overlap, and the code penalises embeddings where the manuscript describes hidden states. The manuscript was withdrawn before final workshop disposition. Result kind: study — An experimental write-up reporting a negative result — neither an artifact nor a resource. AI contribution: authored — The model ran the whole research process: hypothesis, code, experiments and manuscript. First-year window: First year observed in full to 2026-03-12. The AI Scientist-v2 report appeared +29 days as process documentation. The published human/code review is quality evidence, not independent replication of the scientific finding. The manuscript was withdrawn before final workshop disposition. Patents: US20230107460A1 addresses compositional generalization in RL, a different task; it is included as a weak neighbor, not strong anticipation. Review limits: The experiment was not rebuilt: disjoint compositional splits, a corrected representation definition, a comparison against activation regularization, and multiple seeds are all outstanding. The leakage figure is the claimant’s own reported audit finding, not an independently rerun measurement. The manuscript was withdrawn before final workshop disposition. Still open: Rebuild disjoint compositional train/test splits following the published guidance on lexical exposure, fix the representation definition, compare the same regularizer against activation regularization and Yin et al., and rerun with multiple seeds. What this result is not:
Ranked pre-cutoff comparisons:
| 5 yr1 yrCLAIM 12 Mar 202517 Aug 2017Merity, Keskar & Socher — Regularizing and Optimizing LSTM Language ModelsPenalizes changes between successive recurrent activations (temporal activation regularization).23 Aug 2017Merity, McCann & Socher — Revisiting Activation Regularization for Language RNNsThe dedicated activation-regularization paper: penalising the change between successive activations, as its own contribution.3Jul 2023Yin et al. — Consistency regularization training for compositional generalizationExplicit representation/prediction consistency regularization for compositional generalization.431 Oct 2017Lake & Baroni — Still not systematic after all these years: On the compositional skills of sequence-to-sequence recurrent networksSequence-model failure on held-out compositions.52020Kim & Linzen — COGS: A Compositional Generalization Challenge Based on Semantic InterpretationA standard controlled compositional-generalization benchmark predating the study.621 Dec 2022Kim, Linzen & Smolensky — Uncontrolled Lexical Exposure Leads to Overestimation of Compositional Generalization in Neural NetworksShows that uncontrolled lexical overlap between train and test inflates apparent compositional generalization — the exact error the claimant’s code review later found in this study.716 Jun 2020Klinger et al. — A study of compositional generalization in neural modelsControlled compositional-generalization studies.826 Jan 2023Lepori, Serre & Pavlick — Break it down: Evidence for structural compositionality in neural networksStructural compositionality within neural networks.92018Loula, Baroni & Lake — Rearranging the Familiar: Testing Compositional Generalization in Recurrent NetworksSmall controlled tests of compositional generalization in recurrent networks, closer in scale to the tiny arithmetic task than the large benchmarks.106 Apr 2023US20230107460A1 — Compositional generalization for reinforcement learningCompositional generalization in reinforcement learning.12 Mar 2025AI Scientist team — Human review and code review of Compositional RegularizationQuality problem weakens generalization inference; does not prove the narrow negative observation false.10 Apr 2025Yamada et al. — The AI Scientist-v2: Workshop-Level Automated Scientific Discovery via Agentic Tree SearchDoes not establish novelty or validity of the scientific conclusion. | Novelty: known result.1 new · 1 unverified Nothing new was established
|
No claims match this view
Try another search or broaden your filters.