PUBMETAGet the briefing
Independent research readingFree & open · Est. 2026

Research, under scrutiny.

What was tested. What the evidence supports. Where uncertainty remains.

Explore the library
The research deskPrimary sources. Methods. Limits.
Biology / Lead readPublished brief

Protein watermarks preserve tested binding function—not tamper-proof provenance

SynthIDBio tests sequence and structure watermarks as provenance signals. Tested binders retain function, but modification robustness and restricted detection limit what the signal establishes.

David Stutz et al. · Nature · 2026

Inspect the methods & limitations
Study design
Experimental binder validation and computational watermark evaluation
Evidence base
Experimental binder comparisons across three targets; computational structure tests. Exact total sample sizes and uncertainty intervals are not reproduced in this brief.
The boundary

Watermarks can lose detectability after downstream modifications; robustness and detection trade-offs limit deployment. Both approaches encode watermark presence rather than a detailed provenance history, and detection relies on restricted keys/models. Tested binder function does not establish general functional preservation, biological safety or therapeutic efficacy.

Read the evidence brief Peer-reviewed journal article
Also on the desk Source-linked reading
Biology · 2026Published brief

AI-designed carrier domains work in peptide assembly lines—but context decides success

Laboratory tests show that generated carrier domains can support peptide production, but their success depends on the surrounding assembly line.

Study limitations

Compatibility depends strongly on the neighboring domain: success does not transfer uniformly across architectures. The threefold result is a best-case improvement, not an average across designs; some screens use relative LC–MS peak areas. Simulation-derived mechanisms are hypothesis-generating rather than causally established. No clinical efficacy is tested.

Experimental protein design with computational modelingRead brief
Space · 2026Published brief

A safety filter reins in residual reinforcement learning for simulated spacecraft formations

The authors combine a nominal controller, a structured residual policy, and a recursive online safety filter for coupled translation-and-attitude control. The available abstract reports better tracking and training convergence than pure policy learning and single-step shielding, with near-zero simulated safety violations—but supplies no numerical results or evidence of flight or hardware testing.

Study limitations

Publisher Cloudflare controls blocked direct inspection of the landing page and PDF, leaving equations, proofs, tables, figures, numerical results, and experimental details unavailable. The number of spacecraft, orbit regime, simulator, scenarios, episodes, seeds, uncertainty distributions, comparator implementations, filter-intervention rates, runtimes, ablations, and out-of-distribution tests could not be verified. The abstract reports simulations only—no hardware-in-the-loop, air-bearing, robotic, or on-orbit experiment. Near-zero violations are not violation-free performance, and moderate computation cost does not establish compatibility with an onboard processor or control cycle. No code, data, independent replication, or consolidated artifact was located.

Simulation study of safety-aware residual reinforcement learningRead brief
Know what you’re reading. Published briefs are versioned editorial records. Prototype briefs and landmark references are labeled separately; references link to the original study.

More to investigate.

All research
Computing · 2026Published brief

A three-model pipeline generates and scores synthetic UML diagrams, but its benchmark and validation remain preliminary

The authors connect a small instruction model, a reasoning LLM, PlantUML, and three vision-language evaluators across nine diagram types. They report a 4.23 automated score and correlation with human ratings, but the corpus is synthetic, the principal comparison is only five examples, and several dataset and result counts are difficult to reconcile.

Boundary All requirements were model-generated rather than collected from real projects. Only English prompts were tested. The paper reports no external end-to-end benchmark, repeated runs, decoding parameters, random seeds, uncertainty intervals, or fully documented reasoning ablation. Human-study assignment, blinding, inter-rater reliability, and diagram-level results are not reported. The visible tables total 14,957 records rather than 15,000 and do not support a balanced corpus; failure and success counts are also difficult to reconcile. Public artifacts are fragmented, no consolidated DOI-bound release or executable pipeline repository was verified, and no independent reproduction was found.

Read brief
AI · 2026Published brief

Personal DNA–RNA training improves expression prediction for familiar genes, but not for unseen loci

In GEUVADIS lymphoblastoid cells, xDecoder adapted genomic foundation-model embeddings using paired personal genomes and RNA measurements. Predictions improved across new individuals for genes represented during training, while DNA-only transfer to genes on held-out chromosomes remained near zero. Matched ATAC-seq rescued part of that failure, but an ATAC-only control showed that much of the gain came from measured chromatin state itself.

Boundary The primary personalized cohort was small, the data came from one immortalized lymphoblastoid-cell context, and there was no independent cohort, tissue, or prospective replication. The gene set was enriched for known European cis-eQTLs, the unseen-locus result used one fixed chromosome split, and transfer weakened in YRI individuals. Personal sequences omitted indels and structural variants. Main models used final-epoch selection without a general validation set. The code repository lacks a tagged release, a fully pinned environment, and some final-paper analysis scripts. The results are not suitable for clinical or reliable per-gene use.

Read brief

We read the methods. Not just the headlines.

The PubMeta standard

Loading research library…

03 Fields of inquiry

One method. Many scales.

From molecular structures to planetary systems.

01

AI + COMPUTING

Models, systems, benchmarks, and the infrastructure shaping machine intelligence.

AI · Computing
02

LIFE SCIENCES

Genomes, cells, organisms, and the tools used to observe living systems.

Biology · Medicine
03

PLANET + MATTER

Climate, energy, materials, and measurements across planetary scales.

Climate · Materials
04

FRONTIERS

Observations and instruments extending what we know beyond Earth.

Space

Explore AI × Biology coverage

04 The PubMeta Method

Evidence before
excitement.

Every brief uses the same inspection frame, so readers can see not only what a paper says, but why—and how confidently.

Follow the project
  1. 01

    Start at the source

    We link the primary paper, identify its publication status, and distinguish author claims from our interpretation.

  2. 02

    Inspect the design

    Study type, comparison, sample or data, measurement, and analysis come before conclusions.

  3. 03

    Map the uncertainty

    Limitations, missing evidence, conflicts, and the boundary of generalization stay visible.

  4. 04

    Keep a public record

    Sources, corrections, metadata verification, and meaningful updates remain attached to the brief.