Claude Found ART. Ten Repeat Searches Missed Its Defining Array.

The most revealing number in Anthropic’s enzyme-discovery paper is ten. That is how many additional search campaigns the authors ran after the original success. None identified the defining ART repeat array in their audit.

That does not make the biology disappear. On September 23, Anthropic announced that Claude agents had helped identify array-associated reverse transcriptases, or ART: a previously uncharacterized natural system with DNA repeats reminiscent of CRISPR. Human scientists then investigated it, including laboratory measurements of RNA produced from the array. Its biological function remains unknown.

Together, those findings make a more useful story than either “AI discovered the next CRISPR” or “the result was a fluke.” A model can notice an important pattern when the evidence reaches it, yet fail to find that evidence reliably during open-ended research. And a genuine observation can survive scrutiny while its proposed mechanism remains unproven.

RohitAI’s read: the engineering opportunity is in the transitions between those states. Research-agent products need to measure what their models actually inspect, preserve interesting observations when explanations fail, and send experimentally useful questions to scientists. More agents and more polished reports cannot substitute for those decisions.

Finding the data, recognizing the pattern, and establishing what it does are three different achievements. ART is a reason to measure all three.

A new arrangement, not a ready-made gene editor

A reverse transcriptase, or RT, is an enzyme class known for copying RNA into DNA. Anthropic describes ART as an RT associated with two other features: a nearby partner gene and a non-coding DNA repeat array. The system occurs mainly in bacteriophages, viruses that infect bacteria. Claude identified a natural arrangement; it did not design a new enzyme from scratch.

The distinction matters for credit, too. The 2021 MarsHill genome paper had already identified the underlying retron-like RT, an unusual region of the protein, and a long upstream non-coding region. Those researchers proposed an associated RNA. The new contribution concerns the repeat architecture and dedicated partner, followed by investigation of the wider family—not the first observation of the underlying sequence.

The ART preprint combines reanalysis of earlier phage-infection measurements with new laboratory work. In the historical dataset, array-derived RNA accounted for up to 8% of phage RNA at one measured time point. New experiments also detected distinct short RNAs from the array. That is evidence of expression, not proof that the RT uses those RNAs to perform a useful operation.

These boundaries come directly from the authors’ discussion. They explicitly leave RT activity, RNA substrates, partner interaction and biological function unresolved. “CRISPR-like” describes an architectural resemblance; it is not a capability specification.

The announcement quotes an encouraging response from CRISPR researcher Feng Zhang. That is expert interest in a preprint, not independent replication. As of this September 23 assessment, the public evidence reviewed does not include an independently replicated ART mechanism or a peer-reviewed ART paper.

Alongside the result, Anthropic introduced its life-sciences research group and Bay Area laboratory. The company says the group formed in spring 2026. September 23 is the announcement date, not a disclosed date for the discovery or the laboratory’s opening.

What happened during those 21.5 hours

The technical report identifies Mythos 5 as the model running the discovery campaign. This is not a discovery attributable to yesterday’s Opus 5.5 release, although Opus 5.5 appears in later recognition tests. Readers following our Opus 5.5 coverage should keep the research model and the newly released product separate.

The agents worked inside a custom orchestration system with distinct worker, supervisor, curator and editor roles. Plans, scripts, results and reviews were shared through a version-controlled record. Follow-up questions could become new tasks. This was organized computational research, not one spectacular answer to a chat prompt.

The search funnel began with roughly 1.939 billion protein clusters and identified 198,290 RT clusters. The agents investigated recurring neighboring protein families, eventually promoting seventeen for deeper analysis. Those are different units: a database cluster, a candidate partner family and a final research report should not be combined into one discovery counter.

Source: campaign methods and accounting. The announcement rounds several of these figures. Its general discussion of hundreds or thousands of candidate reports describes the team’s broader workflows; this campaign finished with nineteen.

Human scientists supplied the research brief and conducted the subsequent laboratory work. Follow-up family analysis also used interactive Claude Science sessions. Anthropic says all laboratory work in this program is performed by people. The 21.5-hour search therefore cannot support a claim that an end-to-end biological discovery was completed autonomously in one day.

Nor can the token total tell us the price of a discovery. It combines uncached input, output and cache writes while omitting cache reads. Preparation, external compute, scientific labor and experiments are separate costs. Multiplying the headline number by one retail token rate would manufacture precision that the report does not provide.

The reruns reveal a gap that a leaderboard can hide

The authors relaunched the campaign ten times with the same harness and brief. Their repeat-search analysis found no rediscovery of the defining array. Its audit looked for identifiers associated with the original campaign, so an ART locus outside that identifier set could escape detection. This is evidence of a repeatability problem under the tested setup, not a universal estimate of discovery probability.

They then asked a different question: what happens when a model receives ART material directly? Seven Claude models ran under five input conditions, with 100 attempts per model and condition. A Mythos 5 judge scored reports against ten author-selected features. These were selected-input recognition and characterization tests, not another genome-wide discovery trial.

The input format changed the result. Given the DNA loci directly in context, the four strongest tested models identified the array in at least 90% of attempts. When files and tools were supplied instead, performance could be much worse: Opus 5 recognized it in 32% of attempts under one condition. That 32% is not a score for every model or for Claude generally.

Across the tool-enabled attempts of those four models, 39% never read a continuous DNA stretch of at least 200 nucleotides. Having a file available did not mean its relevant contents reached the model. Greater DNA exposure correlated with better recognition; the analysis does not establish that forcing a particular read length would causally fix open-ended discovery.

• Search repeatability: Can the system locate an important pattern in a large, unselected collection?

• Conditional recognition: Can it interpret the right material once someone supplies it?

• Biological replication: Do independent experiments support the underlying observation or mechanism?

None of these scores can stand in for the other two. Ten computational misses do not refute measured RNA expression. Strong recognition on selected examples does not establish reliable discovery. A model-judged rubric also needs external expert checks and negative cases before anyone treats it as a general scientific quality measure.

Our earlier analysis of AI-generated mathematics separated correctness, novelty and acceptance. ART adds a specifically experimental wrinkle: scientists can verify that something unusual exists long before they can explain its function.

A tool-call log is not an observation record

Here is the first product implication. Most agent systems can tell you that a retrieval tool ran successfully. Research systems need to tell you what evidence the model actually received. Those are different records.

Imagine two agents with identical folder access. One sees a filename and a summary; the other reads the relevant sequence region in context. A permissions dashboard treats them alike. A discovery evaluation should not. The ART result suggests that evidence exposure belongs beside model version and tool availability when diagnosing failure.

A useful observation record would identify the dataset version, the selected region or artifact, the content delivered to the model, any truncation, and important uninspected regions. This is a proposed engineering control, not something the ART paper demonstrates will solve the problem.

It creates a better debugging question. Instead of immediately escalating to a more expensive model, ask whether the current model ever saw the evidence necessary for the task. Then test a targeted exposure change against the original workflow under the same budget. Do not simply dump more raw data into every prompt: coverage and attention still have costs.

Keep the anomaly when the explanation fails

The discovery trace contains a second, less obvious design lesson. An agent initially pursued a nearby gene as a possible RT partner. It rejected that association as spurious but retained an interesting observation about the RT family and queued further investigation. A later worker inspected related sequences and noticed the array.

That is a richer workflow than “generate hypothesis, ask critic, accept or reject.” A hypothesis can fail while an observation remains worth explaining. If rejection closes the entire case, the system can discard precisely the oddity that would have led somewhere useful.

I would make that distinction explicit in the research record:

Rejected explanation: why the original story failed
Surviving observation: what remains supported
Competing explanations: what could still account for it
Next question: what evidence would distinguish them
Continuation limit: who owns the decision and the budget

This is not permission for endless retries. A surviving observation needs identifiable evidence and a bounded next question. The point is to stop treating rejection as deletion.

The same distinction improves candidate selection for the bench. The most persuasive report is not necessarily the best experiment. I would favor a feasible test that separates competing explanations over a confident narrative whose proposed result would leave every explanation intact. That is a decision rule derived from this case, not a selection strategy Anthropic has shown to be optimal.

The opportunity includes old data and specialist models

ART also makes a case for revisiting public data. The RT was already described, and the infection measurements came from a 2022 study. Some of the analytical opportunity existed before the current agent workflow. Better interpretation can make old collections scientifically productive again without pretending the data appeared when an AI company noticed them.

For research-software builders, that suggests a concrete product: versioned reanalysis of neglected collections, with credit to the original investigators, changed annotations and unresolved observations preserved. The useful output is not another summary of a paper. It is an inspectable difference between what the collection supported before and what a new analysis adds.

A separate Minerva preprint provides relevant context. It uses a specialized genome-language-model approach and reports RNA arrays associated with a different RT system, UG27, including complementary-DNA hairpin products. It is neither independent confirmation of ART nor a head-to-head demonstration that specialized models beat general agents. The authors also provide a public repository.

My architectural bet is a combination: specialist models for systematic sequence screening, general agents for contextual interpretation and follow-up, and experiments to decide which biological claims hold. Test that combination; do not assume it wins. The ART search-versus-recognition gap is precisely why the detector and the interpreter deserve separate evaluation.

A scientific workbench cannot certify its own science

Anthropic’s Claude Science documentation describes a beta desktop workbench for macOS, Windows and Linux, available with Pro, Max, Team and Enterprise accounts. Organizational plans require enablement. The application supplies an analysis environment and versioned artifacts, but a subscription does not recreate the ART team’s indexed database, custom harness or research judgment.

Its reviewer documentation draws an especially important boundary. The reviewer checks whether claims match execution records and artifacts. It does not rerun analyses or decide whether the selected method was appropriate for the research question. A consistent record can still describe an unsuitable analysis.

That gives builders three distinct checks to implement: whether the claimed work actually ran, whether the method answers the question, and whether the scientific conclusion survives an independent test. Another model reading the same record helps with some errors; it does not automatically supply independent evidence.

Research access is separate again. The Life Sciences Verification Program launched September 17 in beta for organizations, initially through first-party API access and Team and Enterprise plans. More permissive biology access is not implied by ordinary app availability. Its published policy requires 30-day retention and says monitored data cannot be used for training or accessed by Anthropic’s life-sciences researchers.

The new in-house laboratory could give Anthropic valuable feedback about which digital reports deserve experimental effort. That is a strategic inference, not evidence that customer research trains its models. For customers, the relevant questions are the actual access grant, project scope and data terms—not whether a product page says “science.”

What I would test before scaling a research agent

A useful pilot should expose where the workflow fails, not merely reproduce its most attractive demonstration. Start with public or appropriately authorized datasets and a domain expert who can assess the method. Then build two evaluation tracks: open-ended search and analysis of fixed evidence. Keep their results separate.

1. Version the complete system. Record the model, prompts, corpus snapshot, tool versions, orchestration rules and review criteria. A model-name comparison is weak evidence if the data or harness changed too.

2. Instrument evidence exposure. Record what source material entered context, not just tool success. Sample failures manually to distinguish missing evidence from misinterpretation.

3. Repeat the whole search. Use a fixed campaign budget and include unsuccessful or terminated runs. Track candidate diversity and redundant rediscoveries, not only the single winning branch.

4. Include known and negative cases. Test whether agents overclaim novelty or invent explanations when no useful pattern is present. A selected positive family cannot tell you the false-positive rate.

5. Separate observation from mechanism. Give each claim its own evidence and review status. Finding a repeat, measuring RNA and establishing enzyme activity must not inherit one shared “verified” label.

6. Reserve capacity for independent checks. Budget expert review, re-execution and experiments alongside inference. Preserve corrected and rejected reports so later readers can reconstruct how conclusions changed.

Then compare three workflow variants: the baseline agent, the same agent with targeted evidence-exposure controls, and a specialist-detector-plus-agent pipeline. Choose the comparison set and evaluation criteria before looking at the outputs. Otherwise a workflow can appear better simply because its reviewer learned to prefer its style.

The scorecard should include independently supported observations, unsupported claims, reviewer hours, experiment suitability and total resources consumed. Report unresolved candidates separately. Scientific work can be valuable before it is fully characterized, but that is not a reason to count uncertainty as completed success.

What would change this assessment

For ART itself, the next meaningful evidence would narrow its mechanism: what the RT actually does, whether the observed RNAs participate, how the partner relates to it, and what the system contributes to the phage. Independent replication would strengthen confidence in the relevant findings. Engineering usefulness is a subsequent question, not an automatic consequence.

For the agent workflow, look for repeatable discoveries across unselected problems, external assessment of novelty and a clearer breakdown of human effort. A better selected-input score alone would not settle the search problem.

My near-term prediction is that serious scientific-agent evaluations will start reporting evidence coverage and whole-campaign repeatability alongside final-report quality. ART supplies a concrete reason: even a capable model cannot recognize a pattern it never meaningfully encounters.

Quick answers

Did Claude discover a replacement for CRISPR?

No. Anthropic reports a natural RT-associated system with an intriguing repeat architecture and RNA-expression evidence. Programmable editing has not been demonstrated, and the system’s function remains unknown.

Does failing to rediscover ART invalidate the biology?

No. Computational search repeatability and experimental replication are different questions. The reruns expose a weakness in the search workflow; they do not erase the sequence observations or measured RNA.

Can a Claude Science subscription reproduce this result?

Not on the evidence available. The application is one part of the workflow. Reproducing the campaign would also require the relevant data, orchestration, model access and evaluation conditions. Buying the workbench is not buying a validated discovery pipeline.

ART is a promising beginning with unusually useful failure data attached. The right response is to take both seriously: pursue the biology, and build research agents whose successes can be traced, challenged and found again.