Weekly Arc: The check that doesn't need a reviewer

Last week's issue closed on a measurement of the reviewer rather than the model. In a randomized trial Anthropic ran, data analysts adopted whichever option its coding agent had been assigned to favor, 86% of the time. Nearly every safeguard proposed for agent-run science ends with a person checking the agent's output, so that number is a measurement of almost every safeguard. This week brought answers, from two directions that do not meet. In software, a check arrived that needs no reviewer at all. In biology, what arrived instead was a paper trail.
Where a check can skip the reviewer
Anthropic reported that Claude formally proved security properties in the compiled code of 123 open-source libraries, turning up bugs that had gone unnoticed for years. Formal verification means a mathematical proof that a program does what its specification says, checked by a machine rather than read by a person. The technique has existed for decades and stayed rare, because writing the proofs took more expert time than most projects could justify. That cost is the thing an agent removes.
The property worth noticing is not the bug count. It is that nothing in the loop waits on a human's judgment at the end. A proof either goes through or it fails, and anyone who doubts the result can run the checker again themselves. Set that against the week's two other security results. Claude found remote-code-execution bugs in browsers and operating systems using an unreleased internal model, and OpenAI built Patch the Planet to route agent-discovered vulnerabilities to maintainers ahead of disclosure. Both of those still terminate in a person deciding whether a finding is real and what to do about it. Only the proof skips that step.
Biology's substitute is a record
Biology has no checker to run. The reference for a claim about a cell is an experiment, and no amount of logging substitutes for performing it. What the field produced this week instead is a record of how a claim was made.
PaleoRigor, a research agent for ancient microbiome data, writes its work out as a reproducible evidence trail. A bioRxiv preprint described the Findings Workflow, which wraps Anthropic's command-line coding agent in a structure that records what each analysis step ran, on which data, and who signed off. LabWeaver ties analysis agents to a lab's own documented knowledge, on the argument that grounding is what makes agent-run analysis trustworthy enough to use.
None of this establishes that a result is correct. It establishes that a second party can reconstruct how the result was produced, which is a weaker claim and the honest ceiling for now. Why the weaker claim still matters is visible in MCseg, which hands the agent the choice of which cell segmentation method to use and how transcripts get assigned to cells. In spatial transcriptomics, that choice largely determines the answer. Once method selection sits inside the agent, the record of what it picked is the only place the decision is visible at all.
The tests moved up to judgment
The other response this week was to raise the thing being graded. rTaste scores whether an agent can tell a promising early-stage research direction from a weak one, before any experiment runs, and separates picking an idea from generating one. BEAKER grades the software that plans and runs experiments in self-driving chemistry labs. FrontierPhysics grades agents on physics research carried from start to finish. MedPhysBench scores two things at once: whether a model gets a medical physics calculation right, and whether it catches a wrong answer put in front of it.
Read together, these say something specific about where the work is going. Judgment was supposed to be the part that stayed with the scientist once execution was handed off. It is now a scored quantity. The caution comes from a bioRxiv preprint arguing that extended reasoning makes autonomous bioinformatics worse, not better. Extended reasoning is the mode in newer models that works through a problem at length before answering, and the expectation has been that more of it buys more reliability. Here it does not. Whatever taste turns out to be, it will have to be built into how the pipeline is structured rather than deliberated toward at the end of it.
The reference that actually binds
The most consequential evaluation of the week was not a benchmark at all. The Wall Street Journal reported that drugmaker chief executives do not accept Silicon Valley's claim that AI will cure disease. The disagreement is about evidence. The pharmaceutical industry measures itself against clinical readouts and approvals, and the AI claims so far rest on scores and internal results.
That is an audit with teeth, and nobody designed it. A benchmark compels nothing; a buyer who declines to pay compels a great deal. Two properties separate a check that holds from one that does not. The first is whether a second party can re-run it, which the formal proofs have and the provenance trails approximate. The second is whether anyone is obliged to act on the result, which only the buyers have. Nothing released this week has both.
So two reference standards are now running side by side. One is built by the labs making the claims and improves every month. The other is held by the people being sold to and moves on the timescale of a trial. The question for the coming stretch is which one the field's arguments get settled against, and whether any of this year's verification work shortens the distance between them.
- Agents as bench scientists: PharmAgents2 carries a program from target to candidate with nobody passing results between steps and El Agente Fármaco scales the same pattern, while DiabAgent puts an agent inside a blood-glucose control loop; watch for the first of these to report a bench or clinical outcome rather than a pipeline.
- The open stack: a red team attacked biosynthesis-planning agents to find where their safeguards give way in the same week Anthropic published five principles for agents acting on their own; the open question is who outside the vendor tests whether the principles hold.
- The infrastructure nobody owns: Anthropic's biology lab reported the first 3D structure of a prokaryotic vault particle, adding a structural biology result to its own publication record; the thing to watch is what happens the first time one of these findings needs independent replication.
Reply with what you're seeing. A human reads them. Forward freely.
|