Google Research unveiled the Science One framework on July 30, 2026, centered on a mechanism called "Chain-of-Evidence" (CoE), which requires every conclusion made by AI in scientific research to be linked to its sources, code, and experimental records.

Article author, source: 0x9999in1, ME News



TL;DR

• Google Research unveiled the Science One framework on July 30, 2026, centered on a mechanism called "Chain-of-Evidence" (CoE), which requires every conclusion made by AI in scientific research to be linked to its sources, code, and experimental records.
• The existing AI research system has long-standing issues: fabricating citations, unreproducible results, and discrepancies between method descriptions and code. The baseline system’s hallucinated citation rate reaches up to 21%, with a reproducibility rate of scores as low as 42%.
• Science One's results are: zero hallucinations in citations, perfect score on verification, and the highest consistency between methods and code. It has been disclosed that all 337 generated citations are authentic, all 12 reproducible experiments match correctly, and 14 out of 15 papers are consistent with the code.
• The key design is "generate and cite simultaneously," not "write first, then add citations." Citations are not sourced from the model’s memory but from real retrievals via the Semantic Scholar API.
• More importantly, Science One did not sacrifice competence for honesty—it earned two gold and two silver medals on MLE-Bench and achieved the best results at the time on Parameter-Golf.
• The true significance of this lies not in another AI writing a paper, but in it being the first to treat "verifiability" as a foundational element of its architecture, not an afterthought.

Let me ask you a tough question: Can you trust a paper written by AI?

In the past two years, AI research systems have made astonishing progress.

It can read academic papers, formulate hypotheses, run experiments, and write an entire paper from start to finish with a complete structure and professional tone. A single glance won’t reveal any flaws.

But the issue lies within this "glance."

A quick glance is fine, but what about digging deeper?

Google Research delved deeply into this issue. They took five leading autonomous research systems, applied them to five system optimization tasks, and generated a total of 75 papers. Then they rigorously questioned each one: Is the cited paper real? Does the reported score hold up when the code is re-run? Does the code actually implement the algorithm you claim to have used?

The results don't look good.

Every baseline system has made at least one systematic error. The rate of hallucinated citations reaches as high as 21%—meaning that one in five citations may point to a paper that does not exist. The pass rate for score verification is as low as 42%. Consistency between methods and code varies between 20% and 80%.

Just think about that 42%. More than half of the papers report results that cannot be reproduced when the code is re-run.

This is not a minor bug. This is a collapse of trust.

Because the fundamental logic of scientific research is "reproducibility" and "traceability." The value of a paper lies not in how smoothly it reads, but in whether others can follow its citations, code, and data step by step to verify that it isn’t lying. AI has maximized the surface quality of papers, but has removed this foundational pillar.

The shinier the surface, the more hidden the problems. That’s where the real danger lies.

Google's answer: Not based on trust, but on blockchain

The idea behind Science One is, in simple terms, very straightforward.

Since you can't trust AI to be honest on its own, don't expect it to be. Put chains on it.

This chain is the "chain of evidence."

Google provided a particularly precise analogy: in the database field, there's a concept called ACID, which doesn't care how you build your database—it only defines the properties a transaction must satisfy to be considered "reliable." The same applies to the chain of evidence. It doesn't dictate how you build a research agent; it only specifies: for every output you generate to be trusted, it must meet certain criteria.

One principle, two halves.

The first half is integrity: every statement in the paper—whether a citation, a number, a description of a method, or a conclusion—must be supported by a documented chain of evidence.

The second half is correctness: the chain must truly support the phrase it's attached to.

In plain terms: everything you say must be backed by a source, and that source must genuinely prove what you’re saying.

How do you implement it? The key is timing.

How does the current system work? First, write a lengthy paper, then go back and add citations and evidence afterward. It’s like crafting a story first and then searching for evidence to support it. Where the evidence doesn’t fit, you just make it up—that’s how hallucinated citations arise.

Science One works in reverse. At the very moment it generates this statement, it simultaneously builds the evidence chain. The claim and the evidence grow together as one, not stitched together afterward.

Swapping this sequence is the essence of the entire matter.

Break it down: three modules, each handling a segment of the lie.

The Science One architecture is divided into three components, each precisely targeting a potential form of fraud.

First block: Issue investigator, responsible for addressing citation fraud.

It does not rely on the model’s “memory” to cite sources. Model memory is a breeding ground for hallucinations—it can “remember” papers that don’t exist. Science One directly integrates with the Semantic Scholar API to construct an accurate citation graph, reading up to 100 full-text PDFs per topic to generate a structured research brief. Every citation in the final paper originates from this real API retrieval, completely eliminating dependence on model memory.

Citations aren't "imagined"—they're "verified." Cutting this thread severs the root of fabricated citations.

The second module, the Discovery Engine, handles falsification of experimental records.

It employs a parallel "explore-exploit" strategy, simultaneously testing solutions across multiple branches. In each round, one solver implements a solution, while a dedicated evaluator assigns a score. High-performing branches are refined iteratively, and all original evaluation outputs are compiled into a strict, read-only record.

Read-only. These two words are crucial. Once the original score is recorded, it cannot be changed. Want to report a better grade when writing your paper later? Sorry, the ledger is here—no changes allowed.

The third role is a paper writer and claim verifier, responsible for ensuring that the methods and code match.

Before rendering the paper, it first structures each factual claim by attaching an inline evidence tag to each sentence, linking it to a specific artifact in the workspace. Then, a dedicated verifier cross-checks each claim against its cited source.

What if it doesn’t match? Here’s a particularly clever detail: instead of simply and crudely deleting it, it’s “retracted”—phrased more conservatively, bringing it back within the bounds supported by the evidence. Where the evidence reaches, that’s where the statement stops.

No exaggeration, and no waste either. This sense of balance is more sophisticated than a one-size-fits-all approach.

Transcript: The Cost of Honesty—Are You Getting Dumber?

At this point, a sharp question should arise.

By putting so many restrictions on AI, forcing it to nervously seek evidence for every statement, doesn’t it become afraid to explore boldly? In the name of “honesty,” are we sacrificing its “capability”?

This concern is very reasonable. It's a classic contradiction between "security" and "capability."

Science One's answer is: No.

First, look at the Honesty column. Under the same audit protocol—Google calls it the CoE Audit, an automated forensic-style reviewer that rechecks every citation against real-time academic databases—Science One leads in all four integrity checks. It has zero hallucinated citations; every citation points to a real, retrievable paper. It achieves perfect score verification and the highest consistency rate between methods and code. According to disclosed data, all 337 citations it generated are real, all 12 reproducible experiment results match exactly, and 14 out of 15 papers align with the actual code.

Compare those flashy names in the baseline—like “hybrid neural-symbolic solvers”—they sound impressive, but once you dig into the code, it’s just a simple deterministic heuristic rule. Names are for show; the code is the truth.

Now let’s look at the “Capable” section.

On the five tasks of the ADRS benchmark, Science One matched or exceeded human expert performance, achieving the best results among all systems in two tasks (Cloudcast and EPLB).

Beyond this circle, it took on even tougher external challenges. In MLE-Bench, it competed in five rigorous Kaggle competitions spanning medical imaging, fine-grained recognition, and 3D perception, earning two gold medals and two silver medals. In the 3D object detection competition, other systems failed entirely, while it achieved a championship-level score. In Parameter-Golf—a real-time LLM training competition with strict constraints on hardware and file size—other systems couldn’t even submit a valid entry, yet it not only met all constraints but also set the best performance as of April 27, 2026. Crucially, it achieved this through genuinely novel algorithmic breakthroughs, not by tuning hyperparameters or guessing.

So the answer is clear: honesty, without making it dumb.

This is precisely the point I value most. If Science One achieved cleanliness at the cost of being overly cautious, it would have little value—because no one would use a clumsy yet honest research tool. But it proves that verifiability and high performance can be achieved together. This transforms the idea of treating honesty as a hard constraint from a moral appeal into an economically sound engineering choice.

My assessment: It's important, but don't rush to mythologize it.

After saying so many good things, it’s time to pour two buckets of cold water.

First pot: Google itself clearly states at the end that Science One is an experimental research prototype and not a tool ready for production use.

This isn’t humility—it’s a boundary. The five tasks it performs fall within domains like system optimization, where there are clear evaluators and automated scoring. In other words, quality can be objectively judged by machines. But what about true frontier research? The value of many hypotheses and the significance of many conclusions simply cannot be scored by any evaluator in the short term. Evidence chains can prevent “fabricated citations,” but they can’t prevent “choosing a mediocre problem.” It enforces honesty—it doesn’t govern taste.

Second point: The evidence chain itself has costs. Reading 100 PDFs per topic, building evidence for every sentence, tagging it, and verifying it again—this process isn’t cheap. While pursuing the truth, it consumes computational power and time. When scaled up, how to account for these costs remains unanswered.

But after pouring out both buckets of water, I still believe this is an underappreciated matter.

Why?

Because it rephrased the question.

Over the past two years, everyone has been asking: Can AI do scientific research? How strong can it be? People have been comparing problem-solving abilities and benchmark scores. Science One raises a different question: Can the research produced by AI be trusted?

These two issues are not on the same track.

As more systems become capable of producing elegant solutions, problem-solving ability alone is no longer a differentiator. At that point, what truly sets them apart is whether their outputs can be trusted. Google has elevated "verifiability" to the status of a first-class citizen, asserting that it must be an architectural constraint, not an afterthought—I believe this assessment is correct and timely.

Because trust is built slowly but can collapse quickly. If AI research, in its early stages, exhausts the entire academic community’s trust by relying on a flood of papers that appear impressive but are riddled with flaws, the cost of regaining that trust later will be unbearable. Science One is reinforcing the steel rebar in the foundation before the structure collapses.

Epilogue

Back to the original question: Can you believe an AI-written paper?

Before Science One, the honest answer was—don’t be too confident. The better it’s written, the more cautious you should be.

Science One didn’t make AI smarter. Instead, it did something more fundamental and more difficult: it enabled AI to leave verifiable evidence for every statement it makes.

There are already enough smart AIs.

AI that takes responsibility for itself has just begun.

This step is small—it’s just a prototype, and there are still many unresolved issues. But we’re on the right track. Sometimes, the true significance of something isn’t in how fast it runs today, but in where it sets the benchmark.

Science One shifted the benchmark from "whether it looks like it was written well" to "whether it can withstand scrutiny."

This move is worth it.

Reference materials

1. Google Research Blog. "Science One Framework: A Verifiable Autonomous Research Framework via Chain-of-Evidence," July 30, 2026.
2. Google Cloud AI Research. arXiv:2605.26340 "Science One Framework / Chain-of-Evidence" paper.
3. AlphaSignal. "Google's Science One Framework Fixes AI Research's Fake Citation Crisis," July 2026.
4. Zhiding Network. "The Science One Framework: A Verifiable Autonomous Scientific Research System Based on Evidence Chains," July 31, 2026.
5. Sakana AI. arXiv:2504.08066 "The AI Scientist v2" (baseline comparison).
6. OpenAI. GitHub project "Parameter-Golf" benchmark.
7. arXiv:2410.07095 MLE-Bench benchmark paper.
8. arXiv:2510.06189 "Automated Design of Research Systems (ADRS)" benchmark.