The Layer Eats the Model

Neutral Layer series

This week Anthropic released Claude Science, and I think it matters more than any model release this year, even though it isn’t a model at all. Claude Science wasn’t introduced as a new model. It’s a specialized working environment built on top of the existing Claude models, one that connects to more than sixty databases across genomics, proteomics and cheminformatics, runs multi-step research pipelines, and records, for every result, exactly how that result was produced.

The most interesting signal here is not that Anthropic went into science. The most interesting signal is that even the company that owns a frontier model doesn’t sell scientists raw intelligence. It sells them an environment where that intelligence gets access to data, tools, action history and verification. That’s the shift in one sentence: value no longer ends at the model’s answer, it begins where the answer becomes checkable.

The lab just confirmed the layer thesis

This entire series has really been about one idea. Value is migrating out of the model and into the layer above it, the callable layer that knows a domain, owns the rubric and takes responsibility for the verdict, while the model itself slowly turns into infrastructure, something like electricity: necessary, interchangeable, invisible.

Until this week, that was an argument I had to make with examples from other people’s markets. Now you can see it in Anthropic’s own product strategy, just by looking at the lineup. Claude Code was built for developers. Claude Cowork handles multi-step knowledge work across documents, files, research and operational tasks. And now Claude Science, for researchers. One model at the bottom, three vertical environments on top, each with its own domain, its own data and its own logic of verification.

When the creator of an asset starts building on top of its own asset, that is the most honest signal you will ever get about where the margin actually lives. Nobody understands the economics of a model better than the people paying to train it, and if they’re moving up the stack themselves, it’s because things are getting crowded down below.

What the layer actually sells

Strip the brand off Claude Science and look only at the architecture, and you’ll find three sources of value. This list is worth memorizing, because it describes the anatomy of any serious decision layer.

The first is domain access: deterministic connections to specialized databases and tools that don’t depend on the model’s probabilistic memory, because infrastructure is accountable for them, not generation. The second is provenance: the code, the environment, the message history and a plain-language explanation of every step are all preserved, because a result without a trace of its origin is worth nothing in science, and Anthropic made that trace part of the product. The third is verification: a separate mechanism that checks citations and calculations before a result becomes the basis for the next decision, which means the layer doesn’t just generate an output, it tests that output against an external standard before handing it to a human.

Now notice what’s missing from this list. Intelligence. At the frontier, the model itself is less and less a sufficient differentiator on its own. Intelligence stays at the bottom, equally available to every customer and every competitor, while the product value moves entirely into the layer: into the data, the trace, the verification.

The Indifference Test passes here too

In an earlier essay I proposed a simple test for neutrality: would the system’s verdict change if you swapped the commercial interests of the parties involved? If it would, you’re not looking at a decision layer. You’re looking at a storefront pretending to be analysis.

Claude Science is built so that it doesn’t care which hypothesis wins. The citation checker doesn’t know whose paper it’s checking, and the reproducibility pipeline doesn’t know whose grant is on the line. That’s precisely what makes the instrument fit for science: the verification mechanism itself doesn’t depend on which hypothesis happens to be commercially or institutionally convenient.

The exact same principle applies in beauty retail, only the stakes look different. A skin analysis with a brand sponsor behind it stops being analysis and becomes advertising with extra steps. Which is why neutrality can’t live in your marketing materials. It has to live in the system’s response itself, in a place where a machine can check it. In SKINBOT’s API response, the fields ranking_basis, sponsored and brand_weighting sit right in the schema, so any integrator, any universal agent can verify them programmatically, with no negotiations and nothing taken on faith. Write-access to the rubric belongs to the logic alone. Brands don’t get in.

What this means for anyone building on models

Looked at soberly, the Claude Science release carries three conclusions for any founder building a product on top of language models.

If your product’s value is best described as “we use AI” you don’t have value, because everyone uses AI, including the lab itself, which will build a vertical layer in your domain faster than you can the moment that domain looks big enough. So the defense isn’t simply being a layer. The defense is being a layer in a domain the lab will never enter. Anthropic will build an environment for science and for code, but it is not going to build a neutral decision layer for beauty retail that stays compliant with GDPR, Russia’s 152-FZ and the Gulf’s PDPL at the same time, reasons at the level of INCI ingredient lists, and carries contractual liability for a recommendation made in a specific store. For a lab, that’s too narrow. For a single company, it’s exactly deep enough to live in.

And the third conclusion, the most interesting one: proof is becoming the interface. Claude Science doesn’t win on being smarter. It wins on showing how a result was produced. The next generation of agentic commerce will select its sub-agents on the same principle, not by how good the demo looks but by how verifiable the verdict is, right there in the response schema, because verdict aggregators need layers they can trust programmatically, with no human in the loop.

Relevance expires. Proof doesn’t

I’ve written before that relevance should expire: a recommendation without a date and a context slowly turns from an asset into a liability. Proof works the other way around. The trace of how a result came to be doesn’t go stale. It stays checkable for as long as the result itself exists.

So the layer thesis can now be stated in full: the model provides intelligence, the layer provides proof, and proof is what gets paid for. This week, Anthropic voted for that formula with its own product

Ekaterina Shalel, SKINBOT