Grounding, Not Geometry

If you build domain-specific AI in 2026, there's a question you should be nervous about and usually aren't: does your embedding still earn its keep against a frontier model? The model has very likely ingested the same public corpus you trained on, plus everything around it, and it reasons across the whole space. It's an uncomfortable question, so most projects quietly avoid it and benchmark their embedding against… a weaker embedding. I wanted to actually ask it, on a domain I'd mapped in detail: olfaction, then flavor.

One clarification up front, because it does a lot of work. This is about entity embeddings — vectors for materials and molecules, learned from how they co-occur — not the text/document embeddings most "do you need a domain model?" papers study. And it's about the case where a structured catalog already exists. If you have no catalog, an embedding may be your only option; that's not the situation under test.

Two things people bundle

"Is my resource useful?" is the wrong question, because it bundles two very different artifacts:

• a structured catalog — names, identifiers, expert descriptor tags, a regulatory layer. You query it symbolically: look it up, filter, intersect. It is a database.
• a learned embedding over the same entities — vector similarity, interpolation, a geometry. You query it by nearest-neighbor. It is a model.

Keep them apart and the real question appears: does the learned geometry add anything over the catalog, for a model that already has both within reach? The clean contrast is a single comparison — same model, same grounding, the only difference being whether it also holds the embedding.

Two rules, so I couldn't fool myself

Pre-registration. I wrote down the primary comparison and the direction I expected before running anything. (I expected the embedding to help a little. It didn't — good that it was on record.)

A contamination ladder. This is the part these evaluations usually get wrong. The embedding, the model, and web search all drank from overlapping wells. If your "ground truth" is mined from the same co-occurrence data the embedding trained on, every arm wins tautologically and you've measured memorization, not capability. I tagged every test set from "circular" to "unmemorizable" and let only the clean tiers carry the conclusion: objective regulatory facts; the field-standard odor dataset; peer-reviewed-adjacent labels. Three frontier models, paired confidence intervals, two domains.

Grounding wins — over the raw model and web search

Ask a frontier model whether a fragrance material is an EU-declarable allergen and it gets 60–85% right. The reason it fails is instructive, because the ground truth moved. For years the declarable set was the famous 26 allergens; in 2023, Regulation 2023/1545 expanded it to 80 — 56 new substances — phasing in through 2026. A model trained across that boundary blends the old list and the new one, confidently, and you can't tell from the answer which one it used. Worse, it will cheerfully recommend materials that are now simply illegal: Lilial (butylphenyl methylpropional) was a workhorse lily-of-the-valley floral until the EU classified it as a reproductive toxicant and banned it from cosmetics in March 2022 — yet ask for a muguet note and a model will still reach for it.

Hand the model a grounded regulatory lookup and recall goes to 100%, for every model, with no false positives. Web search is a flaky substitute — strong for one model, weak for another — not something to lean on when the answer has to be right. It holds beyond facts: on constrained substitution — find five compliant replacements that hold the smell — giving the agent the catalog and regulatory tools more than doubled its compliant, on-target output versus working from memory. The win is the database, surfaced to the model. Not anything learned.

The embedding doesn't earn its keep

So where does the learned geometry land? On the field-standard odor dataset, a trivial structural nearest-neighbor baseline already reaches the neighborhood of the published state-of-the-art, and the frontier models roughly match that. Against that backdrop:

• Standalone, the embedding doesn't beat a non-learned baseline. On held-out odor-descriptor prediction it ties plain structural nearest-neighbors — within a hair, sometimes a step behind. On a separate odor-similarity retrieval task it loses to simple symbolic tag-matching. It ties or it loses; it doesn't win.
• In an agent's hands, it actively hurts. Adding the embedding-similarity tool lowered the substitution agent's compliant-and-relevant output versus the catalog tools alone — consistently, across all three models, every confidence interval excluding zero. The agents called the tool; it just kept handing them things that are used together rather than things that smell alike, and led them off course.

I re-ran the whole shape in a second domain — food, with a peer-reviewed flavor embedding — and got the same line: it loses to flavor-tags standalone, and grounding the model in real flavor data is what lifts it. Not a quirk of one domain.

A note on leakage — calibrated

The most interesting wrinkle, and the one I'd want a careful reader to keep. My first standalone result looked like a clean win for the embedding — until I noticed the lens had trained on the very descriptor links I was holding out to score it. Circular. I rebuilt a descriptor-free version and the "win" inverted into a loss.

But how general is that trap? I checked, with a small self-contained experiment: train the embedding two ways — with the labels in, and with the test labels held out, the way the knowledge-graph field rebuilt its benchmarks (FB15k became FB15k-237) once it found exactly this kind of leakage — and compare. On a dataset where structure already predicts the labels well, the inflation was tiny; the de-leaked embedding just tied the structural baseline. On the usage-based map, where structure wasn't doing all the work, the same class of leak was large enough to flip the conclusion.

So the honest statement is calibrated: evaluation leakage can flip your result, but its size is setup-dependent — large when the embedding carries weak signal of its own, negligible when structure already saturates the task. You don't know which regime you're in beforehand, so you control for it either way.

What I take from it

The claim is deliberately scoped: entity embeddings, chemosensory domains, applied tasks, the case where a catalog exists. It is not "embeddings are useless" — they're the right tool when there's no catalog, and for structural problems like link prediction, where the field rightly favors them. And it's not about text embeddings, which are a different object.

Within that scope, the practical lessons travel:

• Separate the database from the model. Most of your value is probably the grounded catalog — the thing that fixes the model's factual gaps. Ship that first.
• Test the embedding against the actual frontier model, not a weaker embedding — and isolate its marginal value over the catalog the model already has.
• Web search is not a substitute for a curated catalog on facts that must be right.
• Build a contamination ladder before you trust a number. If your gold comes from the corpus your embedding trained on, you're measuring memorization — and the leak's size is unpredictable, so control for it regardless.
• Pre-register the direction, so you can't talk yourself into the answer you wanted.

The one-line version, within scope: the value was in the grounding, not the geometry — the database, not the model. The most useful thing was running the test clean enough to find that out, and keeping the database and the model separate enough to see which one was doing the work.