The Empirical Layer
Empirical layer
unknowns, discovered dynamicallyA computation nobody named in advance — verified once it's found, then kept.
Semantic layer / ontologies
known knownsThe metrics and relationships your org already agreed on — dbt, Cube, LookML, a real ontology.
Your data
Warehouse tables, files — whatever you bind.
empirical layer · noun
A governed, versioned record of what has actually been computed and verified about your data — built once, reused by every agent after.
Coined here. As far as we can tell, this is the first one — the Open Derivation Spec is the proof, not just the claim.
Why “empirical”? Because nothing in it is asserted — only what's been checked against real data, by a real person, survives.
§ 01 · Why agents need one
Ask an agent the same question twice against raw tables and it re-derives the join, the threshold, the definition of “active” — a little differently each time. Nobody checked the number before your application started trusting it, and nothing it worked out survives the session: the next agent, yours tomorrow or a teammate's this afternoon, starts from zero and pays to rediscover the same thing. Skip the warehouse and paste the dataset into a prompt instead, and it fails outright — no context window fits real data.
A derivation fixes the actual problem: a Python function, computed once, reviewed once, then kept. Stack enough of them up and you have an empirical layer — a growing, versioned model of your data that a person or an agent can add to, so nobody has to guess twice.
§ 02 · Not a replacement for your semantic layer
Semantic layers and ontologies — dbt, Cube, LookML, a real ontology — are essential, and elbi isn't trying to replace them. They're built for known knowns: the metrics and relationships your org has already named and agreed on. But an agent doesn't only ask about what's already named — it proposes things nobody defined in advance: a one-off score, a novel segment, a derivation nobody wrote down yet. Those are unknowns, discovered dynamically, and a semantic layer was never built to verify a computation it never anticipated. That's the empirical layer's job.
That's the stack up top. What the unknown looks like once an agent — or a person — actually finds it:
Raw access, every prompt
Every prompt re-scans the rows and re-runs the join, whether the answer changed or not. Nothing is cached.
Result
Recomputed every time
because nothing carries forward.
Any agent with read access to your database
The elbi pattern
The computation runs once, under review. Every later call reuses the cached result — already signed off.
Result
Cached once
so every later call is instant.
Data contracts · certificates · result history · lineage
§ 03 · Portable, not platform-locked
Every major data platform is shipping some version of this now: a governed model that gives an agent meaning instead of raw rows. They're also making sure it only runs on their stack. That's not an accident.
elbi is the portable version. Every certified derivation — yours or an agent's — adds to the same growing model, and it runs wherever you do. No platform to buy into first.
Stop explaining this in Slack threads
Or just install it and let the certificate do the talking.