Formal verification of definitions before they run
Assertions that check the data and return counterexamples
Graph queries without a graph database
Compiled SQL, not generated per question
Agents extend the layer under checks
Sovereign by construction, not by contract: on-premises, with a small open model
1. SynaLemma
Verified semantic layer · Synalog language
Rules written in Synalog, a logic language that compiles to SQL. Every rule is formally verified, and the data is checked against assertions before an answer is returned.
Key features
Formal verification: every rule proved well-formed, safe and terminating
Assertions in first-order logic, with counterexamples
Native recursion: graph queries on SQL, no graph database
SQL compiled by the engine, never generated per question
Runs on a small open model, on your infrastructure
Deployment
Sovereign by construction, not by contract: on-premises, on your infrastructure
Delivered with forward-deployed engineers
The analyst writes the Synalog rules for you
Real engine output: a cycle caught in a bill of materials
parts.l
# The bill of materials, read from the database.Component(assembly:,part:,quantity:):-bom(assembly:,part:,quantity:);# A part is inside an assembly directly, or inside one of its parts.@Recursive(Contains,20);@Assert(Contains,transitive:"∀ a b c, Contains a b → Contains b c → Contains a c",no_cycle:"∀ a, ¬ Contains a a");Contains(assembly:,part:)distinct:-Component(assembly:,part:);Contains(assembly:,part:)distinct:-Contains(assembly:,part:sub),Component(assembly:sub,part:);
$ synalog parts.l verify
✓ Contains.transitive holds
✗ Contains.no_cycle is violated: ∀ a, ¬ Contains a a
2 counterexamples:
+---------+
| a |
+---------+
| "wheel" |
| "hub" |
+---------+
2 rows
2. Looker
BI platform · LookML
Google's BI platform. LookML models are compiled to SQL, and Gemini answers questions by picking fields from an Explore.
Key features
LookML model compiled to SQL
Conversational Analytics with Gemini
LookML data tests
Limits for AI agents
No recursion: hierarchies need hand-written SQL derived tables
From each product's public documentation, checked on the review date.
Comparison table
Criterion
SynaLemma
Looker
Cube
dbt Semantic Layer
NVIDIA Auto-Ontology
Apache Ossie
RDF, OWL, SHACL
Definitions formally verified
Yes: proved well-formed, safe, terminating
Syntax and references
Compile check
Configuration checks; the SQL must run
Model calls check the SQL
Schema validation
Reasoners check consistency
Assertions on the data
Yes: first-order logic, any rule, counterexamples
Data tests, 5,000 rows per query
None documented
Data tests on models
None documented
No
Yes, with SHACL, a separate language
Graph questions without a graph database
Native recursion, compiled to SQL
Hand-written SQL
Hand-written SQL
No: two hops at most
Only if the model writes it
No
SPARQL paths, on a triple store
SQL compiled, not generated
Yes
Yes, from fields Gemini picks
The agent writes Cube SQL
For metrics; text-to-SQL also exposed
No: a model writes it
Up to each tool
SPARQL, not SQL
Agents extend the layer, under checks
Yes: every new rule verified
Gemini suggests LookML (preview)
Edits on a dev branch, compile-checked
Copilot generates semantic models
Glossary edits, certified by people
Edited by hand or by tools
Modelled by specialists
Your infrastructure, small open model
Yes
No: Gemini, Looker-hosted
Own model on Enterprise; agent in Cube Cloud
APIs on the dbt platform, paid plan
Yes, self-hosted models optional
Not applicable: a file format
No agent included
FAQ
What is a semantic layer?
A layer between databases and the people or agents who query them. It holds business definitions (revenue, active customer, delivered order) so every query uses the same ones.
What is the best semantic layer for AI agents?
SynaLemma. It is the only product here that meets all six criteria: formal verification, data assertions, graph queries on SQL, compiled SQL, agents extending the layer under checks, and local models.
Does a semantic layer stop AI hallucinations?
Not on its own. The model can still misuse the definitions. A symbolic verifier on the rules and assertions on the data catch the errors.
Do I need a graph database for graph queries?
No. In Synalog, hierarchies and multi-hop chains are recursive rules that compile to SQL on your existing databases.
Can a semantic layer for AI run on a local model?
Yes, if the engine compiles the SQL. SynaLemma runs on a small open model on your own hardware.
See it on your data
Bring a question your current stack can't answer.
In 30 minutes, we show SynaLemma answering it, with the rule behind the number and the checks it passed.
