CENTARIPre-training

CENTARI

Pre-training

A relational foundation model for the enterprise.

DRI: Matteo Carrabba
Architecture
TRAINING INPUTSCAPABILITIESSchemasZSL declarative IaCActionsZFlow invocationsWorkflowsDurable executionEventLogsOCEL-format tracesMigrationsETL from legacy ERPsCENTARI-2Relational foundation model350B PARAMETERS · SCHEMA-AWARE · MULTI-TURNGreenfield buildTables · actions · formsModify instanceBrownfield changesRoot causeDiagnose failuresData migrationETL from messy sourcesEvaluated by ENT-Bench · 4 task families · L1–L4 difficulty ladder
Training timeline
TRAINING TIMELINE20252026Q1Q2Q3Q4Q1Q2Q3Q4CENTARI-1CENTARI-2Pre-trainFine-tuneInternal AlphaPre-trainFine-tuneRL / EvalsGACheckpoint reviewNOWshippedactiveGA targetmilestone

Relational Foundation Model

Trained on enterprise process data — schemas, actions, workflows, execution logs. Understanding the structure of business operations.

CENTARI-2

Next-generation model in pre-training. Trained on OCEL 2.0 process logs, schema definitions, workflow executions, and migration patterns. Cell-level tokenization preserves relational structure.

CENTARI-1

First-generation model in internal alpha. Greenfield build, instance modification, root cause analysis, data migration.

ENT-Bench

Open benchmark suite for enterprise AI. 4 evaluations, L1–L4 difficulty ladder, robustness variants. 3-layer scoring pipeline.

Reinforcement Learning

Active RL training on data migration tasks. v0.1 achieved 0.79 validation reward. v0.2 targeting expanded task bank with sandboxed execution.

Enterprise Understanding

Not code generation — understanding business processes, data relationships, and operational patterns at a structural level.

More detail on CENTARI-1 coming July 1, 2025.