Varve - AI Risk & Decision Intelligence Platform

How Varve Works

Lineage Archaeology Pipeline

Composing DataHub primitives — Lineage, Ownership, and Governance tags — to prove where your next production outage will originate.

01DataHub Graph Primitive

datahub.lineage.read

Historical Lineage Ingestion

Varve traverses DataHub's historical graph layer by layer — not just current dependencies, but when each feature, threshold, or pipeline step changed.

SELECT event_id, actor, node_urn, timestamp FROM datahub_lineage_events

02Lineage Archaeology

archeology.correlation.engine

Cross-Model & Actor Correlation

Correlates events across un-related models. A threshold shift on Model A made by an engineer who touched Model B before departing is caught automatically.

JOIN lineage_events ON actor_id AND departure_date_window

03Deterministic SQL Logic

precedent.sql.join

Deterministic Precedent Join

Joins lineage changes directly against your organization's real incident history. The reasoning is deterministic SQL; the LLM only formats the sentence.

incidents.root_cause_event_id -> lineage_events.event_id

04Hash-Chained Ledger

datahub.aspect.emit

Idempotent DataHub Write-Back

Every decision lands back onto DataHub's lineage node as idempotent metadata and gets permanently committed to an append-only verification ledger.

datahub.emit_metadata(urn, aspect='varveRiskPattern')

Live Triage Console

Interactive Risk Archaeology Console

Select any production model to inspect its DataHub lineage trail, precedent breakdown, and evidence verification tier.

Ranked Production Models (3)

INSPECTOR TARGET

`customers`

HIGH RISKORG-VALIDATED

FINDING SUMMARY

Departing engineer J. Alvarez altered risk threshold logic mirroring prior incident (c3d4e5f6)

Immediately revert threshold logic in the customers transformation, validate with code review, and reprocess downstream partitions.

DataHub Lineage Archaeology Trail

urn:li:dataset:(urn:li:dataPlatform:dbt,b2fd91.order_entry_db.order_entry.customers,PROD)

Incident Precedents

2 Events

Actor Lineage

J. Alvarez→

Audit Hash

Target Owner: jonny1 (Data Owner)

Zero Faith Required

Append-Only Audit Ledger

Every decision, severity resolution, downgrade, and DataHub write-back is hash-chained. Run verify_ledger.py to prove history has never been altered.

service/scripts/verify_ledger.py

$python service/scripts/verify_ledger.py --db db/varve.db

[+] Walking 28 hash-chained records...

[✓] Block 1043: SHA-256 match (0a8f72... -> b14e59...)

[✓] Block 1044: SHA-256 match (b14e59... -> 5c91a3...)

[✓] Block 1045: SHA-256 match (5c91a3... -> e83a64...)

[Ready] Click verification trigger to walk chain live.

#1001INCIDENT_CONFIRMED

12 mins ago▲

Target: System StateHash: e6cab438...2b23

Previous Block Hash:00000000...0000

Full Block Hash:e6cab438...2b23

Payload Metadata:

{
  "block_id": 1001,
  "action": "INCIDENT_CONFIRMED",
  "target_model": "System State",
  "proof": "SHA-256 Validated",
  "status": "VERIFIED"
}

#1002FINDING_CREATED

10 mins ago▼

Target: Finding 98bd1ace...Hash: e7c1f099...004e

#1003OWNERSHIP_ROUTED

8 mins ago▼

Target: Finding 98bd1ace...Hash: 2f0ec667...069e

Questions & Answers

Frequently Asked Questions

Everything you need to know about Varve's DataHub integration, evidence tiers, and audit ledger.

Varve uses DataHub's native OpenAPI and GraphQL APIs to traverse lineage graphs, entity URNs, ownership metadata, and schema aspects without requiring invasive agent installations.

Tells you which production ML model will break next