TrustGate
Captured from a real TrustGate + DataHub run

Can I trust order_history for production_ml?

TrustGate investigates DataHub metadata, reconciles conflicting dbt/Snowflake evidence, proposes a fix, waits for human approval, writes back to DataHub, re-fetches, and reassesses the asset.

Asset
order_history
Purpose
production_ml
Sources
dbt · snowflake
Mode
demo (captured run)
Run the assessmentGitHub

Demo Mode uses captured outputs from a real TrustGate run against DataHub OSS. The full MCP investigation and DataHub writeback workflow can be run locally from the public repository.

TrustGate Assessment ConsoleDemo mode
Asset
order_history
Purpose
production_ml
Sources
dbt + snowflake (siblings)
Platform
datahub oss
verdict pending
  1. 01Investigate via MCP
  2. 02Reconcile dbt vs Snowflake
  3. 03Verdict: BLOCKED (PII contradiction)
  4. 04Human approves mutation
  5. 05Writeback to DataHub
  6. 06Re-fetch metadata
  7. 07New verdict: approved_with_conditions

Awaiting investigation

Run it to see the blocked verdict, the proposed fix, the human approval gate, and the DataHub writeback that changes the verdict.

How it works

Why this is an agent that does real work

Reads Context

Uses DataHub MCP to investigate metadata, schema, glossary terms, lineage and sibling assets.

Reasons Over Conflicting Evidence

Builds a canonical profile while preserving source disagreement and missing evidence.

Takes Governance Action

Generates and executes an approved metadata remediation rather than stopping at recommendations.

Contributes Context Back

Writes TrustGate assessment and remediation context back to DataHub so the next human or agent inherits the result.

QuestionCan I use order_history for production ML?
Evidencedbt and Snowflake disagree.
DecisionBlocked.
ActionTrustGate proposes a PII metadata correction.
ControlHuman approves it.
WritebackTrustGate updates DataHub.
VerificationTrustGate re-fetches the metadata.
New decisionApproved with conditions.

Architecture

How the pieces fit together

TrustGate reads the DataHub context graph through the MCP server, builds a canonical profile, applies purpose-aware policy, and — only with human approval — writes back and verifies.

Interfacelocal cli
TrustGate CLI / Demo
Protocolmodel context protocol
DataHub MCP Server

read: search + metadata retrieval · write: approved property mutation

Metadata Sourcedatahub oss
DataHub Context Graph
Schema
Glossary Terms
Tags
Structured Properties
Lineage
Siblings
dbt representation
Snowflake representation
TrustGate Coredeterministic evaluation
Canonical Profile Builder
Purpose-Aware Policy Engine
Assessment Result
Outputartifacts
Reports
JSON + Markdown
Remediation Loophuman gated
Human-approved remediation
DataHub Writeback
Re-fetch
Reassessment

Artifacts

Outputs from the run

Representative outputs stored in the public repository.

Verified Run Artifacts

Before

blocked

  1. Human-approved remediation
  2. DataHub writeback
  3. Live metadata re-fetch

After

approved_with_conditions

View Before AssessmentView After AssessmentView Artifact Guide
Examples overviewexamples/README.mdBefore-remediation assessmentorder_history_production_ml_assessment.jsonBefore-remediation Markdown reportorder_history_production_ml_assessment.mdAfter-remediation assessmentafter_remediation/order_history_production_ml_assessment.jsonAfter-remediation Markdown reportafter_remediation/order_history_production_ml_assessment.mdPost-remediation canonical profileorder_history_live_canonical_profile.jsonRepository examples folderEvery representative output produced by the run.

Testing

Built to be inspectable

41 / 41

tests passing

2

sibling platforms reconciled

5

schema fields evaluated

11 → 10

conflicts after remediation

83.33%

evidence coverage

Required

human approval for mutation

TrustGate separates confirmed facts, assumptions, conflicts and missing evidence rather than fabricating certainty.

Run it yourself

How to run locally

setup
git clone https://github.com/SwavnaSahoo/trustgate-ai.git
cd trustgate-ai
python3 -m venv .venv
source .venv/bin/activate
python -m pip install -e ".[dev]"
export DATAHUB_GMS_TOKEN="<YOUR_TOKEN>"
Live assessment
.venv/bin/python -m src.cli live-assess \
  order_history \
  --purpose production_ml
Preview remediation
.venv/bin/python -m src.cli remediate \
  order_history \
  --purpose production_ml
Human-approved remediation
.venv/bin/python -m src.cli remediate \
  order_history \
  --purpose production_ml \
  --approve

Tech stack

DataHub OSS 1.7.0DataHub MCP ServerPythonPydanticdbtSnowflakepytest