The evidence-first dataset release agent

Know what leaves.
Prove why it can.

LATCH turns DataHub context and real BigQuery evidence into a governed release: unsafe training data stays held, safer alternatives are measured, and every consequential step remains human-controlled.

Start a release review See the release loop
Configured technical evidence—not legal certification.
A concrete release problem

The table looks useful. The boundary changes everything.

A support team wants to fine-tune an external model using customer conversations. The source includes useful prompts and responses—but also governed identifiers, embedded contact details, ineligible records, rare cohorts, and sensitive lineage.

  • Meaning from the graphOwnership, classifications, schema, and column lineage.
  • Proof from the rowsBounded aggregate checks execute inside BigQuery.
  • Authority stays humanNo model can approve, materialize, or publish.
One governed path

A release advances only when its evidence does.

  1. 01Request
  2. 02Understand
  3. 03Test
  4. 04Transform
  5. 05Verify
  6. 06Publish

There is no “AI says safe” shortcut. Proof states come only from executed policy checks and persisted evidence.

DataHub is load-bearing

The catalog controls the transaction—and inherits the result.

Remove DataHub and LATCH cannot compile the release contract. Missing ownership, classifications, MCP capabilities, or required column lineage stops the flow rather than silently weakening it.

Official MCP tool discovery Canonical context snapshot Output properties and lineage writeback
READContext graphOwner · tags · terms · lineage
Release PassportEvidence · approvals · hashes
WRITEGoverned outputProperties · field lineage · receipt
Evidence you can inspect

Every conclusion keeps its technical receipts.

Check resultExpected condition and observed aggregateAppears only after an executed warehouse job
DataHub causePII.DIRECT_IDENTIFIER
Warehouse proofjob ID · query hash · bytes
Allowed repairEMAIL_REDACTION@1

Evidence schema only. Operational values appear only after real execution.

Built for a sensitive boundary

Context, proof, and authority stay separate.

DataHub-grounded context

Required graph context controls policy compilation and remains traceable in every release.

Warehouse-native verification

Raw conversations remain in BigQuery. LATCH persists aggregates, hashes, and job references.

Human-controlled consequences

Gemini explains and compares. Authenticated reviewers approve transformation and publication.

From catalog context to defensible release

When data must cross a boundary, make the proof travel with it.

Start a release review