DATAHUB AGENT HACKATHON REBUILD

Fix the metadata that makes an agent unsafe to trust.

CAT reads a real DataHub catalog through the official MCP server, traces the impact of a metadata gap, writes only an exactly approved correction, and proves the live graph matches.

See the live evidence Run the proof command

Real local stack: DataHub OSS + mcp-server-datahub@0.6.0. No simulated MCP adapter is used in this proof.

CONTEXT SOURCEOfficial MCPentity · schema · lineage

HUMAN CONTROL2 approvalsone fingerprinted scope each

FAILURE CAUGHT120 / 159partial write rejected

FINAL READBACK116 / 116exact correction verified

LIVE EVIDENCE SNAPSHOT

One dataset. One detected truncation. One bounded correction.

This snapshot was generated from the local DataHub OSS catalog by official MCP reads of entity metadata, schema fields, and upstream lineage.

Read

Get entity + schema fields from DataHub.

urn:li:dataset:(urn:li:dataPlatform:demo,analytics.daily_revenue,PROD)

Verify

The write returned, but live readback exposed a 120-character truncation.

net_revenue · DECIMAL

Trace

Check the immediate upstream impact context.

commerce.raw_orders

Correct

Apply only the separately approved replacement and verify every character.

Correction verified

ENFORCED CAPABILITY GATE

DataHub can ask. CAT still has to prove it can act.

Before CAT proposes any provider action, it checks a small, explicit registry: what is actually connected, what it is allowed to do, whether approval is required, and where the receipt goes.

DataHub resultnet_revenue write truncated at 120 characters→Registry decisionseparate correction approved→Outcomeexact readback verified

claude codenot connected

Anthropic Claude Code

code architecture review

Approval: required before external call
Credential health: not authenticated
Last test: No provider call recorded

authentication required

n8nnot connected

n8n Community Edition

draft local workflow

Approval: required before workflow activation
Credential health: not applicable for local draft
Last test: No provider call recorded

local runtime unreachable

groknot connected

xAI Grok / SuperGrok

independent strategy review

Approval: required before external call
Credential health: not detected
Last test: No provider call recorded

keychain credential not found

agentic commerceisolated inactive

Agentic Commerce Hackathon

No active actions

Approval: not available
Credential health: not applicable
Last test: No provider call recorded

competition work isolated until explicitly activated

LIVE LOCAL APPROVAL LOOP

One decision. You stay in control.

rejected

DATAHUB-BOUND PROPOSAL

Review the bounded metadata remediation.

This browser-only widget demonstrates the approval boundary for a bounded Claude Code review. It does not initiate a provider call.

Capability: claude code
Allowed action: code architecture review
Provider state: Browser demo only

APPEND-ONLY LOCAL RECEIPTS

1Proposed

2Approved

3Safe attempt

{
  "receipt_id": "cap-DATAHUB-REM-001-claude_code-rejected",
  "occurred_at": "2026-07-25T20:55:47.560Z",
  "request_id": "DATAHUB-REM-001",
  "capability_id": "claude_code",
  "action": "code_architecture_review",
  "state": "rejected",
  "reason": "capability_not_connected_or_disabled",
  "approval_id": null,
  "outcome": null,
  "registry_protocol": "cat-capability-registry-v1",
  "external_side_effects": "none"
}

Stored only in this browser. No credentials, provider calls, or external side effects originate from this widget.

The registry is CAT’s source of truth—not a marketing list. A disconnected, disabled, or unregistered provider is rejected before any call. Every proposal, approval, rejection, attempt, failure, and outcome has an append-only receipt; this rendered page never initiates a provider call.

VERIFIED CORRECTION

Short enough for DataHub. Complete enough to trust.

CAT compared the first write with live readback, rejected the partial result, required a second exact approval, and verified the correction.

LIVE 116-CHARACTER DESCRIPTION FOR net_revenue

“USD gross revenue less discounts, refunds, and chargebacks; excludes sales tax, shipping, tips, and processing fees.”

✓ Includes discounts, refunds, and chargebacks
✓ Excludes tax, shipping, tips, and processing fees
✓ Exact readback verified under 8615032d1cca…

REPRODUCIBLE, NOT THEATRICAL

Refresh the evidence against local DataHub.

The command seeds the controlled catalog, calls the official MCP server, then evaluates CAT’s action surface without exposing credentials or calling providers.

1python scripts/seed-datahub-remediation-demo.py

Creates the two demo datasets and their lineage in local DataHub.

2npm run datahub:remediation:live

Runs official MCP reads and regenerates the evidence packet.

3npm run datahub:capability-gate

Proves the registry blocks unverified capability calls and writes the gate receipt.

4npm run build

Builds this site with the refreshed, inspectable proof snapshot.

This snapshot performs read-only MCP calls. CAT prepares a proposed change; it does not write metadata without an explicit human approval.