Data Quality Gate

Check lead CSV files before your agent writes to a CRM

Current status — November 1, 2026

Hosted Quality engine quality-1.0.0; active MCP Registry descriptor 1.0.1. Five trials once per verified organization, then $20 plus applicable tax for 100 checks. Separate organization key required. Quality's existing n8n submission awaits owner Terms confirmation and creator eligibility; submission and acceptance are separate.

Agent setup · Product skill · Task-to-tool reference. Skills support Codex and Claude Code and grant no upload, business-meaning approval, spending or destination-write permission. Historical test notes below retain their original scope; synthetic/internal qualification is not paid adoption.

Data Quality Gate checks every row of an original lead, order or inventory CSV/JSONL against explicit contracts you review and approve. The workflow proceeds only on continue, then downloads and verifies the exact checked original before a destination write.

Try five checks free · Integration guide · Reproducible lead guides

Five trial checks once per verified organization, then 100 prepaid checks for $20 plus applicable tax. No quality subscription, automatic overages or replenishment. A supported completed block or warning uses one check. Unsupported or incomplete evaluations and platform failures do not; duplicate retries reuse the original job.

The fixtures here are synthetic, generated and checked by the maintained engine. No customer files, credentials or private execution logs belong in this repository. Internal qualification and synthetic examples do not establish independent customer adoption, customer ROI or comparative superiority. Qualification evidence is bound to the intended installed source before publication.

Reference client

From this examples repository, use Python 3.12 or later and the standard-library client:

export QUALITY_API_KEY=... # from your verified account; keep it private

python client.py --file leads.csv

Review the exact preset before approving and authorizing a check:

python client.py --approve-contract --authorize-check

Resume the original quote/key after an uncertain response:

python client.py --resume --authorize-check

The retry file contains no API key but is private; keep it and customer outputs out of version control. --contract-id reuses an existing approved contract; --baseline-id supplies its reviewed reference. Exit 0 returns the exact checked original; exit 2 stops for constraint/quote review; exit 5 stops for failed/unfinished evaluation. A timeout never authorizes writing. Delete local customer outputs according to your policy.

n8n setup

Import an inactive workflow, then:

1. Verify your organization and create a key. In HTTP Header Auth configure header X-API-Key with that key, named Quality organization X-API-Key. Select it on quality HTTP nodes.
2. In /quality/app, edit/save/review/approve a contract. Baseline recipe also needs an explicitly approved reference and threshold/action.
3. Set Config baseUrl, contractId, baselineId when needed, the actual original fileContent, its format and a fresh stable idempotencyKey. For a binary upstream file, supply its original UTF-8 bytes to Upload; do not substitute the parsed report. Configure the incoming trigger yourself for your account/service.
4. Review the workflow and set operatorConfigured: true. Configure a separate destination credential named Your CRM/database destination and an endpoint accepting the original CSV/JSONL. Set destinationEnabled: true only after destination review. A real CRM adapter may parse the checked file after fingerprint verification; preserve identifiers and decimal strings.
5. Enable n8n's built-in crypto module for the runner (NODE_FUNCTION_ALLOW_BUILTIN=crypto). The SHA256 verification node fails closed if it is unavailable. Never disable fingerprint verification to bypass this setup.

Recipes:

Polling stops after 25 status reads. HTTP errors/timeouts, warnings, blocks, unknown states, incomplete reports and changed original downloads reach Review required. That node stops the execution; it does not send external notifications. Configure reviewer notification separately with your own account/authorization. Only continue reaches a destination automatically.

The destination receives downloaded original file bytes. Validation report JSON is never sent as the dataset. Source retention is 24 hours even though reports last seven days; obtain the original promptly. Keep the same quote/key on retries; new corrected bytes require a new key and quote.

Discovery and publishing

server.json describes the active Registry listing ai.contentsweep/data-quality-gate version 1.0.0 at /v1/quality/mcp/, with a required secret organization key. capabilities.json, llms.txt and ard.json describe bounded implemented capabilities. Registry publication requires actual qualified HTTPS API/worker/MCP and n8n evidence matching the installed source fingerprint, exact descriptor review and owned DNS authentication. The existing active listing was read back on November 1, 2026; changed future descriptors still require reviewed qualification.

Use the creator submission packet after shared creator eligibility and the existing first-template queue are verified. Do not overwrite supplier-feed examples or publish private prospect packets, customer files, API keys or raw n8n execution logs.

scripts/publish-quality-examples.py prepares a separate allowlisted payload for founder-sweep/data-quality-gate-examples without network mutation. --publish requires exact owner-reviewed repository/payload/source hashes and matching deployed HTTPS runtime proof; customer/private/unmanaged material is rejected. It checks remote HEAD even after an uncertain push. scripts/publish-quality-registry.py has a separate exact descriptor/namespace/HTTPS qualification gate. Historical local qualification did not publish these assets; the repository and Registry listing were subsequently published. Changed material needs new exact review and matching release proof.

Tutorial series

Start with lead CSV preflight, a duplicate at record 12,002, and warning review. The owned website then covers missing identifiers, unexpected statuses, exact numeric ranges, approved row-count drift and correction replay. Each guide offers full synthetic inputs, explicit rules, observed output and local reproduction instructions. Publication follows a finite reviewed backlog; exhausted weeks maintain existing material.