Structured-output contract
Strict schemaAction, confidence, target, and reason must all pass validation.
Executed technical proof · simulated provider responses
AI reliability harness / A-001
Put a control layer around the AI step.
This importable n8n workflow proves the checks that keep malformed, low-confidence, or disallowed model output from quietly triggering the next business action.
Control path under test
01AI responseSimulated fixture
02ValidateSchema + allowlists
03Retry onceBounded recovery
04Route or reviewHuman gate preserved
01 / What this proves
Reliability is the logic around the model, not a perfect prompt.
Recovery boundary
One retryA bad first response gets one bounded retry, not an endless loop.
Safe default
Human reviewLow-confidence, malformed, and disallowed output cannot continue automatically.
IMPORTANT LIMIT
This is a control-layer test, not an invented AI client result.
Provider responses are simulated so anyone can import and execute the harness without an API key. A real engagement replaces those fixtures with the approved model node or API while preserving the same validation, retry, routing, and approval contract.
02 / Authoritative execution
Seven checks passed in an isolated n8n run.
TestInputExpected evidenceResult
A-01Valid JSONAccepted on first attemptPASS
A-02Code-fenced JSONNormalized, validated, acceptedPASS
A-03Invalid first responseOne retry; valid retry acceptedPASS
A-04Confidence 0.42Held for human reviewPASS
A-05Disallowed actionBlocked by action allowlistPASS
A-06Malformed responseSafe human-review fallbackPASS
A-07All six fixturesZero external actionsPASS
Executable proof
Import the AI-step reliability harness.
The final report returned all_passed: true: two first-attempt accepts, one accepted retry, three human-review routes, and zero external actions across six response fixtures.
03 / Reusable safety gate
Run the same checks against your own AI output.
Webhook product
AI Output Safety Gate
POST an AI response and a small policy. The workflow returns a deterministic continue or human_review route, validation issues, normalized output, and bounded-retry evidence. It never performs the requested action.
One-time digital purchase. Delivered after Stripe confirms payment. Processed by AIMart.trade; your statement will show AIMART.TRADE.
COMPATIBILITY
Validated before marketplace publication.
The complete workflow converted without warnings and every node passed indie.money's public pre-publish support check. It has not been registered, listed, or represented as earning revenue.
04 / How this becomes client work
Connect only after the decision contract is approved.
Define allowed outputs
Agree on schema, actions, targets, and confidence thresholds.
Attach the model
Use the approved provider, prompt, credentials, and test data.
Test failure paths
Malformed, low-confidence, hostile, timeout, and retry cases.
Release behind approval
Keep consequential actions human-gated until evidence supports more.
Have an unreliable AI step?
Turn this proof into one bounded workflow diagnosis.
Review the exact $49 boundary before payment, or request the free written fit check. No production credentials or client data are needed.