Contract
Required schemaAction, confidence, target, and reason must all be present and valid.
Eight-lesson course + importable n8n workflow · $19 once
AI Output Safety Gate / OWF-AISG-001
Stop unsafe AI output before the next action.
Put a deterministic validation step between an AI response and the workflow that would act on it. Bad, uncertain, or disallowed output goes to human review instead of quietly continuing.
Open the $19 secure checkout Try the free browser demoInspect the executed proof firstAsk a pre-purchase questionRead the buyer licenseNeed only the contract and checklist? Start at $5
One-time digital purchase. No subscription or included API key. Automatic protected ZIP delivery after Stripe confirms payment. Processed by AIMart.trade; your statement will show AIMART.TRADE.
01 / The control layer
Use AI for judgment without letting malformed output decide.
Policy
Allowlists + thresholdThe caller controls allowed actions, allowed targets, and minimum confidence.
Failure default
Human reviewInvalid JSON, low confidence, or a blocked action cannot continue automatically.
02 / Simple webhook contract
POST the candidate output. Receive a decision signal.
Example requestCaller policy + AI output
{
"candidate_output": {
"action": "route_lead",
"confidence": 0.92,
"target": "sales",
"reason": "High-intent request"
},
"allowed_actions": ["route_lead"],
"allowed_targets": ["sales"],
"minimum_confidence": 0.8
}Decision responseNo action executed
{
"accepted": true,
"route": "continue",
"attempts": 1,
"retry_used": false,
"issues": [],
"external_action_executed": false
}03 / Included evidence
Tested against the failure paths that matter.
CaseExpected routeEvidenceResult
G-01Valid outputcontinuePASS
G-02Low confidencehuman_reviewPASS
G-03Disallowed actionhuman_reviewPASS
G-04Malformed JSONhuman_reviewPASS
G-05Valid retrycontinue after one retryPASS
G-06Every resultzero external actionsPASS
IMPORTANT LIMIT
The workflow returns a route; it does not perform the action.
A separate, explicitly approved n8n step must consume the continue decision and perform any consequential action. This keeps credentials and action nodes outside the downloadable gate.
N8N RUNTIME PROOF
The exact v1.0.1 workflow imports and runs in n8n 2.33.7.
On August 10, 2026, the delivered JSON imported into an isolated credential-free n8n instance, stayed inactive, and round-tripped without a semantic change. Its exact validator Code node passed all 8 fixtures: 3 continued, 5 routed to human review, and zero external actions ran.
Workflow SHA-256: D5CFFD0A1850BB93A60651C93B2AEC916CE3A0A69EF8DE56958400AC2A0E1242. Version 1.0.1 includes the stable workflow ID required by the current n8n CLI import path. This evidence applies to the tested version; it is not a guarantee for untested n8n releases or buyer-specific downstream actions.
04 / Preview the course
Check the syllabus and first lesson before buying.
Eight-lesson syllabus
Approximately 100 minutes
- 01Separate judgment from consequence
- 02Write the action contract
- 03Normalize before validating
- 04Apply policy and confidence
- 05Make human review a real route
- 06Bound retries and replays
- 07Wire the n8n gate
- 08Verify before production
Free sample lessonLesson 1 — Separate judgment from consequence
An AI model is useful because it can make a judgment from messy input: classify a message, propose a destination, summarize a case, or draft a response. The same flexibility makes its output unsuitable as direct authority for a consequential tool.
The central design move in this course is to split the system in two:
- The model proposes a small decision object.
- Deterministic code decides whether that object may continue.
The gate does not perform the proposed action. It returns continue or human_review plus the reasons for that route. A later workflow step can require its own permissions and approval.
This separation limits the blast radius of malformed JSON, unexpected actions, overconfidence, prompt injection, or a model simply misunderstanding the situation. It also makes failures observable. Instead of an action node receiving arbitrary prose, a reviewer receives a structured result with issues such as action is outside the allowlist.
Think of the gate as a turnstile, not a robot manager. A turnstile checks a defined credential and opens or stays closed. It does not decide where the person goes afterward.
Exercise
Choose one workflow you understand. Write down:
- the judgment an AI model could help make;
- the worst reversible mistake;
- the worst difficult-to-reverse mistake;
- the smallest proposed-action object that a deterministic gate could inspect; and
- the human who should review uncertain cases.
Example: an inbound inquiry classifier may propose route_lead to sales, but it should not send an email or modify a CRM record inside the validation step.
Checkpoint
You are ready for Lesson 2 when you can describe the AI output as a proposal rather than an instruction, and when the proposed action remains separate from the tool that could execute it.
The preview publishes one complete lesson. The validator source, fixtures, workflow JSON, remaining lessons, and protected ZIP stay inside the paid package. The exact buyer license is public before checkout.
Get all eight lessons + workflow — $19Review permitted useWhat the purchase includes
A complete implementation course and importable workflow.
Work through eight concise lessons, run the synthetic regression fixtures, inspect the validator source, then import the n8n JSON. The ZIP also includes exercises, checkpoints, disclosures, a buyer-use license, and a SHA-256 integrity manifest.
05 / Before buying
Straight answers about the product boundary.
What does the $19 purchase include?
The one-time purchase includes the eight-lesson implementation course, source code, runnable fixtures, contract tests, and the importable credential-free n8n workflow.
Does it include an AI model or API key?
No. It validates output produced elsewhere, so you keep model choice, prompts, and credentials in your own environment.
Does it automatically send email or update a CRM?
No. It only returns continue or human_review. You decide what an approved downstream step may do.
Can I inspect evidence before buying?
Yes. The public reliability page separates the executed harness from the exact delivered-package manifest, fixture result, workflow node count, and ZIP checksum.
How is the file delivered?
Stripe redirects a completed payment to a protected receipt page. The server validates that exact paid session before releasing the course ZIP. The package contains its own file-hash manifest.
What does the buyer license allow?
You may use and modify the lessons, validator, fixtures, and workflow for your own personal or internal business automations. You may not resell, redistribute, sublicense, publish, or give away the bundle or a substantially similar derivative package. Read the exact Version 1.0 license.
Is this a subscription?
No. It is a one-time digital purchase with no subscription and no included API usage.
A small control before a large consequence
Put a review boundary around the AI step.
Eight lessons, tested code, fixtures, and the importable workflow. One-time purchase with no subscription or included API usage.