Trust & Verification

AI-generated designs require human verification before manufacturing. This chapter establishes a Human-in-the-Loop (HITL) protocol for working safely with AI assistance.

Manufacturing Standards

AI output is a starting point, not a finished product. All designs must be verified against your company's quality standards before release to production.

---

The Trust Spectrum

Not all AI outputs require the same level of scrutiny:

Output Type	Trust Level	Verification Required
Documentation questions	High	None - informational only
Design suggestions	Medium	Review for applicability
Generated scripts	Low	Review before running
Bulk modifications	Very Low	Verify each change

Rule of thumb: The more elements affected, the more verification needed.

---

HITL Protocol: Three-Step Verification

Step 1: Review Before Running

When AI generates a script:

1. Read the code - Understand what it will do
2. Check quantities - Verify counts and ranges
3. Confirm targets - Are the right elements affected?

Example: AI wants to update all wires to 18AWG

Before running, ask yourself: Do ALL wires really need to be 18AWG? What about power wires that need 14AWG?

Step 2: Verify After Running

After executing a script:

1. Visual inspection - Does the design look correct?
2. Spot check properties - Select random elements, verify values
3. Run validation - Ask AI to check for issues

"Check for any wires with incorrect gauge for their current rating"

Step 3: Manufacturing Review

Before releasing to production:

1. BOM verification - Cross-check part numbers with suppliers
2. Wire list audit - Verify all connections match requirements
3. Drawing review - Ensure annotations are accurate
4. Sign-off - Document who verified what

---

High-Risk Operations

Be especially careful with these AI-assisted tasks:

Bulk Updates

When AI modifies multiple elements at once, be especially careful:

// CAUTION: This script updates ALL wires to 22AWG
const allWires = runScriptCommand("queryWire", {});
allWires.elements.forEach(wire => {
  runScriptCommand("updateWire", {
    query: wire.name,
    update: { coreSizeAWG: "22" },
  });
});

Safer approach: Query first, review the list, then target specific elements:

// Better: Only update signal wires (filter by name pattern)
const allWires = runScriptCommand("queryWire", {});
const signalWires = allWires.elements.filter(w => w.name.startsWith("SIG"));

// Review what will be changed
scriptConsole.log(`Will update ${signalWires.length} signal wires`);
scriptConsole.table(
  signalWires.map(w => ({ name: w.name, current: w.coreSizeAWG }))
);

// Only then update each one
signalWires.forEach(wire => {
  runScriptCommand("updateWire", {
    query: wire.name,
    update: { coreSizeAWG: "22" },
  });
});

Deletions

AI erasure commands are immediate:

// WARNING: No undo within script
runScriptCommand("eraseComponent", { query: "J101" });

Mitigation:

• Save project before running erase scripts
• Use Ctrl+Z immediately if wrong elements deleted
• Ask AI to list what will be deleted before erasing

Auto-Generated Components

AI can create components from descriptions, but:

• Pin counts may be assumed incorrectly
• Part numbers should be verified against datasheets
• Electrical specifications need validation

---

Verification Checklists

Before Releasing BOM

Check	Method
Part numbers exist	Search DigiKey/Mouser
Specifications match	Compare to datasheets
Quantities correct	Cross-reference design
Lead times acceptable	Check supplier availability

Before Releasing Wire List

Check	Method
Wire gauges appropriate	Calculate current capacity
Colors per spec	Verify against customer requirements
Lengths include service loops	Add margin for assembly
Terminal crimp specs correct	Match wire gauge to terminal

Before Releasing Drawing

Check	Method
All connections shown	Compare to wire list
Dimensions accurate	Verify scale and measurements
Labels readable	Check font sizes
Revision marked	Include date and author

---

AI Confidence Indicators

When asking the AI about your design, notice hedging language:

AI Says	Meaning	Action
"This will..."	High confidence	Verify anyway
"This should..."	Medium confidence	Definitely verify
"I think..."	Low confidence	Research independently
"I'm not sure..."	Very low confidence	Don't proceed without confirmation

---

Audit Trail

Maintain records of AI-assisted design decisions:

What to Document

1. Prompts used - What did you ask the AI?
2. Scripts executed - Save generated code
3. Verification performed - What did you check?
4. Changes made - What did AI modify?

How to Document

Option 1: Chat history export

• Save chat transcripts for project records

Option 2: Script library

• Save verified scripts with descriptions
• Note what verification was performed

Option 3: Project comments

• Add notes to design elements about AI involvement

---

Recovery Procedures

If AI Made Wrong Changes

1. Immediate: Ctrl+Z to undo
2. Recent: Use undo history menu for multiple steps
3. Older: Load from auto-save or export backup

If Design Data Corrupted

1. Export current project as JSON (backup current state)
2. Load previous known-good version
3. Report issue to support with steps to reproduce

---

Best Practices Summary

DO

• Review scripts before running
• Verify bulk changes element-by-element
• Cross-check part numbers with suppliers
• Document AI-assisted decisions
• Use targeted queries instead of blanket updates

DON'T

• Run AI scripts on production data without backup
• Trust part number suggestions without verification
• Skip the verification step for "simple" changes
• Use AI-generated designs without human sign-off

---

Compliance Considerations

For regulated industries (automotive, aerospace, medical):

• AI-generated content may require additional documentation
• Maintain clear traceability between requirements and design
• Consider AI as a tool, not a decision-maker
• Establish company policies for AI-assisted design

---

Next Steps

• Prompt Recipes - Proven prompts that produce reliable results
• Scripting Basics - Understand what AI scripts do
• Debugging Scripts - Fix issues when scripts fail