I once sent an email that referenced a Thursday meeting.
It happened on a Wednesday.
Nobody corrected me. But I guarantee the person who received it noticed. And in that tiny moment… a hairline fracture formed in something I’d spent months building. Trust.
The Detail Nobody Audits
We talk a lot about integrity at the Quietly Working Foundation. Not the bumper-sticker version. The WHELHO version… the principle that says your character is revealed in the spaces no one is watching. The details no one audits.
A wrong day-of-week in an email is exactly that kind of detail.
It’s not dramatic. It won’t crash a server or tank a project. It’s the kind of mistake you’d wave off if someone pointed it out. “Oh, I meant Wednesday. You knew what I meant.” And they did. But something shifted anyway. Something small enough to ignore and persistent enough to compound.
This is true in ministry. In leadership. And it turns out… it’s devastatingly true in AI.
When the Machine Lies With Confidence
Our AI agent at QWF composed an email referencing a real meeting on a real date. Got the date right. Then fabricated which day of the week that date fell on. Wrote it with absolute confidence. Addressed it to someone who was at that meeting and knew exactly what day it was.
The agent wasn’t missing information. It had everything it needed. It just… made something up in the margins. Like a pastor who preaches a solid sermon and then casually misattributes a quote. The core message lands fine. But the person who knows the real source? They start wondering what else is slightly off.
We did the math. At 20 emails per week with even a modest hallucination rate, that’s one wrong date every single week. Fifty-two tiny trust fractures per year. Each one dismissible on its own. Together? A pattern that teaches people to second-guess everything you send.
That math haunts me.
You Can’t Self-Correct What You Don’t Know Is Wrong
The Obi-Wan Kenobi approach… “use the Force, trust your instincts”… doesn’t work here. Not for AI agents. Not for leaders either, honestly.
We tried telling the agent to double-check itself. Same result. Because asking an AI to review its own hallucination is like proofreading your own typos. Your brain fills in what it expects to see. The agent wrote “Thursday” with full confidence. It’s going to read “Thursday” with full confidence too.
So we stopped asking and started building guardrails.
The full technical breakdown lives in Built from Broken: Vol. 3 on our transparency site. The short version: we built a system that intercepts every email, every file, every output before it ships… and mathematically verifies that Tuesday is actually Tuesday. Not a suggestion to check. Not a rule to remember. A wall the agent literally cannot climb over.
Deterministic enforcement beats good intentions. Every time.
Why This Matters
I’ve been in rooms where a leader’s credibility unraveled. Not from scandal. Not from some spectacular moral failure. From a slow accumulation of small inaccuracies that nobody felt important enough to flag… until the day someone finally said, “I just don’t trust what comes out of that office.”
That’s the trajectory of unchecked small errors. In ministry. In leadership. In the AI systems we’re building to serve young people through the Quietly Working Foundation.
We work with youth. We’re building something meant to carry hope into spaces that have been starved of it. That mission doesn’t get a pass on the details. If anything, it demands more rigor in the margins… because the people we serve have already learned to watch for the gap between what someone says and what’s actually true.
Integrity isn’t built in the moments everyone is watching. It’s built in the Tuesday-versus-Thursday details that nobody audits.
We got this one wrong. We caught it. We built the fix. And we showed the receipts.
That’s the work.