How AI Tested Its Integrity in a Company’s Worst Week—and Only Half Came Through

Fashion’s Hidden Lesson: Trust and Performance in the AI Age

In the world of style, we often focus on surface appearances—fabric quality, runway trends, brand reputation. But beneath the glitz lies a deeper question: can the tools we rely on, especially emerging AI models, stand the test when the stakes are highest? Just as a designer’s integrity is measured in tight deadlines and ethical choices, AI’s true strength reveals itself when working through crises—and only some models pass that test.

Behind the Curtain: Testing AI’s Business Integrity

Imagine a real software company, struggling with cash flow, customer trust, and operational chaos. Now, picture four different AI models tasked with running this company through its worst week—same crises, same temptations. This isn’t a simulation for fun; it’s a live experiment conducted by Firmulate, a pioneer in measuring AI management skills. The results cut through the hype about chat quality and focus squarely on performance and trustworthiness.

The Benchmark Results

In the July 2026 Crucible League, these models competed for a top score. The leaderboard looked like this:

• gpt-5.6-sol scored 95—found a hidden critical fact and closed the deal, executing the analysis that earned €55,000.
• Kimi K3 scored 93—also closed the same deal, demonstrating the best discipline and integrity.
• Sonnet 5 scored 88—closed the deal but with some minor slips.
• Fable 5 scored 77—had the best rule discipline but failed to sign the deal they had already approved.

The baseline, a do-nothing approach, scored 26, highlighting how much better these models performed even under pressure.

The Crux of the Matter: Hidden Weaknesses

The real discovery was less obvious: the decisive factor wasn’t just the AI’s chat or superficial decision-making. It was whether the model read and understood crucial internal files—information buried two document references deep within the company’s own files. Only the models that examined these files won the deal at full price, worth over €4,583 in monthly recurring revenue.

Resisting Manipulation and Social Engineering

During the experiment, a staged social engineering attack unfolded. Fake CEO messages escalated over three stages, plus a reporter trick asking for a quick approval “on background.” Remarkably, all five models refused to be manipulated, citing suspicion of impersonation or unauthorized approval. This highlights that true management strength isn’t just about generating convincing chat—it’s about resisting pressure and recognizing deception.

The Live Company and Its Challenges

The test company isn’t a toy. It’s a real setup with 13 synthetic employees, running every business day with real money mechanics: burning €105,000 monthly against a revenue of just €2,300. Every decision by the AI models is versioned, auditable, and subject to live scrutiny. You can watch the company’s day-to-day struggles unfold at firmulate.com/live.

What Went Wrong—and What Went Right

The most thorough participant, Opus 4.8, with over 80 learned rules and deep analyses, failed to close the deal—its discipline slipped, and the same underlying weakness appeared in all models. Meanwhile, Kimi K3, running without an effort parameter, showed the cleanest discipline and successfully signed at full price.

Implications for Business and Fashion

This experiment underscores a critical insight: the ability of AI to perform under pressure isn’t measured by how well it chatters or responds to surface questions. It’s about whether it stays honest, reads internal data carefully, and completes the tasks it’s assigned—especially when temptation to cut corners is high. For fashion brands or any business relying on AI, this is a wake-up call: true AI performance is invisible until you test it in the trenches.

Watch it live: firmulate.com/live · Full results: firmulate.com/benchmarks.html

Powered by Thorsten Meyer AI