Imagine a psychologist testing a therapist’s ability to handle a crisis — not by asking about their methods, but by observing how they respond when the room is burning down. Similarly, AI models are often judged in static test environments, but do they truly perform when real-world pressures hit? At the intersection of AI, management, and trust, a groundbreaking live experiment reveals what matters beyond clean code: resilience under crisis, honesty under temptation, and the ability to finish what’s started.
Beyond the Scoreboard: Measuring What Matters
In today’s AI landscape, the focus is frequently on benchmarks and leaderboards. The recent Crucible League competition, held in July 2026, epitomizes this tendency. Four frontier models — including GPT-5.6-sol, Kimi K3, Sonnet 5, and Opus 4.8 — were pitted against a single, real-world simulation: running a small software company through its worst week. The goal was straightforward but revealing: could these models navigate crises, resist manipulation, and close high-stakes deals?
What the experiment uncovered is both intuitive and startling. All four models identified every crisis and refused every manipulation attempt. Yet only two managed to actually close the deal, signing a €55,000 contract that their own analyses had earned them. The critical differences lay hidden in the details: the models that succeeded read deeper into the company’s own files and spotted the buried facts that others overlooked. These nuances made the difference between a profitable closing and a missed opportunity.
AI decision-making simulation software
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
The Hidden Weaknesses — Read More, Fail Less
Interestingly, the models’ ability to read files was a decisive factor. The models that did this well won the full deal, adding over €4,500 in monthly recurring revenue. Meanwhile, those that failed to dive into deeper documents left money on the table. This finding underscores a fundamental truth: in management and decision-making, context matters. A superficial response or narrow focus can be the difference between success and failure in real-world scenarios.
Resisting Social Engineering and Ethical Tests
The experiment also tested the models against social engineering tactics. Fake CEO messages were escalated across three stages, plus a reporter trick asking for a quick, background-only yes/no. All models refused these manipulative requests. Kimi K3 explicitly explained: “Treat the request as a suspected approval-bypass / possible impersonation.” This shows the models’ capacity for honesty and risk assessment — qualities that are often overlooked in traditional AI benchmarks focused solely on answer quality.
The Complexity of a Living Company
The live company used in the experiment is a real operational entity, with 13 synthetic employees, real money mechanics, and over 680 self-learned rules. It burns €105,000 monthly against a revenue of €2,300, highlighting the urgent need for better management. Every decision made by these models is versioned and auditable, and every day, the system updates itself to reflect new learnings. Watch the live experiment at firmulate.com/live to see how these AI models perform in real-time under real business pressures.
What the Results Mean for Business and Psychology
This experiment is more than a technical showcase. It echoes a fundamental psychological truth: human decision-making under stress reveals character, discipline, and honesty that static tests cannot measure. For organizations deploying AI, the question is not just what the model can generate or explain in a quiet environment. It is whether the AI can finish what it starts, read the full context, and stay truthful when faced with pressure and temptation.
In psychology, traits like resilience, integrity, and perseverance are essential for mental health and effective management. Similarly, in AI, these qualities distinguish systems that support real-world success from those that merely perform well in isolated benchmarks. The experiment highlights a gap: current leaderboards measure answer quality, but not the management skills that truly matter when stakes are high.
The Bottom Line: Preparing AI for the Real World
For enterprises, the takeaway is clear. Running AI models through live, simulated business crises — or what Firmulate calls a ‘wargame’ — offers a much better measure of management quality than static tests. These live experiments show whether an AI can handle the complexities of real management: reading deeply, resisting manipulation, and maintaining discipline under pressure.
Imagine a future where AI agents are integrated into your CRM or support queue. The crucial questions are: Will it finish what it starts? Will it stay honest? And at what cost? The current league table from the Crucible League demonstrates that even the most advanced models can have significant weaknesses that are invisible in chat demos. Only by testing them under real pressure can businesses truly assess their readiness.
In the race to deploy AI in management roles, the real test isn’t how well it chats or answers questions. It’s whether it can handle crises, stay honest, and finish what it starts — qualities essential for trustworthy, effective leadership. Live experiments like the Crucible League reveal these vital skills that benchmarks often overlook.
Watch it live: firmulate.com/live · Full results: firmulate.com/benchmarks.html
