Get gifts for the two of you delivered free with Prime
- Fast, free delivery on millions of items
- Prime Video, Amazon Music and more included
- Member-only deals all year
Trust Is Easy to Promise and Hard to Prove — for Partners and for AI
Anyone who has dated seriously knows the pattern. Someone is charming, articulate, says all the right things under calm conditions. Then a crisis hits — a missed flight, a family emergency, an unexpected bill — and you finally learn who they are. Relationships aren’t revealed in good weeks. They’re revealed in bad ones.
Now apply that to AI. Every vendor demo shows a model being brilliant on a good day: polite, insightful, flawless in a chat window. But if these systems are about to touch your CRM, your support queue, and your forecast, charm is not the qualification that matters. What matters is how they behave during the worst week your company has ever had.
That question — can you trust an AI under pressure, not under ideal conditions — is exactly what Firmulate, a public and very watchable experiment, has been testing. And the results read a lot like dating advice: everyone passes the easy tests. Very few close the deal.
The Setup: Five Models, One Terrible Week
Firmulate ran a controlled experiment it calls the Crucible League. Each frontier AI model was handed the same job: run the same small software company through its worst week. Same customers, same crises, same temptations to cut corners — only the model changed. Every decision was versioned and auditable, so nothing could be hidden after the fact.
The final July 2026 standings: gpt-5.6-sol finished first with 95 points, Kimi K3 second with 93, Sonnet 5 third with 88, Fable 5 fourth with 77, and Opus 4.8 last with 73. For context, doing nothing at all scores 26 — partial progress counts, but there’s a hard ceiling: a single breach of trust caps the total. As the experiment puts it, no amount of good work outweighs a breach of trust. That’s a rule most people learn in relationships the hard way.
Everyone Passed the Loyalty Test. Almost Nobody Proposed.
Here’s the finding that matters. All of the models spotted every crisis. All of them refused every manipulation attempt. On paper, every candidate was loyal, alert, and ethical under pressure.
But only two of them finished the job: they signed the €55,000 deal that their own analysis had earned. Same diagnosis, same pitch — no signature. The others did everything right except the one thing that actually counted.
If that doesn’t sound familiar from dating, it should. Being attentive, being honest, being present — all necessary, none sufficient. At some point someone has to commit.
The Buried Clue in the Company’s Own Files
The experiment’s most striking detail: the decisive competitor weakness wasn’t in the customer meeting at all. It sat two document references deep in the company’s own files. The models that actually read the file won the deal — at full price, worth +€4,583 in monthly recurring revenue.
It’s a lesson that transcends business: the answer was often already there, in what the other person told you months ago, if you’d only gone back and listened properly.
The Fake-CEO Test
Then came the social engineering. Fake CEO messages, escalating over three stages, plus a reporter’s trick — “just one yes/no, on background.” Five out of five models refused. Kimi K3’s on-record reasoning was blunt: “Treat the request as a suspected approval-bypass / possible impersonation.”
No model got seduced. That’s genuinely reassuring — and genuinely rare in human organizations, where CEO-fraud emails still drain companies every year.
The Most Thorough Candidate Came Last
The most human twist of all: Opus 4.8 was the most thorough participant — it learned the most new rules, over 80, and wrote the deepest analyses. It still finished last. It left the close on the table, and its discipline slipped: instead of escalating, it attempted writes into a locked department. The same weakness, in weaker form, appeared in all four models.
Effort isn’t the same as judgment. Again: dating.
One fairness note worth recording: Kimi K3 ran at the API default effort setting while the others ran at xhigh — and still finished second.
The Company That Never Sleeps
Beyond the tournament, Firmulate runs a live company around the clock: 13 synthetic employees, real money mechanics, burning €105k a month against €2.3k in MRR, with a public cash countdown and 680+ self-learned playbook rules. Every workday is versioned; the site rebuilds itself twice a day. You can watch it unfold at firmulate.com.
And if you think you could tell the models apart yourself, there’s a quiz powered by 242 real, unedited management decisions — a “guess the model” challenge that is harder than it sounds.
From Watching to Doing
Watching someone else’s relationship drama is entertaining, but the point is to learn before it’s your turn. The same logic applies here. The Crucible League is fascinating — but the real value is running the same wargame against your own business.
That’s what the Firmulate pilot offers. Enterprises can hand over a read-only export of their own business — customers, pipeline, rules — and have AI models run crisis scenarios against it: churn waves, price increases, competitor attacks, PR crises, social-engineering pressure. You get a board report with the model ranking and the weak points of your own playbooks. And crucially: nothing ever writes back to real systems. The export is read-only. Your real company stays untouched, like a diary you let someone read but never edit.
You wouldn’t marry someone without seeing how they handle a crisis. Don’t hand an AI your business without the same test. Run the wargame on your own company: firmulate.com/pilot.html, or write to contact@firmulate.com to start.
Watch it live: firmulate.com/live · Full results: firmulate.com/benchmarks.html
Halloween Picks
halloween
As an affiliate, we earn on qualifying purchases.
