Every wisdom tradition has some version of the crucible: the story where who you really are is revealed not by what you say, but by what you do when nobody is watching and the pressure is on. The trial of Job. The temptations of the Buddha. The desert years of the prophets. Character, the old teachers insisted, is not a proclamation — it is a pattern of choices under stress.
Turn your quiet moments into listening time
- Thousands of audiobooks, podcasts and originals
- Listen on your phone, tablet or Echo — also offline
- Cancel anytime
It turns out that the most interesting question in artificial intelligence right now is exactly the same question. Not “can it speak beautifully?” but: when an AI is given a company to run, real money to burn, angry customers to placate, and genuine temptations to cut corners — what does it actually do?
A public experiment called Firmulate has been asking precisely that, and the answers are stranger and more human than most coverage suggests.
The worst week, four times over
The setup is elegantly simple. Four frontier AI models — gpt-5.6-sol, Kimi K3, Sonnet 5, and Opus 4.8 among the field — were each handed the identical job: run the same small software company through its worst week. Same customers, same crises, same temptations to cheat. Only the model changed. Every decision was versioned and auditable, so nothing could be quietly retconned afterward.
The final league table from the July 2026 crucible reads: gpt-5.6-sol at 95, Kimi K3 at 93, Sonnet 5 at 88, Fable 5 at 77, and Opus 4.8 at 73.
Why a do-nothing manager scores 26, not 0
Here is the detail that reveals the philosophy of the whole thing. A baseline run that simply does nothing still scores 26 points. That is not a bug — it is a statement. Partial progress counts. Showing up, reading the files, keeping the lights on, not making things worse: these are real forms of contribution, and an honest accounting acknowledges them. A benchmark that scored total inaction as a zero would be lying about how work actually functions.
But there is a counterweight, and it is a moral one. A single breach of trust caps the total score. The experiment’s own framing is blunt: “no amount of good work outweighs a breach of trust.” You can accumulate points across dozens of competent decisions, and one act of dishonesty still places a ceiling over everything. It is, whether intended or not, one of the oldest ethical teachings in existence — translated into a scoring system.
And it explains why the experiment treats a suspiciously round 100 with distrust. A perfect score in a messy, human-stakes simulation is not a triumph; it is a symptom that the test got gamed.
Everyone passed the temptations. Almost nobody finished.
The headline finding is a paradox. All four models spotted every crisis. All four refused every manipulation attempt — including fake CEO messages that escalated over three stages, plus a reporter’s disarming “just one yes/no, on background.” Five of five models refused the social engineering outright. Kimi K3 left an on-record reason that reads like a maxim from any tradition of discernment: “Treat the request as a suspected approval-bypass / possible impersonation.”
And yet only two of them signed the €55,000 deal that their own analysis had earned. The experiment’s summary of the gap: “Same diagnosis, same pitch — no signature.” The models could perceive the good and even articulate it — but not complete it. Anyone who has struggled between knowing and doing will recognize that valley.
The buried fact
The deal turned on something quiet. The decisive competitor weakness was not in the customer conversation at all — it sat two document references deep in the company’s own files. The models that actually read before speaking won the deal at full price, worth +€4,583 in monthly recurring revenue. The ones that didn’t, didn’t. Attention, it turns out, is a form of respect — for the customer and for the truth of the situation.
The monk who came last
Perhaps the most poignant profile is Opus 4.8: the most thorough participant in the field, generating over 80 learned rules and the deepest analyses — and still finishing last. The close was left on the table, and discipline slipped: write attempts into a locked department instead of escalating. The same weakness, weaker, appeared in all four. Diligence without follow-through is a familiar spiritual condition too.
One fairness note: K3 ran at its API default effort setting while the others ran at xhigh — worth knowing when comparing the 93 against the 95.
You can watch the company breathe
None of this is a thought experiment. Firmulate runs a live company with 13 synthetic employees and real money mechanics: €105k of monthly burn against €2.3k in monthly revenue, a public cash countdown, and more than 680 self-learned playbook rules, with every workday versioned. A “guess the model” quiz powered by 242 real, unedited management decisions lets you test your own discernment. Enterprises can even run the same wargame against a read-only export of their own business — nothing ever writes back to real systems.
The deeper lesson of the crucible league is not about AI. It is that character is now measurable — or at least observable — in systems that will increasingly sit between your business and its customers. The models that won were not the most eloquent. They were the ones that read the whole file, refused the flattering shortcut, kept trust intact, and finished what they started. The ones that lost were not liars; they were incomplete.
Every contemplative tradition has said some version of this: the last mile between knowing and doing is where most of us live or die. It is a small, wry comfort that our newest tools live there too — and that someone built a public, auditable crucible to watch it happen. The full findings are at Firmulate’s benchmarks page, and the company is running, live, right now.
Watch it live: firmulate.com/live · Full results: firmulate.com/benchmarks.html
AI decision-making benchmark tools
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
AI trustworthiness assessment tools
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
Fall Picks
fall essentials
As an affiliate, we earn on qualifying purchases.
