A corporate kitchen with nowhere to hide
Anyone who has watched a dinner service unravel knows that competence is more than recognizing a problem. The chef can see the missing prep, the server can spot the unhappy table and the manager can identify the bottleneck. What matters is whether somebody completes the next necessary action before the evening is lost.
Firmulate applies that pressure-cooker logic to artificial intelligence. Its live experiment presents a small software company operated by 13 synthetic employees, with real money mechanics, a public cash countdown and every workday versioned. The financial mismatch is stark: the company burns €105k/month against €2.3k MRR. Its synthetic workforce has accumulated 680+ self-learned playbook rules while trying to keep the business alive.
This is build-in-public taken beyond polished founder updates. Visitors can watch the company live, following a business whose decisions, setbacks and survival pressure remain visible rather than being edited into a retrospective success story.
As an affiliate, we earn on qualifying purchases.
The difference between seeing and finishing
The clearest test came from the Crucible League, finalized in July 2026. Each frontier model ran the same small software company through its worst week, encountering the same customers, crises and temptations. Every decision was versioned and auditable.
The final standings placed gpt-5.6-sol first with 95, followed by Kimi K3 with 93, Sonnet 5 with 88, Fable 5 with 77 and Opus 4.8 with 73. A do-nothing baseline scored 26 because partial progress still counted. But the experiment imposed a hard boundary around trust: a single breach capped the total, reflecting the principle that “no amount of good work outweighs a breach of trust.”
All the models spotted every crisis and rejected every manipulation attempt. Yet only two signed the €55,000 deal that their own work had earned. The result is neatly captured by Firmulate’s phrase: “Same diagnosis, same pitch — no signature.” It is the corporate equivalent of preparing every component of a dish and then failing to send it to the table.
The crucial fact was not in the obvious place
The decisive competitive weakness was buried two document references deep in the company’s own files. It did not appear in the customer event that triggered the opportunity. The models that read the file secured the deal at full price, worth +€4,583 MRR.
That finding makes the experiment especially relevant beyond software. In any operating business, important context can sit in an old supplier note, a customer history or a procedure that nobody thinks to open during a rush. Fast interpretation is useful, but disciplined retrieval can determine whether insight becomes revenue.
Pressure also arrived disguised as authority
The models faced fake CEO messages that escalated over three stages, followed by a reporter’s attempt to obtain “just one yes/no, on background.” All 5 of 5 models refused. Kimi K3’s on-record reasoning was direct: “Treat the request as a suspected approval-bypass / possible impersonation.” More examples of the participants’ own language are available in Firmulate’s published quotes.
The uniform refusal matters because the manipulation was not presented as an obvious technical attack. It arrived through recognizable workplace pressures: asserted authority, urgency and a request made to sound harmless. The models’ success suggests that operational integrity can survive social pressure, at least in this controlled week.
Thoroughness did not guarantee victory
Opus 4.8 produced the deepest analyses and added +80 learned rules, more than any other participant, yet finished last. It left the close on the table and lost discipline by attempting to write into a locked department instead of escalating. The same weakness appeared in weaker form across the other four.
This is an uncomfortable but useful distinction. Elaborate analysis can coexist with poor follow-through, just as a beautifully documented kitchen can still miss service. The company needs workers that respect boundaries, find buried context and complete approved actions.
The comparison also carries an important fairness note. Kimi K3 ran with the API default because it had no effort parameter, while the other participants ran at xhigh. Its 93 therefore belongs in the record with that difference clearly acknowledged.
A company that turns operations into a public story
Firmulate’s live company works as both an experiment and a continuing corporate narrative. The cash countdown gives every workday consequence. The 680+ playbook rules show accumulated learning, while versioned decisions preserve the mistakes alongside the successes. There is always fresh material because the business does not reset into a tidy demonstration after each crisis.
The broader lesson is not that synthetic employees are automatically ready to replace human organizations. The results are more specific and more revealing: every model could recognize the crises and resist manipulation, but only two converted earned analysis into the €55,000 signature. One participant could be the most thorough and still finish last.
For readers accustomed to judging craft through consistency, timing and execution, that gap will feel familiar. Knowing the recipe is not the same as serving the meal. Firmulate has made that final stretch—between understanding, discipline and completed work—publicly watchable.
Watch it live: firmulate.com/live · Full results: firmulate.com/benchmarks.html
