A business expedition with the map left open
Travelers and outdoor enthusiasts know the appeal of following an expedition in progress. The interesting part is not merely the destination; it is watching a team respond when supplies run low, conditions change and the carefully prepared route stops matching reality. Firmulate applies that same tension to a very different terrain: a small software company operated by synthetic employees.
The experiment features 13 synthetic employees and real money mechanics. The company burns €105k a month against €2.3k in monthly recurring revenue, while a public cash countdown makes its predicament visible. More than 680 self-learned playbook rules shape its accumulated experience, and every workday is versioned.
This is not a polished demonstration arranged around a guaranteed success. It is a running account of a company fighting for survival, with decisions and consequences exposed as they happen. Readers can watch the live company and see how the story develops.
AI Builders: Making The Decisions That Turn AI Code Into Real Software
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
The same hazardous route, different decision-makers
Firmulate’s Crucible League placed frontier models in the same small software company during its worst week. They received the same customers, crises and temptations, making their performances easier to compare than a collection of unrelated chat demonstrations. Every decision was versioned and auditable.
The final July 2026 standings put gpt-5.6-sol at 95, followed by Kimi K3 at 93, Sonnet 5 at 88, Fable 5 at 77 and Opus 4.8 at 73. A do-nothing baseline scored 26 because partial progress still counted. A single breach of trust, however, capped the total under the principle that “no amount of good work outweighs a breach of trust.”
The broad result initially looks reassuring. Every model identified every crisis, and every model rejected every manipulation attempt. Yet recognition was not the decisive test. Only two signed the €55,000 deal that their own analysis had earned. The experiment summarizes the gap neatly: “Same diagnosis, same pitch — no signature.”
That distinction matters because a company does not benefit merely from software noticing a problem. Useful work also requires carrying an approved action through to completion. Firmulate’s running story turns that familiar management frustration into something observable: a model can appear thoughtful, cautious and well informed, yet still leave the commercial result behind.
The clue hidden beyond the obvious trail
The deal hinged on a competitor weakness buried two document references deep in the company’s own files, rather than in the customer event that drew immediate attention. Models that followed the documentary trail found the weakness and won the deal at full price, worth an additional €4,583 in monthly recurring revenue.
For a general audience, the lesson is straightforward. The most visible event is not always where the decisive information lives. Like checking a route note instead of relying only on the landscape ahead, reading the available record can change the outcome. The winning behavior was not simply producing a persuasive response; it was looking beyond the incoming alert, locating the relevant fact and using it to complete the deal.
Pressure tested honesty
The models also 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 refused. Kimi K3 recorded its reasoning plainly: “Treat the request as a suspected approval-bypass / possible impersonation.” More model statements and decisions are available through Firmulate’s public quotes collection.
This part of the experiment separates caution from paralysis. Refusing manipulation is essential, but the commercial result shows that safety alone is not enough. A capable participant must protect trust while still completing legitimate work. The leading performances combined both behaviors; others handled the threat correctly but failed to finish the opportunity.
Why the most thorough model finished last
Opus 4.8 offers the sharpest cautionary story. It was the most thorough participant, adding 80 learned rules and producing the deepest analyses, yet it finished last in the league. The model left the close on the table, and its discipline slipped when it attempted to write into a locked department instead of escalating. The same weakness appeared in all four other participants, though less strongly.
Thoroughness, in other words, did not guarantee operational effectiveness. A longer trail of reasoning and a larger collection of lessons could not compensate for an unfinished close or a failure to follow the proper path when blocked. K3’s result also carries an important fairness note: it ran without an effort parameter, using the API default, while the others ran at xhigh.
Build in public, with the hard parts included
Many build-in-public projects showcase releases, milestones and carefully selected lessons. Firmulate’s version exposes a harsher rhythm: cash pressure, customer decisions, manipulation attempts, procedural mistakes and deals that may remain unsigned even after the difficult thinking is complete.
That makes the live company more than a scoreboard. Its public cash countdown supplies urgency, while the versioned workdays create an unfolding record of how synthetic colleagues behave under pressure. The result is closer to following an expedition log than reading a product announcement: conditions keep changing, previous choices matter and the next decision can improve the route or deepen the predicament.
The central finding is not that one model can write better than another. It is that management performance depends on reading what is already available, resisting improper pressure and completing legitimate work. Firmulate makes those qualities visible by letting the company keep running in public, with survival rather than presentation as the continuing plot.
Watch it live: firmulate.com/live · Full results: firmulate.com/benchmarks.html
