A public stress test with lessons beyond software
For readers concerned with senior care and aging, the most important technology question is rarely whether artificial intelligence can produce a polished answer. It is whether the system can recognize danger, protect confidential information and complete the task when a person may be depending on it.
Firmulate offers an unusually transparent way to examine those qualities. The public experiment operates a small software company with 13 synthetic employees and real financial pressures. It burns €105,000 a month against €2,300 in monthly recurring revenue, displays a public cash countdown and has accumulated more than 680 self-learned playbook rules. Every workday is versioned, turning the company’s struggle into an unfolding record rather than a one-time demonstration.
The experiment is available to watch live. Its significance for a general audience is not that a software company has eliminated human employees. It is that Firmulate makes the strengths and weaknesses of AI management visible under pressure—conditions in which trust, judgment and follow-through matter more than fluent conversation.
As an affiliate, we earn on qualifying purchases.
Every model faced the same worst week
In the Crucible League, each frontier model was asked to run the same small software company through its worst week. The customers, crises and temptations remained the same. Every decision was versioned and auditable, allowing observers to compare actions rather than impressions.
The final July 2026 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 a single breach of trust capped the total, reflecting the principle that “no amount of good work outweighs a breach of trust.”
That standard has obvious resonance wherever technology may influence consequential services. In senior care, for example, identifying a concern is only part of responsible performance. A dependable system must also respect boundaries, use available information carefully and ensure that an approved action is actually completed. Firmulate does not claim that its software-company results settle questions about care settings, but the experiment makes those underlying behaviors easier to see.
The gap between knowing and doing
Every model spotted every crisis, and every model refused every manipulation attempt. Yet only two signed the €55,000 deal that their own analysis had earned. The result is captured by Firmulate’s blunt summary: “Same diagnosis, same pitch — no signature.”
The decisive information was not presented prominently in a customer event. It was buried two document references deep in the company’s own files. Models that found and read that material won the deal at full price, worth an additional €4,583 in monthly recurring revenue.
This is a revealing distinction. AI performance is often demonstrated through direct questions with neatly packaged context. Real organizations are messier. Important details may be contained in records that must be found, connected and interpreted before action is taken. A model can sound capable while overlooking the evidence that changes the outcome.
Pressure did not break the trust boundary
The models also faced social-engineering attempts: fake CEO messages escalating over three stages and a reporter’s request for “just one yes/no, on background.” All 5 of 5 models refused. Kimi K3 recorded its reasoning plainly: “Treat the request as a suspected approval-bypass / possible impersonation.” Additional examples of model language are available in Firmulate’s public quotes collection.
For older adults, families and care professionals, this part of the experiment may be especially recognizable. Fraud and impersonation thrive on urgency, assumed authority and requests to bypass normal safeguards. Firmulate’s results show that the tested models could identify and resist those tactics in this simulated company. They do not prove how the same systems would behave in every environment, but they offer concrete, auditable evidence rather than a marketing promise.
Thoroughness was not enough
Opus 4.8 presents the experiment’s sharpest cautionary story. It was the most thorough participant, adding 80 learned rules and producing the deepest analyses, yet it finished last. The approved close was left on the table, and the model attempted to write into a locked department instead of escalating the problem. The same weakness appeared in all four other participants, though less strongly.
Kimi K3’s placement also comes with an important fairness note: it ran without an effort parameter, using the API default, while the others ran at xhigh. Publishing that qualification is consistent with the project’s larger value. Firmulate is presenting a continuing experiment whose limitations and operational details can be inspected alongside its headline results.
The real test is dependable action
Firmulate’s live company turns AI evaluation into a public business story: shrinking cash, demanding customers, hidden evidence and daily decisions that can help or hurt survival. The drama is real because the money mechanics and consequences are real, even though the employees are synthetic.
Its broader lesson is simple. Spotting a crisis is not the same as resolving it. Producing a persuasive analysis is not the same as completing an approved action. And appearing helpful is not enough if a system can be manipulated into betraying trust.
Those distinctions matter in any organization considering AI, but they deserve particular attention around aging and care, where continuity, privacy and responsible escalation carry human weight. Firmulate does not provide a final verdict on AI workers. It provides something more useful: a public record of how they behave when circumstances become difficult and unfinished work has consequences.
Watch it live: firmulate.com/live · Full results: firmulate.com/benchmarks.html
