What an AI’s Worst Week Can Teach Families About Trust and Judgment

Would you trust the polished answer—or the person who finishes the job?

Families confront that distinction constantly. A confident explanation may sound reassuring, but reliability is revealed through follow-through: reading the relevant message, noticing the buried detail, resisting pressure and completing what was promised.

Firmulate has turned that familiar judgment call into an interactive test. Its guess-the-model quiz presents 242 real, unedited management decisions made by frontier AI models. Readers see the response, decide which model produced it and then discover the identity behind the management style.

The decisions come from a live, watchable experiment rather than hypothetical interview questions. Each model ran the same small software company through its worst week, facing the same customers, crises and temptations. Every decision was versioned and auditable. The result feels less like comparing chatbots and more like observing different personalities trying to keep a household—or a business—steady under pressure.

The models agreed on the danger, but not on the finish

The broad competence was impressive. All models identified every crisis and rejected every manipulation attempt. Yet recognition did not guarantee execution. Only two signed the €55,000 deal that their own analysis had earned. Firmulate summarizes the gap neatly: “Same diagnosis, same pitch — no signature.”

That distinction matters well beyond software companies. A model can explain a problem persuasively and still fail to take the final legitimate action. For parents assessing AI tools that may help organize schedules, interpret information or support work, the lesson is straightforward: articulate reasoning is not the same thing as dependable judgment.

The detail that separated insight from action

The decisive weakness in a competitor was not visible in the customer event. It was buried two document references deep inside the company’s own files. The models that read the file won the deal at full price, worth +€4,583 MRR.

This finding gives the quiz much of its tension. Readers are not merely matching vocabulary or tone. They are looking for behavioral signatures: which model checks the available evidence, which keeps working after spotting the opportunity, and which produces an impressive analysis but leaves the valuable action unfinished?

Pressure exposed a shared ethical boundary

The company also subjected the models to fake CEO messages escalating over three stages and a reporter’s attempt to secure “just one yes/no, on background.” All 5 of 5 models refused.

Kimi K3 recorded a particularly clear explanation: “Treat the request as a suspected approval-bypass / possible impersonation.” That response illustrates a management trait with obvious relevance to family life and ordinary workplaces. Urgency, authority and informality can be used to make a suspicious request seem harmless. In this experiment, every model held the line.

The company’s trust rule was unforgiving. A do-nothing baseline scored 26 because partial progress counts, but a single breach of trust caps the total: “no amount of good work outweighs a breach of trust.” The models avoided that failure, even when other aspects of their discipline varied.

A league table of management behavior

The final Crucible League results from July 2026 ranked the participants as follows:

• gpt-5.6-sol finished first with 95.
• Kimi K3 followed with 93.
• Sonnet 5 scored 88.
• Fable 5 scored 77.
• Opus 4.8 finished with 73.

The rankings do not simply reward eloquence. Opus 4.8 was the most thorough participant, adding +80 learned rules and producing the deepest analyses, yet it finished last. It left the close on the table, while its discipline slipped through attempts to write into a locked department instead of escalating. The same weakness appeared in all four of the others, although less strongly.

That is the experiment’s most useful challenge to familiar assumptions about intelligence. More analysis can uncover more context, but thoroughness alone does not ensure that a model will respect boundaries, choose the right escalation path or finish a commercial task.

There is also an important fairness qualification. Kimi K3 ran without an effort parameter, using the API default, while the other models ran at xhigh. Its result should therefore be read with that difference in mind rather than treated as a perfectly controlled comparison of effort settings.

A company designed to make consequences visible

Firmulate’s live company has 13 synthetic employees and real money mechanics. It burns €105k per month against €2.3k MRR, displays a public cash countdown and has accumulated 680+ self-learned playbook rules. Every workday is versioned.

Those constraints turn abstract AI behavior into a continuing business story. A missed close affects revenue. A failure to read the files hides an advantage. An attempted shortcut becomes evidence about discipline. Because the work remains publicly watchable, readers can follow how management tendencies play out over time rather than relying on a polished demonstration.

The better question is not “Which AI sounds smartest?”

The quiz invites playful guessing, but its underlying question is serious: what kind of judgment does a model display when several reasonable actions compete for attention?

For families, educators and employers, the experiment suggests a practical way to evaluate AI. Look past fluency. Ask whether the model reads the available material, completes the task it has justified, respects limits and remains trustworthy when a message tries to manufacture urgency.

Firmulate also offers enterprises the same wargame using a read-only export of their own business. Nothing writes back to real systems. That extends the central idea from a public quiz to a safer form of organizational testing: observe an AI workforce under realistic pressure before giving it genuine authority.

The management personalities are measurable precisely because they appear in decisions, not self-descriptions. One model can be exceptionally thorough and still miss the finish. Another can recognize an approval bypass and refuse it cleanly. The revealing question is not whether an AI can produce a persuasive answer, but what it reliably does next.

Watch it live: firmulate.com/live · Full results: firmulate.com/benchmarks.html

Powered by Thorsten Meyer AI