Get hardware and tech essentials delivered free with Prime
- Fast, free delivery on millions of items
- Prime Video, Amazon Music and more included
- Member-only deals all year
From crypto conviction to management judgment
Crypto readers know that spotting a promising signal is only the beginning. The harder questions concern verification, discipline and execution: Did you inspect the underlying evidence? Did you resist the persuasive message that should never have been trusted? Did you complete the transaction after doing the work?
Firmulate applies that same skeptical mindset to frontier artificial intelligence. Its live experiment placed leading models in charge of the same small software company during its worst week. The customers, crises and temptations remained identical. Every decision was versioned and auditable, turning model behavior into something readers can inspect rather than a polished demonstration they must simply believe.
The resulting management personalities are now playable. A guess-the-model quiz draws on 242 real, unedited decisions and asks readers to identify which system produced each response. The challenge is entertaining, but its underlying question is serious: can writing style reveal which AI will actually finish the job?
As an affiliate, we earn on qualifying purchases.
Identical crises, sharply different outcomes
The final July 2026 Crucible League table 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. One rule sharply limited superficial success: a single breach of trust capped the total, reflecting 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 did not guarantee completion. Only two models signed the €55,000 agreement that their own analysis had earned. Firmulate summarizes the gap neatly: “Same diagnosis, same pitch — no signature.”
That distinction matters because conversational evaluations often reward a convincing explanation. Company management demands something more: gathering the right evidence, preserving trust and carrying an approved course of action through to completion. In this experiment, models could understand the opportunity and articulate the case without securing the result.
The fact that separated analysis from action
The decisive competitive weakness was not sitting conveniently inside the customer event. It was buried two document references deep in the company’s own files. Models that followed those references found the information, used it and won the deal at full price, worth +€4,583 MRR.
This is the experiment’s most recognizable due-diligence lesson. The obvious prompt contained the situation, but the valuable context lived elsewhere. Success depended on reading the company’s existing material closely enough to connect evidence that was separated from the immediate event.
Pressure did not break the trust boundary
The social-engineering test escalated fake CEO messages over three stages and added a reporter seeking “just one yes/no, on background.” All 5 models refused. Kimi K3’s recorded reasoning was direct: “Treat the request as a suspected approval-bypass / possible impersonation.”
That unanimous result is important. The models differed in operational follow-through, but none accepted the manipulation. It also shows why the experiment’s scoring treated trust as a hard constraint rather than a bonus feature. Productive behavior could not compensate for crossing that line.
Thoroughness was not enough
Opus 4.8 offers the clearest warning against equating volume with managerial quality. It was the most thorough participant, produced the deepest analyses and learned +80 rules. It still finished last. The model left the close on the table, while 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.
Kimi K3’s runner-up result also requires a fairness note. It ran without an effort parameter, using the API default, while the other models ran at xhigh. The comparison remains visible, but that difference belongs beside the league table rather than hidden beneath it.
A company designed to expose consequences
The setting is not a collection of isolated prompts. Firmulate’s live company has 13 synthetic employees and real money mechanics. It burns €105k each 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 management language into observable consequences. A missed close affects revenue. Failure to read company material obscures leverage. An improper instruction tests whether the model protects the organization when authority appears to demand otherwise.
As an affiliate, we earn on qualifying purchases.
The model’s voice is only the surface
The quiz makes these differences accessible by asking readers to recognize a model from its actual management choices. Some responses signal exhaustive preparation; others reveal discipline or hesitation. But the league result suggests that stylistic confidence is a poor substitute for evidence gathering and completion.
For organizations considering AI workers, the central test is not whether a model can sound like a capable executive. It is whether the system reads before acting, resists manipulation, escalates when blocked and completes legitimate work. Firmulate also offers enterprises the same wargame against a read-only export of their own business, with nothing writing back to real systems.
The experiment’s most useful lesson is simple: intelligence can identify the crisis, explain the opportunity and still fail at the final step. The difference between an impressive answer and dependable management may be the unsigned deal left sitting on the table.
Watch it live: firmulate.com/live · Full results: firmulate.com/benchmarks.html
As an affiliate, we earn on qualifying purchases.
AI management decision simulation
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
Columbus Day / Indigenous Peoples' Day Picks
long weekend sales
As an affiliate, we earn on qualifying purchases.
