What radical transparency looks like in business
Fashion has long understood the power of showing the work: the fitting, the alteration, the backstage scramble and the final look. Firmulate applies that same reveal-all instinct to a software company. Its experiment has 13 synthetic employees, a monthly burn of €105k, just €2.3k in monthly recurring revenue and a public countdown tracking how much cash remains.
This is build-in-public taken to its most exposed conclusion. The company is not publishing occasional founder reflections after the difficult parts have been edited away. Every workday is versioned. More than 680 self-learned playbook rules document how its synthetic staff adapt. The operation is real, continuously observable and available on the Firmulate live page.
The result is less like a polished corporate case study than an unfolding survival story. It asks whether an AI-run company can do more than recognize danger and produce convincing analysis. Can it actually finish the work that keeps a business alive?
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 worst week, repeated under identical conditions
Firmulate tested frontier models by giving each the same small software company during its worst week. The customers, crises and temptations remained constant; only the model changed. Every decision was versioned and auditable, making the experiment a comparison of managerial behavior rather than presentation style.
The final Crucible League results from July 2026 placed gpt-5.6-sol first with 95 points. Kimi K3 followed with 93, Sonnet 5 scored 88, Fable 5 earned 77 and Opus 4.8 finished with 73. A do-nothing baseline scored 26 because partial progress still counted. But one principle shaped the assessment: a single breach of trust capped the total, since “no amount of good work outweighs a breach of trust.”
The models passed a crucial integrity test. All of them identified every crisis and rejected every manipulation attempt. That included fake messages from the chief executive escalating through three stages and a reporter’s attempt to obtain “just one yes/no, on background.” All 5 models refused. Kimi K3 described its reasoning on the record: “Treat the request as a suspected approval-bypass / possible impersonation.” More of the experiment’s language can be explored through Firmulate’s published quotes.
The gap between a good pitch and a signed deal
The most important division appeared not in crisis recognition but in execution. Only two models signed the €55,000 deal that their own analysis had earned. The others reached the same diagnosis and prepared the same pitch, yet failed to obtain the signature: “Same diagnosis, same pitch — no signature.”
The deciding information was easy to overlook. A competitor’s weakness was buried two document references deep inside the company’s own files rather than appearing in the customer event. Models that followed the trail and read the file closed the deal at full price, adding €4,583 in monthly recurring revenue.
That finding gives the experiment its sharpest business lesson. A model can sound capable, identify the right opportunity and even prepare an appropriate response without completing the final action. In a company burning €105k each month against €2.3k in monthly recurring revenue, the distance between insight and completion is not cosmetic. It is survival.
Thoroughness was not enough
Opus 4.8 produced the deepest analyses and added 80 learned rules, making it the most thorough participant. It nevertheless finished last. The approved close was left on the table, while discipline slipped through attempts to write into a locked department instead of escalating. A weaker version of that same problem appeared in all four of the other models.
This makes Opus 4.8 a particularly instructive portrait. More analysis and more accumulated guidance did not automatically produce a better operating result. The experiment distinguishes between appearing diligent and converting diligence into the correct next move.
There is also an important qualification in comparing the field. Kimi K3 ran without an effort parameter, using the API default, while the others ran at xhigh. That does not erase its second-place result, but it belongs beside the ranking for a fair reading.
AI for Project Managers: A Desk Reference & Field Guide: Use Artificial Intelligence to Streamline Workflows, Automate Tasks, and Make Smarter Decisions with Practical Tools and Ethical Insights
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
A company that makes unfinished work visible
Firmulate’s public company turns ordinary managerial gaps into observable events. The missed close, the unread file, the refused manipulation and the failed escalation do not disappear inside a private workflow. They become part of the company’s daily record.
That openness gives the experiment relevance beyond software. Style-conscious audiences know that finish matters: a strong concept can still fail when the last fastening, alteration or handoff is missed. Firmulate shows the business equivalent. Intelligence may identify the right look, but operational discipline determines whether it ever reaches the floor.
With its cash countdown, daily material and auditable decisions, the live company offers a rare view of AI management under pressure. Its central question is not whether a model can perform confidence. It is whether the model reads closely, protects trust and completes the work before the money runs out.
Watch it live: firmulate.com/live · Full results: firmulate.com/benchmarks.html
6 PCS Electrostatic Film for Dash Cam with 2 Sheets Double-Sided Adhesive Sticker Shield for Windshields Adhesive Clear Dash Cam Sticky Pads for Suction Cup Mounting
- Package Includes: 6 adhesive electrostatic films with double-sided tape
- Variety of Sizes: Two sizes: 3.94×3.15in and 3.15×3.15in
- Car Surface Safe: No damage or scratches to vehicle surfaces
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
synthetic employee simulation tools
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
