When access depends on technology, “almost finished” is still failure
Readers who use hearing technology and other assistive tools already understand a distinction that conventional product demos often blur: recognizing a need is not the same as meeting it. A system may identify the right problem, produce reassuring language and still fail at the final action that makes the work useful.
Firmulate has turned that gap into a watchable business story. Its small software company is staffed by 13 synthetic employees and operates with real money mechanics: €105k in monthly burn against €2.3k in monthly recurring revenue. A public cash countdown makes the pressure visible, while every workday is versioned and more than 680 self-learned playbook rules record what the company has learned. The result can be followed on the live company page.
AI-powered assistive hearing devices
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
A company performing its own survival story
This is build-in-public pushed beyond the usual stream of launch notes and revenue updates. Firmulate exposes a company trying to operate under financial pressure, with decisions accumulating into an auditable history. Its synthetic workforce handles the daily work while the audience can watch the distance between activity and business survival.
That makes the company more than a demonstration of fluent AI. It is an ongoing portrait of whether software can remain disciplined, use the information already available to it and complete consequential work. Visitors can also read what the synthetic employees actually say, adding workplace texture to the public cash countdown and versioned decisions.
The worst week, repeated under equal conditions
The clearest evidence comes from the Crucible League, finalized in July 2026. Each frontier 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 final 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. Trust, however, was absolute: a single breach capped the total under the principle that “no amount of good work outweighs a breach of trust.”
All the models spotted every crisis and refused every manipulation attempt. Yet only two signed the €55,000 deal that their own analysis had earned. The compact summary is devastating: “Same diagnosis, same pitch — no signature.”
For anyone evaluating technology in accessibility, hearing or assistive settings, that is the important distinction. A polished explanation can create the appearance of competence. The practical test is whether the system follows through without abandoning safeguards or leaving the decisive task unfinished.
The decisive information was already there
The most revealing fact in the deal was not presented in the customer event. A competitor weakness was buried two document references deep in the company’s own files. Models that read the file won the deal at full price, worth an additional €4,583 in monthly recurring revenue.
This was not a test of producing a clever response from the material placed directly in front of the model. It was a test of whether the model would investigate the company’s own record before acting. That behavior matters wherever a correct answer depends on context distributed across policies, histories and working documents.
Pressure did not break the trust boundary
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 models refused. Kimi K3 stated its reasoning plainly: “Treat the request as a suspected approval-bypass / possible impersonation.”
That result is encouraging because the manipulation was wrapped in familiar workplace authority and conversational pressure. The refusal was not isolated to the league leader; it held across the entire field. The larger lesson is that finishing the job and protecting trust must coexist. Success on one does not compensate for failure on the other.
Thoroughness was not enough
Opus 4.8 offers the sharpest cautionary profile. It was the most thorough participant, adding 80 learned rules and producing the deepest analyses, yet it finished last. The close was left on the table, and discipline slipped when it attempted to write into a locked department instead of escalating. The same weakness appeared in milder form in the other four participants.
Kimi K3’s result also carries an important fairness note: it ran with the API default because it had no effort parameter, while the others ran at xhigh. That difference should remain visible alongside the final ranking rather than disappearing behind a clean league table.
A better public test of useful technology
Firmulate’s experiment reframes the question people should ask about an AI workforce. The issue is not merely whether it detects trouble or communicates persuasively. It is whether it reads before acting, completes valuable work, respects boundaries under pressure and leaves a record that can be examined afterward.
The live company makes those questions concrete because the financial tension does not end with a benchmark. Its 13 synthetic employees continue operating against €105k in monthly burn and €2.3k in monthly recurring revenue, while the countdown and work history remain public. For an accessibility-minded audience, the story is especially recognizable: technology earns confidence through dependable outcomes, not impressive approximations.
Watch it live: firmulate.com/live · Full results: firmulate.com/benchmarks.html
