We wrote an internal plan this week and tried to break it before anyone signed off on it. Ten drafts, three separate machine-run adversarial passes, one human redirect, about a day.
The tool is called logic-critic. It's one of the adversarial skills built into Friday OS, and its only job is finding the flaw in an argument, not the typos, the actual reasoning. We ran a real strategy document through it three times over that day, and each pass caught something the last one missed.
Not every fix came from the tool. Partway through, the accountable executive redirected the plan himself: swap a rotating group of writers for one single owner. That's a human call, not a machine catch, and it's worth saying plainly instead of blurring the two together. The machine is good at finding flaws in reasoning. It doesn't replace someone with the authority to just decide.
The first logic-critic pass found a diagnosis stated as settled fact when it was still a hypothesis, a caveat that had quietly dropped out of an earlier draft, and a rule with no threshold attached to it, so nobody would know when it had actually been triggered.
The second pass caught something we didn't expect: two provisional answers in the plan sounded fine read separately, but they only held up together if a condition were true that the plan itself had already ruled out elsewhere.
The third pass caught the most specific problem of the three. The plan offered two options for where a new initiative should sit, nested under an existing one or standing alone, when a legitimate third option, holding off on the decision entirely until a later date, had been raised earlier in the process and then quietly dropped.
None of the three were things a single human read-through catches reliably, because none of them are typos. Each one needed something actively trying to break the argument instead of just reading it once and nodding along.
Ten drafts, three machine catches, one human call, about a day. That's what it actually took to ship a plan that holds up.
