The first sanctioned insurgency

This is the case study I point chief executives at, told the way a chief executive needs to hear it. Not what got built. Who granted sanctuary, what the insurgent did with it, what kept it from going off the rails, and what the result was, with the caveats left in. Every figure here is lifted from the reviewed record at uncommonwaters.ai; where the record is silent, so am I.

WarriorPath operated as a Mid-America field prototype. It is independent and is not affiliated with or endorsed by the Department of Defense. “Pilot” describes field use, not Navy authorization or endorsement.

The incumbent's condition

A Navy recruiting region for special warfare candidates. A mentor has a region and five weekdays, and attention is the scarce resource, so every week he decides which candidates to spend it on. He decided nearly blind. The legacy tracker recorded status and outcomes: who signed a card, what the recruiter did. It did not show who trains when nobody is watching. On any given Tuesday a candidate who works out only when someone stands over him and a candidate who has been putting in work all week on his own looked identical.

So the organization managed the only number it had: pool size. A pool measured by size hides the candidates who were never going to test, and the answer to a missed month was always the same: more names, more incentives, more pressure. That is the incumbent's condition in one paragraph. It is your company with the nouns changed.

The sanctuary

On December 5, 2025, I started building in earnest for that region. I had no engineering background and had never shipped software. What I had was the thing this whole site is about: enough room to operate. Nobody put my work through a procurement cycle or a steering committee. By January 2026 it was running in NTAG Mid-America. Everything below happened in the ten months since.

The entire cost to leadership was one hour a week. That is what sanctuary looked like here. Not a budget line, not a transformation office. One hour, and the decision not to stop me.

The insurgent

I have never hand-written a line of the code. Agents wrote all of it under my direction. My job was deciding what to build, what was true, and what was allowed to ship. If you are looking for your own insurgents, that is the profile: not the person who can code, the person who knows what the organization needs, can tell when a model's output is garbage, and will eat the cost of saying so.

The first version proved the last point. I built one and killed it. The model output was unreliable in ways I could not gate around at the time, and shipping it would have put something in front of recruiters that lied to them occasionally. I threw it away and rebuilt. The governance layer exists because of what that first attempt taught me.

The doctrine that kept it from becoming a disaster

Anyone can get a model to produce an application. The problem is that you then have an application nobody has a reason to trust. I built the control plane first and the product on top of it: deterministic checks block, models advise, and a human judges. Every implementation is bound to an approved plan by content hash. Privileged operations route to my phone, because agents run with my credentials, and a system that cannot tell my intent from theirs is not a system.

Two refusals did the rest. The system never assigns a candidate a predictive score or ranks people for priority; staff see effort and recency, and the mentor judges. And silence is a flag, not a verdict: about one in four candidates who log nothing contract anyway, so a blank week starts a conversation instead of closing a file. Those are rules of engagement. They were built into the product, not written into a policy.

What the eighty percent experienced

Nobody on the staff learned to prompt anything. Frontier AI is inside the product, not bolted on. Brief Me is a spoken daily standup for field staff; every number is precomputed from the same data the dashboard renders, so the voice cannot invent a figure. The AI Swim Buddy coaches candidates, turning a candidate's own answer into a specific training action grounded in his logged data. The work of the week changed without anyone being asked to change.

The operating change for leadership was small and total: one hour a week, looking at who is actually working and spending attention there. Every candidate leaves that call with an owner and a next step. Pool health replaced pool size as the measure.

The result, with the caveats left in

NTAG Mid-America went from dead last to the #2 NTAG in the nation. That is the official CNRC ranking, confirmed September 2026. The region hit its FY2026 contract goal about six weeks early. WarriorPath did that. I built it, I ran it, and I claim the result.

Underneath the ranking, the mechanism was measured: 45 of 60 forecast-pass dates landed within 14 days of the contract date, and in the final four measured months, 24 of 24 did (measured September 3, 2026). Among decided candidates with at least 60 days in the program, 55% of those logging no work quit, versus 24% of those logging work (Fisher p=0.016; measured August 29, 2026).

The caveats. There was no controlled trial; a federal recruiting region does not run an A/B test on itself. The system also runs in NTAG Jacksonville, and there is no published outcomes study for it, so I make no claim there. The scale is two regions, not a national rollout. The claim was never that it is big. The claim is that it is real, and that one sanctioned insurgent with frontier AI produced all of it.

What this proves for you

Pilots do not fail because the tooling is bad. This one worked because it had the three things yours are missing: sanctuary from one person, an insurgent who could tell truth from output, and doctrine that let the organization trust what shipped. None of those is a technology. All of them are decisions a chief executive can make this quarter.

The engineering behind the governance layer is published separately for technical readers at uncommonwaters.ai, with a sanitized evidence pack of verbatim GitHub exports. If your CTO wants receipts, send them there. If you want to know whether your company could do this, take the qualifier.