Built in my own lab first, then opened to owners.

01 The cases that wrote the rules

Four cases from my own lab, each with the rule it produced

I built these systems for my own work before I offered them to anyone. The failures below are the ones that mattered, dated, with what each one changed. The full account is in the field guide.

• July 8, 2026
  The capability that was never thereWhat an owner would have seen: a feature that reported zero on every run, with no error and no alarm. Underneath, two agents relied on a piece of code from the day they were written, and that code did not exist. The rule it produced: a capability counts as present only when it has run once and been read back.
• August 17, 2026
  Three lies at onceWhat an owner would have seen: a setup job that read like success. It had been refused, and three things hid it at once: an error nobody saw, a success message that printed no matter what, and a self-check that never looked at what it claimed to check. The rule: success is a read-back, never a print. A pass names what it scanned, and a pass over nothing fails.
• September 14, 2026
  Eight agents listening, one clockWhat an owner would have seen: a board with eight agents busy at work. All eight stamps carried the same second, because one script had written one time eight times. The rule: count outcomes, never heartbeats. A system is alive when it has done something a person can point to.
• October 1, 2026
  Healthy, and two weeks behindWhat an owner would have seen: every health light green on a knowledge index that knew nothing newer than two weeks. The rule: a liveness check never stands in for a freshness check. Ask the system about something it should have learned yesterday.

02 What the cases produced

Eight rules in writing, five status words, one chained record

Every system I install ships with the eight rules written down, the same eight my own systems run under. You can score any AI against them on the owner's checklist. My systems are allowed five status words, computed by a machine from evidence, and each outcome goes into a chained record that an edit after the fact would break. How that holds on this site is on the trust page.

03 The practice, by date

From my own lab to a door owners can walk through

• September 29, 2026
  The practice opens its door as Custom AI SystemsSales pages only, written for owners of companies of 5 to 100 people. Every build starts with a free one-page Blueprint, and every rule on the site had already been tested on my own systems.
• October 8, 2026
  The agent door opensAn AI assistant can request a Blueprint for the owner it helps, with that owner's permission, straight to the intake. The same day every page gained a copy written for machines, and the trust page began answering the three questions owners ask before any work starts.
• October 9, 2026
  The cost reading goes publicThe meter on my own systems, read from September 8 to October 9, 2026: ninety-seven of every hundred tokens the models read came from memory. The method and the reading are on the cost page.

04 What stays the same

The same engineer, the same rules

One engineer designs, builds and cares for every system, and stands behind it from the first walkthrough to the monthly tune-up. Who that is, and how to reach him, is on the team page. What that buys you is on why us.