Tag: adr
RBX Journal articles filed under this tag.
- August 8, 2026
The reviewer who rejected itself
Two days ago we told the story of a data plan that shrank until only the essential remained. What came next was more interesting: building what survived under continuous adversarial review, with two models in opposing roles. A contract went through three verdicts, a wrongly numbered migration nearly broke the live database, an SDK was hiding the winner of a write race, and in the middle of it all the reviewer found a defect in text it had itself demanded. A story about how certainty is bought by the pass, and about knowing when to stop buying.
- August 6, 2026
The plan that shrank
We wrote a data platform plan for Robson: sixteen metrics, eight analytical tables, a warehouse, dashboards. We handed it to a second model with a one-line instruction: protect production. The verdict was NO-GO, and every cut it proposed survived verification against the code. What remained fits in five invariants, a two-table contract, and a rehearsal with numbers. The story of how a good plan gets better by shrinking.
- August 5, 2026
The first span
Yesterday we told the decision: one ruler for all of risk. Today it debuted in production, on a real position. This post checks the geometry number by number, walks the two seconds of events between the signal and the protective stop, and explains why the position size went down when the span went up.
- August 4, 2026
One ruler
The operator looked at a screen and asked a one-line question: why is the stop three times farther than the first target? The answer exposed two rulers living in the same system, an environment variable that turned one binary into two products, and ended in the largest engineering effort on Robson since the Rust rewrite. Including the day the fix was nearly worse than the defect.
- August 4, 2026
The missing slot
A month with zero governed losses, and the dashboard showed 3 slots out of 4. The number was right and the rule was wrong. How one operator question became the formal supersession of an ADR, why the obvious fix was unsafe, and what a second model found in our own text before approval.