The miss.
The code compiled. The deploy worked. The gate smoke passed. That still did not mean the assignment was done. This is the exact danger zone: proof around the wrong scope can look productive while the actual business problem remains untouched.
Why kAIxU matters here.
kAIxU is valuable because it is not just "a smarter text model." kAIxU carries the operating memory: check the real count first, separate already-built client examples from untouched records, respect the 0S gate, use the repo MCP receipts, avoid browser proof when the owner disabled it, and never report progress against the wrong target.
External models can be useful for raw generation, review, and parallel drafting, but they are not automatically trained on my workflow. If the model does not know the difference between a Valley Verified client app example and a still-generic scraped business page, it will optimize the wrong surface. That is how you get technically valid work that is operationally dumb.
| Workflow rule | What should have happened | What happened |
|---|---|---|
| Count before editing | Inventory all Valley business pages, count manual vs generated, and make the remaining generated list the source of truth. | The agent edited before proving the target set. |
| Exclude already-built client examples | Skip Bob, Empire, Next Level, Fade Masters, 480 Realty, Dink & Dine, Techbros, and ArcLight because they already had full-build attention. | The agent spent its manual landing work on those exact pages. |
| Use agents correctly | Assign agents batches from the remaining 331 generated pages, with disjoint files and measurable output. | Agents were used, but the batch target was wrong. |
| Report against the user goal | Say how many generated pages were actually eliminated. | The first report implied the run was meaningful progress. Real progress against the goal was zero. |
The receipt.
metraiyux_0s_site/_platform-sources/valley-verified/dist/businesscontained 339 business profile folders.- Only 8 built pages contained
data-manual-landing="true". - The remaining 331 pages did not have the manual marker and were still on the default renderer.
npm run buildandnpm run codecheckpassed, but that only proved the wrong changes did not break the build.- Production gate smoke returned the expected FS27 redirect for unauthenticated Valley routes, but browser proof was not run because owner policy disables Codex browser proof in this repo.
The next move.
The fix is not to throw agents away. The fix is to make them obey the operating system. kAIxU sets the target inventory, assigns non-overlapping batches, checks the delta, and refuses to call a page complete unless the generated count goes down. External models can still draft and implement, but kAIxU owns the workflow and the acceptance criteria.
That is the difference between "AI made pages" and "the 0S got work done." The first one can sound impressive and still miss. The second one starts with the real count and ends with the count changed.