The work got better when the AIs stopped trying to agree
I have been building parts of a commercial intelligence system together with another team and another AI. What has surprised me is that the value does not prima
field notes / working evidence
Field notes.
Short notes from work in progress — what changed, what I learned, and what still needs to be tested.
We made the way of working persistent
For a long time, a meaningful part of the way I work with AI lived in conversation history: how we orient, choose context, pick models, verify work, escalate de
We stopped when more verification stopped reducing risk
Most of today was spent tightening a development execution gate with an independent AI reviewer trying to break it.
A safeguard turned out to be stronger than we expected
We wanted to verify that cross-tenant learning could not accidentally become approved global learning. The running system turned out to have no functioning appl
We tried to prove tenant isolation before tenants existed
We thought the next step in the learning architecture was to prove tenant isolation. Looking at the running system showed something more basic: there was no ope
A declined meeting does not necessarily mean the need was wrong
A declined meeting could mean wrong timing, wrong relationship, incumbent supplier, poor outreach, low priority or no need.
Model cost changed how we structured the work
AI usage increased without a proportional increase in useful progress. We changed the workflow from many small review loops toward larger coherent work packages
One opportunity should not become five observations
A single opportunity can produce several outcome records. If every event becomes a calibration observation, one real situation can create artificial evidence.
Signal volume became less interesting than signal convergence
More events did not necessarily make an account more interesting. Several independent observations pointing toward the same underlying change appeared more usef