FIELD NOTE / INSTAGRAM
Run four policy-change drills.
The complete written thought and the evidence behind it. The video edition will follow its public release.
The written argument is here.
This approved Instagram edition is on the journal now. Its video player and original platform link will appear after each public release is verified.
Run four policy-change drills.
Video caption
Run four policy-change drills.
A green status can hide an unresolved case.
Inspect the persisted case after every injected result.
Incomplete model work stays incomplete.
#EricFieldNotes
Full written post / accessibility read
The provider route is a moving part. A model can refuse, time out, change an error code or return an answer that is structurally valid but missing required evidence. Each condition should create a known business state.
In a test queue, a report has a known correct owner. The model refuses, and the adapter emits an empty object that fits the expected schema. The ticket writer accepts it and closes the report. The schema did its job; the business contract did not.
Force one refusal, one timeout, one unknown provider code and one evidence-free valid JSON result. For each, require a named owner, open or review state, bounded retry, and no effectful action. A separate test process reads the persisted case.
Run these four tests whenever the adapter or route configuration changes. Do this because uncertainty must remain visible until a person or authorized process resolves it.
#EricFieldNotes
Four-beat scene transcript
1. Run four policy-change drills.
The provider route is a moving part. A model can refuse, time out, change an error code or return an answer that is structurally valid but missing required evidence. Each condition should create a known business state.
Visual: A single happy prompt cannot validate continuity.
2. The dangerous result is false completion.
In a test queue, a report has a known correct owner. The model refuses, and the adapter emits an empty object that fits the expected schema. The ticket writer accepts it and closes the report. The schema did its job; the business contract did not.
Visual: A green status can hide an unresolved case.
3. Give each drill an external oracle.
Force one refusal, one timeout, one unknown provider code and one evidence-free valid JSON result. For each, require a named owner, open or review state, bounded retry, and no effectful action. A separate test process reads the persisted case.
Visual: Inspect the persisted case after every injected result.
4. Fail on an invented default.
Run these four tests whenever the adapter or route configuration changes. Do this because uncertainty must remain visible until a person or authorized process resolves it.
Visual: Incomplete model work stays incomplete.
Research and claim limits
- OpenAI Usage Policies (S133)
- Anthropic Usage Policy update (S134)
- OpenAI structured outputs explanation (S08)
The examples identified as illustrative or simulated are design probes, not reported incidents. Vendor specifications do not establish workload performance.