FIELD NOTE / INSTAGRAM
Sovereignty has six layers.
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.
Sovereignty has six layers.
Video caption
Sovereignty has six layers.
An answer without an authorized effect is incomplete.
Record owner, substitute and recovery test.
One local component does not make the chain independent.
#EricFieldNotes
Full written post / accessibility read
When someone says 'we own the AI,' ask what they own. Model weights are one layer. Data rights, compute capacity, service policy, effectful credentials and human escalation determine whether the organization can actually keep a process running.
A local classifier can label a document. If the team cannot export its source records, or a separate cloud system owns the final approval button, the end-to-end task still depends on someone else. The architecture diagram must include those boundaries.
For a representative case, make six rows: weights, data, compute, policy, credentials, people. In each row, name the change authority, what breaks if it vanishes, and a lawful alternate. Then test the highest-risk loss in a disposable environment.
Show which layer you can operate and restore yourself, and which still needs a partner or person. Do this because 'we run the model' is narrower than 'we can complete the business process.'
#EricFieldNotes
Four-beat scene transcript
1. Sovereignty has six layers.
When someone says 'we own the AI,' ask what they own. Model weights are one layer. Data rights, compute capacity, service policy, effectful credentials and human escalation determine whether the organization can actually keep a process running.
Visual: A local model can leave five dependencies untouched.
2. The failure moves to the next layer.
A local classifier can label a document. If the team cannot export its source records, or a separate cloud system owns the final approval button, the end-to-end task still depends on someone else. The architecture diagram must include those boundaries.
Visual: An answer without an authorized effect is incomplete.
3. Score each layer against a real task.
For a representative case, make six rows: weights, data, compute, policy, credentials, people. In each row, name the change authority, what breaks if it vanishes, and a lawful alternate. Then test the highest-risk loss in a disposable environment.
Visual: Record owner, substitute and recovery test.
4. Claim only the control you can demonstrate.
Show which layer you can operate and restore yourself, and which still needs a partner or person. Do this because 'we run the model' is narrower than 'we can complete the business process.'
Visual: One local component does not make the chain independent.
Research and claim limits
- OpenAI Usage Policies (S133)
- Anthropic Usage Policy update (S134)
The examples identified as illustrative or simulated are design probes, not reported incidents. Vendor specifications do not establish workload performance.