Owning weights is not owning the workflow.

Video caption

Owning weights is not owning the workflow. The final effect can fail outside your model. Prove completion when a dependency disappears. #EricFieldNotes

Full written post / accessibility read

Open weights can remove one model-provider dependency. They do not automatically give you legal data rights, reliable compute, tool credentials, business permission or a human rescue path. Sovereignty has to describe the complete accepted task, not where one tensor file lives.

Imagine a local model drafting a case decision, but the source records sit in a hosted application and the action requires a remote API credential. When that API or role disappears, local inference cannot finish the job. The workflow is only as independent as its least replaceable step.

For each task, list where the model runs, where evidence lives, who can alter the service rule, which credential executes the action, and who handles exceptions. Test one dependency loss at a time against an accepted-outcome oracle.

Run a route-loss drill and measure which cases still reach an authorized result. Do this because a self-hosted model changes your control over inference, while the business remains dependent on every other step needed to act.

#EricFieldNotes

Four-beat scene transcript

1. Owning weights is not owning the workflow.

Open weights can remove one model-provider dependency. They do not automatically give you legal data rights, reliable compute, tool credentials, business permission or a human rescue path. Sovereignty has to describe the complete accepted task, not where one tensor file lives.

Visual: Data, compute and action rights may still be external.

2. A local answer may still rely on remote action.

Imagine a local model drafting a case decision, but the source records sit in a hosted application and the action requires a remote API credential. When that API or role disappears, local inference cannot finish the job. The workflow is only as independent as its least replaceable step.

Visual: The final effect can fail outside your model.

3. Map the whole dependency stack.

For each task, list where the model runs, where evidence lives, who can alter the service rule, which credential executes the action, and who handles exceptions. Test one dependency loss at a time against an accepted-outcome oracle.

Visual: Weights, data, compute, policy, credentials and people.

4. Buy sovereignty at the workflow boundary.

Run a route-loss drill and measure which cases still reach an authorized result. Do this because a self-hosted model changes your control over inference, while the business remains dependent on every other step needed to act.

Visual: Prove completion when a dependency disappears.

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.