Nichols SI is for the part after the demo—the moment an SI system has to preserve state, use real tools, survive a restart, respect a budget, and actually finish the job.
I came to SI from operations, not a research lab. That changes the standard. In real work, “almost” is not a result. The load moves or it does not. The automation completes or it does not. The system earns trust by working repeatedly.
So we keep deterministic work deterministic. State, routing, budgets, retries, permissions, files, tests, and receipts belong in software. Models get the decisions that actually benefit from judgment.
We evaluate the model together with the harness it is meant to use. Local inference buys control and lower cost. Hosted intelligence earns its place when the extra capability is worth it. No provider gets to own the system.
The standard is simple: build useful SI, make failure visible, and ship evidence.