You went to Spot VMs for the discount and did the homework: checkpoint code, a drain handler, a preemption survived. Then a job sat in Pending for days across three regions and never got a single L4. Scarcity grew a second half. Interruption is losing capacity you have; obtainability is failing to get capacity at all, and you can't checkpoint your way out of a stockout. Part 1 of the series that builds the probe the Compute Fallback Ladder left unbuilt.
I'm hiring a Senior Developer Relations Engineer for GKE and AI Infrastructure. Instead of describing the work, I built it: a small, playable cluster scheduler that teaches the real problem behind running AI at scale—gang scheduling, memory ceilings, spot preemption, and a TPU slice with actual topology.
The finale of the series: the mature end-state the first three parts build toward. When every server-side capability's trust boundary is a decision on record—the deployment tier you chose, the toggle you set, the contract you signed—saying yes to another team stops being a risk assessment and becomes a lookup. Fine-grained capability governance is what buys you safe expansion. Maturity, not lock-in. Boring, in the way a well-run system is boring.
Parts 1 and 2 argued that a capability's trust boundary is a separate decision from the model's. Part 3 is the runbook that makes the decision stick: the org-policy constraints that gate partner web search and structured outputs, set once at the organization tier by gcloud and Terraform, plus the seams org-policy doesn't reach—VPC Service Controls, request-response logging, the grounding-provider choice. Which toggle, at which scope, with deny-wins precedence, so a good-faith developer can't trip a data path nobody chose.
On February 23, 2026, Google quietly made the retention terms for Grounding with Google Search better: the window dropped from thirty days to up to three, and what's kept narrowed from your prompts and output to short-lived debug logs. The old thirty-day, prompts-and-output language didn't vanish—it now describes Grounding with Google Maps, a sibling capability under the same terms. Here's what the corrected terms actually say, how to verify the date yourself from Google's dated archive, and the durable lesson: govern facts that move—cite the dated primary source, hedge the number, and don't hard-code a digit that will drift out from under you.
The builder's companion to Governed Growth—the same reference repo from the implementer's chair. Why I built a one-tool MCP server to hand an assistant a reliable, governed Google search, how Gemini grounding and citation extraction work in the code, and how a full, human-validated eval picked the default model. The surprise: gemini-3.1-flash-lite matched the larger models on quality, grounded most reliably, and cost the least.
