Ceilings you can defend, a ranked map of how each part fails, and one runbook entry a teammate could follow at 3am. The runbook is finished in Week 5.
What this is
The least glamorous, most career-defining document.
Ops is where LLM features live or die. This artifact puts numbers on your hot path and names how it breaks, ranked by pain. Pairs with Lab 3 if you want to build the wrapper.
💡 In plain English
A budget so you know what "too expensive" means, and a map of what breaks so you are not debugging live for the first time at 3am. Depth beats breadth here: one great runbook entry, not ten vague ones.
Do this, in order
The drill.
1
Budget your hot pathFill the ceilings (cost and latency, typical and worst), find the dominant cost, tick the levers you'll switch on.
2
Walk the request, list failuresAt every hop, retrieval, model, tool, ask: slow, wrong, or down? One row each: component, failure, blast radius, mitigation.
3
Rank by blast radiusSort so the most painful failures sit at the top. A cosmetic glitch and a double-charge are not equal.
4
Write one runbook entryTake your worst failure and write symptom → check → act → prevent, executable by someone who didn't build it.
Fill this in
The budget + one runbook entry.
Request
the hot path: …
Cost ceiling
typical < $… · worst case < $…
Latency ceiling
p95 < … ms · first token < … ms
Dominant cost + levers
dominated by … · levers on: cache / route / right-size / stream / batch
Ceilings are real numbers, typical and worst case.
Failures are ranked by blast radius; each has a specific mitigation.
The runbook entry could be executed by someone who didn't build it.
✕ Decorative
"Be fast, be reliable" with no ceiling.
A flat list where a glitch sits next to a double-charge.
A runbook that says "investigate and resolve".
Wrapping up
The budget, the ranked map, one runbook entry.
It is fine, expected even, that the runbook has a single entry, it is a seed you grow in Week 5 from real traces. The map tells Week 5 exactly what to build evals and alerts for.
Deliverable
Budget table + ranked failure map + one runbook entry. Due Sun Aug 2.
submit on Maven
Stuck?
Use the S5 cheat sheet pricing table to price your worst-case request. A defensible estimate beats a blank.
rewatch S5 + S6
For the instructorHow to teach thisclick to open ▾
The ranking is where the value is, sorting by blast radius tells them what to fix first and what to ignore. And the runbook bar is concrete: a teammate who didn't build the system could follow it at 3am. If they couldn't, it isn't done.
Reassure them a single runbook entry is the goal, not a coverage sprint. Depth over breadth. The prevent line is what turns an incident into a permanent fix, and it is the one people skip.