Write the use-case record that could have stopped the project, and triage five proposals

short · 35 min · Objective 1.3

Task

Start the life cycle where the lesson starts it: before any data or model. Write a use-case record for your lab assistant with a named owner, its data, its agency, the people it affects and kill criteria you could actually measure. Then apply the lesson's stop rules to five proposals, and lay out the gates your assistant must pass, each with an owner and the evidence it needs.

Steps

  1. Write ~/secai-lab/lab/use-case.md for the lab's HR assistant with exactly six second-level headings: ## Purpose and owner, ## Data, ## Agency, ## Affected people, ## Kill criteria and ## Decision. Name the owner by role, not by name.
  2. Under ## Kill criteria, write at least three bullet points, each with a number you can measure in this lab, for example: any restricted chunk retrieved for a non-HR user; more than 10% of triage replies outside the allowed set; any planted instruction obeyed in evaluation.
  3. Write ~/secai-lab/lab/use-case-triage.csv with header id,proposal,verdict,reason, verdict being proceed, redesign or stop, for: P1 a support chatbot answering from the published FAQ; P2 the same chatbot issuing refunds without review; P3 an LLM deciding whether an address is on a fixed blocklist; P4 a model trained on call recordings that customers were told would be used only for quality checks; P5 a pilot whose sponsor leaves next month with no team to run it.
  4. Write ~/secai-lab/lab/gates.csv with header stage,owner,evidence,criterion and one row for each stage: use-case, collection, preparation, selection, evaluation, deployment, validation, monitoring, maintenance, feedback. evidence names a file under ~/secai-lab/lab/ that does or will hold the proof, such as manifest.csv for collection or template-versions.csv for evaluation.
  5. In ## Decision, state whether the assistant proceeds, and which kill criterion your earlier labs have already measured as failing, if any.

Verify

cd ~/secai-lab
grep -c '^## ' lab/use-case.md
awk '/^## Kill criteria/ {k=1; next} /^## / {k=0} k && /^[-*] / && /[0-9]/ {n++} END {print n+0" measurable kill criteria"}' lab/use-case.md
awk -F, 'NR>1 {print $1"="$3}' lab/use-case-triage.csv | sort | paste -sd' '
awk -F, 'NR>1 && $2!="" && $3!="" {n++} END {print n+0" gate(s) with owner and evidence"}' lab/gates.csv
awk -F, 'NR>1 {print $1}' lab/gates.csv | sort -u | wc -l

Six headings; at least three measurable kill criteria; triage P1=proceed P2=redesign P3=stop P4=stop P5=stop, where P4 may defensibly read redesign if your reason names a lawful basis you would obtain; ten gates with an owner and evidence; and ten distinct stages. A gate with no owner is a stage nobody will stop.

Notes

P2 is redesign rather than stop: the same chatbot becomes acceptable with a human approving refunds over a threshold, which is human-in-the-loop from the same lesson. P3 is the one people miss, because a deterministic lookup does the job better than any model. The capstone fills in the evidence column for real.

This is an independent study companion for CompTIA SecAI+ CY0-001 and is not produced by or endorsed by CompTIA.