Capstone: One Customer, End to End

Every part of this path, applied to one customer. The customer is Lexa, the legal-tech company from the Field Engineering exercises: contract review on a closed frontier API, 6,000 to 14,000-token sections plus a 3,000-token instruction block, 400 to 1,200-token JSON outputs, 3 req/s average and 8 req/s peak during EU business hours, a nightly re-review of 40,000 sections, $60k/month and rising, EU residency and zero retention required, 150 expert-labelled sections, and a CISO who has not been involved yet.

Produce each artifact below. Each one has a check you can apply without an instructor. Keep them in .scratchpad/capstone/ (gitignored) or wherever you keep work; nothing here is committed.

#	Artifact	Draws on	Passes when
1	Discovery notes	Field Engineering 1	Every unknown in the brief is a question, each question names the decision it changes, and the missing CISO is flagged as a risk with an owner.
2	Model shortlist	LLM Foundations 3	Three open models, each with total and active parameters, attention variant, and context length from its config.json, and one sentence on why it can or cannot handle 17k-token inputs.
3	Checkpoint report	Frameworks and Models 2	config_explain.py and safetensors_inspect.py --url output for each shortlisted model, KV bytes per token recorded, and the chat template checked for a JSON-output instruction.
4	Sizing	Inference Performance 4, Field Engineering 2	capacity.py run for the chosen model and GPU; GPU count for 8 req/s at the stated lengths; separate sizing for the nightly batch; a serverless vs dedicated vs batch recommendation with the breakeven shown.
5	Benchmark plan and dry run	Field Engineering 3, Inference Performance 4	A plan using Lexa’s length distribution and arrival pattern, and a loadgen.py --mock --sweep run reported as a curve with goodput under the SLO. If you have a GPU, a real vLLM or SGLang run replaces the mock.
6	Eval plan	AI Full Stack 3	Metrics for clause extraction (per-clause precision and recall, JSON validity rate) on the 150 labelled sections, a parity threshold against the incumbent, and the sample-size caveat stated.
7	Fine-tune decision	Training 2	A yes or no on LoRA with 150 examples, the reasoning, and if yes, a memory_calc.py estimate of GPU-hours and cost.
8	POC plan	Field Engineering 4	Success criteria tied to artifacts 4 to 6, owners, a timeline, and exit criteria for both success and failure.
9	Migration plan	Field Engineering 5	Template and sampling parity checks, shadow traffic, canary, cutover, and a rollback trigger.
10	TCO and security brief	Field Engineering 6	12-month cost for current spend vs proposed, with the growth rate applied, EU residency and zero retention addressed, and the CISO meeting on the plan.
11	Escalation drill	Field Engineering 7, Inference Performance 1	A report for this symptom that the performance team could act on without a follow-up: “p99 TTFT jumps from 1.5 s to 9 s at 8 req/s when 14k-token sections arrive together.” Include your hypothesis (chunked prefill, batched-token budget) and the evidence that would confirm it.

Done when: all eleven pass their checks, and you can present artifacts 1, 4, 5, 8, and 10 as a 20-minute customer readout without reading from them.