FDE Practice: Deployment to Renewal
All 16 cards in this deck
Scoping, evals, identity, on-call and the commercial layer — the facts that decide whether a deployment survives. Answer each question before you open it — then run the review above, which schedules every card with spaced repetition so it comes back right before you would forget it.
What are the six axes a scoping session must cover?
Outcome, data, users, constraints, success, risk. An axis you never asked about is where the surprise comes from — usually data or risk.
What makes an assumptions block in a SOW actually protect you?
One line: "delays in the above shift milestone dates day-for-day." Without it, the assumptions are decoration and their nine-week delay is your nine-week overrun.
How do you set a defensible accuracy target?
Measure inter-expert agreement first — how often two of the customer's own experts agree on the same 40 cases. Set the target relative to that ceiling, with a separate cap on high-cost errors.
Why must a golden set be frozen and stratified?
Frozen so scores are comparable between runs; stratified so rare high-cost cases are visible rather than averaged away by the easy majority.
What must be reported beside every LLM-judge score?
Its measured agreement with human graders. A judge is a measuring instrument; without a calibration figure you cannot tell a real regression from a judge quirk.
Which CI assertion catches the failure an average hides?
The per-case regression list being empty. An overall score can improve while the rare, expensive cases silently break.
In SCIM, how does offboarding usually arrive?
As PATCH with active=false, not DELETE. Most identity providers deactivate rather than delete, so a DELETE-only implementation offboards nobody — silently.
What should user records be keyed on?
The identity provider's immutable subject identifier (oid in Entra, persistent NameID in SAML). Email changes on marriage, rebrand or acquisition.
Why map directory groups to roles through a table?
Because their groups get renamed in reorgs nobody tells you about. Indirection turns a Monday-morning outage into a config change their own admin makes.
What is the SLA trap that catches everyone once?
A response-time commitment whose clock starts when a user noticed rather than when you were paged. Get the reporting channel into the same paragraph as the response time.
What is the only honest measure that a handoff is real?
The share of incidents resolved without you, month over month. Flat at zero by month six means the handoff is not happening, whatever the documentation says.
What is the services trap?
Deployment work grows, every engagement is bespoke, nothing is reused, and a software company becomes a consultancy with worse margins — without deciding to. Diagnostic: is deployment N faster than N−1?
What is the generalisation memo?
One page at the end of every deployment: what we built, what was genuinely customer-specific, what should be product, and what that would cost. It is what makes deployment work compound.
When does work on a renewal start?
About nine months before it. By the time procurement schedules the meeting, the outcome is set by whether the system is used, whether the champion is still there, and whether anyone can state the value as a number.
Why is escalating past a data gatekeeper usually wrong?
It gets access in about a week and an obstacle for eighteen months. They will not refuse anything again — they will simply be thorough, on their board's timeline, forever.
What is the strongest move in a security questionnaire?
A specific no with a compensating control, an owner and a date. It outscores a vague yes every time, and volunteering a gap raises your score because finding it is the reviewer's job.