Requirements and quantified non-functional requirements

A landlord hires cleaners for a twelve-floor office. "Clean the offices, empty the bins, restock the washrooms" is the work. "Keep it spotless" is a wish. The contract instead says: 95 of every 100 spot checks pass, on all 12 floors, before 08:00 each weekday.

In the cleaning contract	In requirements
empty the bins, restock the washrooms	functional requirements (step 1)
"keep it spotless"	an adjective, not an NFR (step 3)
95 of 100 spot checks pass	metric and percentile (steps 3–4)
all 12 floors, before 08:00 each weekday	load and window (steps 3, 5)
an inspector checking one floor on a Sunday	a test at the wrong load and window (step 6)
a cleaner on call for 99.5% of opening hours	availability, turned into minutes without cover (step 7)
separate headings: hygiene, supplies, key security, spills	quality characteristics (step 8)
two cleaners off sick: window polishing waits, washrooms do not	which target gives way first (step 10)

In real terms: a requirement is only as good as the check that can fail it; the four parts make that check possible.

Breaks down when: a landlord's inspector sees every floor. A service's users see only their own requests, and one bad request in twenty never shows up on a dashboard average. Wrong inference: "if the average spot check passes, every tenant is happy." The tail is the percentile's job (step 4).

Contrasting case: a contract that only lists the tasks, with no spot checks. Everything is functional; nothing can be failed on quality.

4.1 ExamHall: status tiles that passed their check

EXH-1670: at 09:00 on ECON-201 day, proctors' candidate status tiles ran seconds behind. The release check had passed. The requirement said "under 2 s on average", and it was checked in a rehearsal with 240 candidates. MR !1870 rewrote it with all four parts. The samples below are synthetic, from a seeded generator.

rehearsal       n=240  mean 0.9 s · p95 (rank 228) 1.4 s · max 1.5 s
  old "under 2 s on average":           PASS
  new "p95 <= 2 s at 2,400, 09:00-09:10": NOT A TEST (load differs)
09:00 exam day  n=2400 mean 1.5 s · p95 (rank 2280) 6.5 s · max 9.0 s
  old "under 2 s on average":           PASS
  new "p95 <= 2 s at 2,400, 09:00-09:10": FAIL

The old wording passed twice: the average hid the tail (step 4), and the rehearsal ran at a tenth of the load (step 5). The new wording refuses the rehearsal as evidence and fails the real morning (step 6).

4.2 Proof domain, e-commerce catalog: nines as minutes

The catalog runs its search API at 99.5% and checkout at 99.99%. Allowed downtime is (1 − target) × window:

search API   99.50%  day: 7.2 minutes · 30-day month: 3.6 hours · 365-day year: 43.8 hours
checkout     99.99%  day: 8.6 seconds · 30-day month: 4.3 minutes · 365-day year: 52.6 minutes
99.99% over a 365-day year = 52.56 minutes; SRE book says 52.56: true

Each extra nine divides the allowance by ten, and the window decides the number (step 7). A target with no window has no allowance to check against. The last line reproduces the SRE book's figure.