Sensitivity — does the answer move?

The test

A matrix is twenty-odd judgements and a handful of weights, any of which a reader can dispute. Sensitivity analysis asks the question the reader will ask: if I change the thing I disagree with, does the answer change? The procedure: identify the most contestable weight — the one a stakeholder would argue about — change it (moving the difference onto another weight so the sum stays 1.00), and re-total. Then identify the most contestable score — the judgement with the least evidence behind it — move it by one, and re-total. If the winner survives both, the recommendation is robust. If it does not, the recommendation is not robust and must say so; that is a finding, not a failure.

Worked example — Lakeshore under three weightings

As built (cost 0.30, reliability 0.25, the rest 0.15): A 2.80 · B 3.05 · C 4.10 · D 2.40.

Equal weights (0.20 each) — the weighting a sceptic reaches for first: A 3.20 · B 3.00 · C 4.20 · D 2.60. C still leads, by a point.

Cost-dominant (cost 0.45, reliability 0.10, the rest 0.15) — Priya's weighting if cash is all that matters: A 2.95 · B 2.60 · C 3.80 · D 2.10. C still leads. Notice A overtakes B here — the ranking below the winner moves, which is worth a sentence in the memo, but the winner does not.

A contestable score. C's flexibility score of 4 rests on an upgrade clause not yet negotiated. Drop it to 2: C loses 0.30 and totals 3.80 — still ahead of B's 3.05. Verdict: robust. The lease leads under every tested weighting and survives its most doubtful score by a wide margin. Stage 4 can recommend it with that sentence attached.

Preview — when the answer does move

In module 11's Kestrel case, the weights are margin 0.25, quality 0.20, reliability 0.20, sustainability 0.20, fit 0.15, and the question is which supplier to drop. Raising sustainability to 0.35 and cutting margin to 0.10 moves one supplier — Tundra Line, which failed a labour audit — from 3.30 to 2.70. Its rank changes; it is still not last; and Cedar & Co. is last under every weighting, including equal weights (2.00). The recommendation there is robust for the question asked (whom to drop) and not robust for a question nobody asked (who is second-last) — and the memo should say precisely that.

Re-total the clinic's matrix under one change

Suppose the full scores give: paper 2.20 · reminder-only 3.10 · full booking 3.90 · hybrid 3.75, with no-show reduction weighted 0.30. The receptionists argue that ease of use should weigh more. Move 0.10 from no-show reduction (0.30 → 0.20) to ease of use (0.15 → 0.25) and re-total. Full booking scores 5 on no-shows and 2 on ease: it loses 0.10 × 5 and gains 0.10 × 2, net −0.30, to 3.60. The hybrid scores 4 and 4: −0.40 + 0.40, unchanged at 3.75. The hybrid overtakes. The honest finding is not robust: the choice between full booking and the hybrid turns on how much training the front desk can absorb, and stage 4's recommendation must carry that sentence and, ideally, a question back to Tomas.

State the test cold

Change the most contestable weight; re-total. Change the most contestable score by one; re-total. Winner survives both → robust. Winner changes → say so, say what it turns on, and say what evidence would settle it.