Learning objective
By the end of this lesson, you can build a complete pairwise matrix, preserve one comparison basis and interpret preference cycles.
Why it matters
People find “Which is best?” difficult when a list is long. Comparing two options is easier, but the result becomes meaningless if “best” changes from impact to cost to personal preference between pairs.
Core concept: one question for every pair
Define one comparison sentence:
Which option better satisfies [criterion] within [horizon], based on [evidence rule]?
Compare every unique pair once. Record the preference, rationale and confidence. With six options, there are fifteen pairs.
Visual explanation
Text alternative: one criterion governs every pair. Recorded judgements create a ranking, which is checked for cycles before a next action is authorised.
Worked example
Five automation ideas are compared on “largest verified reduction in customer waiting within eight weeks.” Ticket classification ranks first, but the decisive comparison depends on a vendor estimate and receives low confidence.
The ranking authorises a shadow-mode test, not deployment. Weak evidence remains visible beside the first-place result.
Common mistake
Mistake: deleting a cycle such as A > B, B > C and C > A.
The cycle may reveal changing criteria, context dependence or insufficient evidence. Investigate it before forcing an order.
Quick check
Why should “insufficient evidence” be an allowed response?
A. To avoid completing the matrix.
B. To prevent arbitrary preference from becoming hidden decision evidence.
C. To guarantee a tie.
D. To let AI fill the answer later.
Answer: B. Forced certainty corrupts the ranking.
Practical prompt
Take four comparable options and one criterion. Generate six unique pairs, record each reason and flag any cycle or low-confidence comparison.
Summary
- Use one criterion for every pair.
- Compare comparable options only.
- Record evidence and confidence.
- Treat cycles as diagnostic signals.
- Define what the ranking authorises.
Next lesson
Continue with Preserve Disagreement in Delphi, where several disciplines estimate an uncertain AI outcome independently.