A blinded discrimination test in which assessors receive three coded samples—two alike and one different—and identify the odd sample to determine whether products are perceptibly distinguishable.
A triangle test answers a narrow question: can assessors perceive a difference between two products?
Purpose
Determine whether two samples are perceptibly distinguishable under blinded, controlled conditions. The test does not by itself identify the cause, direction, quality, or consumer importance of a difference.
Inputs
- Two products or formulations to compare
- Three coded samples per assessor: two alike and one different
- Randomized serving orders
- Standardized glasses, volume, proof, temperature, and room conditions
- A predefined interpretation plan
Sequence
- Define the comparison and decision threshold before tasting.
- Prepare samples consistently and code them without revealing identity.
- Randomize the position of the odd sample across assessors.
- Ask each assessor to identify the sample that differs.
- Record responses independently.
- Compare correct identifications with the chosen statistical criterion.
- If a difference is detected, use descriptive or preference testing to investigate its character or value.
Outputs
A documented discrimination result, not a quality ranking.
Controls and cautions
Panel size, randomization, serving order, fatigue, sample carryover, statistical threshold, and assessor independence determine whether the result is interpretable. Use trained sensory or statistical guidance for consequential decisions.
Evidence
Matt Strickland, Cask Management for Distillers, PDF pp. 186–198 / ebook pp. 155–161. The source establishes the practitioner test structure; controlled implementation requires an explicit statistical plan.
Statistical correction from the fresh source audit
The source's non-significance/equivalence interpretation on PDF pp. 188/195 is incorrect. Report a failed discrimination test as insufficient evidence of a difference under this design. Establishing similarity requires its own effect boundary and power plan. Preselect alpha; do not interpret it as the probability that a retained null is false. Repeated judgments must not be assumed independent. Table 7's printed 30% column contains 14 thresholds one below an exact binomial calculation; do not copy it. The correct upper-tail p-value for 24 correct of 30 with chance 1/3 is approximately 2.09016 × 10^-7. See 
