Recommendations are written per species; evidence arrives per accession. The roll-up takes each accession's longest passing year (only accessions with ≥2 distinct test years get a vote — one test is an anecdote) and reports the species interval as the maximum across accessions: the best survival anyone has demonstrated for this species.
Maximum is optimistic by construction — one exceptional accession sets the number. The qualification travels with it: supporting_accessions counts how many accessions independently reached the interval, and support_fraction divides by the total. "18 years, 4 of 6 supporting" and "18 years, 1 of 9" print the same interval and mean very different things.
The alternative — a median or a model across accessions — answers a different question ("what should we expect?") than the max ("what has been proven possible?"). For scheduling recollection of rare plants, proven-possible plus an honest support count is the operationally safer pair.