Five in a Hundred: The Threshold Nobody Derived

Every working science keeps a ledger of evidence, and at the bottom of nearly every page sits the same entry fee: p < 0.05. This paper audits where that fee came from. Fisher, the threshold's founder, printed it as a convenience — 'usual and convenient,' a 'conventional line,' a point it is 'convenient to take' — alongside a sliding scale (below .02 strong; .1 to .9 no suspicion) that contradicts any fixed wall, and in 1956 he wrote that no scientific worker has a fixed level of significance. The antecedents in the provenance record (Cowles and Davis, 1982) are probable-error folklore: Student's three probable errors (1908, tail .043 when converted) and Wood and Stratton's thirty-to-one odds (1910, tail .032). The repair attempts are censused and priced: the ASA's 2016 statement names the misuse and deliberately offers no replacement number; Benjamin et al. (2018) propose 0.005, concede in print that any particular threshold is arbitrary, and support theirs with a Bayes-factor calibration (0.05 corresponds to factors of 2.5-3.4, weak; 0.005 to 14-26, substantial to strong) and a false-positive-rate argument (over 33% at prior odds 1:10, down to about 5%), at a stated cost of about 70% larger samples. A companion row prices Cohen's conventions (power 0.80; effect sizes 0.2/0.5/0.8, offered 'only when no better basis' exists). The honest contrast is Neyman-Pearson, where the error rate is a design choice made per problem by the investigator. The defect charged is narrow: not that 0.05 exists, but that it is spent as if derived. Grade: CONVENTION STANDS — IMPORTED, UNDERIVED, PRICED.