GAO/NSIAD-92-340 graded its Patriot kill evidence in two tiers: a 'strong evidence' tier (~9%) and a non-probative 'lesser evidence' tier (~16%)

patriot-missilegulf-wargao-reportstatisticsverification-methodmilitary-engineering

GAO's evidentiary review of Army-claimed Patriot-Scud warhead kills did not produce a single number; it produced a two-tier structure. The top tier — "strong evidence," about 9% of claimed engagements — required corroborating physical or radar signatures (see claim-gao-nsiad-92-340-patriot-9-percent-strong-evidence-standard). Below it, GAO identified a second, weaker tier: "Another 16 percent of the engagements produced less convincing evidence of success, such as radar tracking data showing that the Patriots came close to the Scuds." Critically, GAO stated this second tier explicitly does not establish a kill: "This evidence does not prove, however, that the Patriots either destroyed or diverted the Scuds."

The two shares are not straightforwardly additive, and the report does not add them. GAO's summary paragraph states the strong-evidence share against one denominator — "About nine percent of Patriot launches produced strong evidence" — and the weaker tier against another in the very next sentence: "Another 16 percent of the engagements produced less convincing evidence of success." The findings section frames the same 9% differently again, as a share of "the engagements that the Army believes resulted in warhead kills." Nowhere does the report sum the two tiers or state a common base for them, so the inviting arithmetic (9 + 16 ≈ a quarter supported, three-quarters unsupported) is an inference the document does not license. What the two-tier structure establishes is definitional rather than arithmetic: it is the backbone of the "strongest evidence" framing used throughout the report and, downstream, in later citations of the ~9% figure — it explains why the number is a floor on confirmed kills rather than a point estimate of the true kill rate, since GAO explicitly declined to treat proximity data as proof.