Applied Forensic Epistemology: Weighing Mediated Evidence with the Instruments of Forensic Proof

This paper defines applied forensic epistemology and builds its apparatus. Applied forensic epistemology is the branch of applied epistemology that runs the calibrated instruments of forensic proof on evidence that reaches the analyst through intermediaries. Its unit is one commitment to a referent outside the claim. That commitment is open to attack at three points: whether the referent exists, whether it can be reached, and whether it binds. Three tiers grade single claims as verified, inferred or open. A fourth tier, established, grades composite propositions. One engine joins three methods: likelihood-ratio weighing, structured comparison of competing hypotheses, and graded confidence assessment. The apparatus faced a pre-registered convergence test in two cycles. In the first cycle, three blind coders from three model families scored fourteen evidence-weighing disciplines. The cycle confirmed its hypothesis. Each of seven structural elements carried three or more independent witnesses at instrument level, meaning fields whose own tools use the element. On the strongest element, the likelihood ratio was about 1.2 × 10¹⁰, twelve billion to one. A red team upheld that result. Three language-model families nominated the second cycle's eleven fields. Each family was asked which traditions most explicitly reject the apparatus's premise. The project chose none of the fields. In those fields the instrument found the elements as shapes, not instruments: a mean instrument count of 0.73 to 1.36 of the seven elements per field, against the measurement disciplines' 3.21 to 3.86. That result matches what was sealed before reading. The second cycle is therefore read as the test's negative control, and the control passed. Both pre-registered discriminators print as sealed. Their inferences are withdrawn as declared corrections. Twelve cells are still at the range, and settling them cannot change the reading (§6). A negative control on a coding instrument found the instrument firing where it should not. A first round fired at two of five features. Both definitions were tightened, a declared change. A second round, across three model families, fired at three of five features, two of them under two families. The paper claims the apparatus and its own measurements, not the findings of the larger study from which it is drawn.