Falsification-First Decipherment: A Decontaminated Inference Framework for Testing Undeciphered-Script Claims, with Linear A as a Worked Null

Decipherment research has a false-positive problem and no shared instrument: a small corpus, an unknown language, and many candidate readings let an analyst "translate" almost anything. The flagship survey lists fifteen challenges of script and data; none addresses the orthogonal inference-control layer — overfitting, multiple testing, contamination — a candidate sixteenth challenge. We build a falsification-first framework for it: a fabricated-language negative-control family, multiple-comparison deflation, a literature index separating published from not-indexed signs, and pre-registration with a mechanically-graded gate. This is a methods paper; we assert no decipherment. The machinery calibrates: on a tested null the gate's false-graduation rate is 0.6% (3/500); the identical morphology test finding no power on Linear A recovers Mycenaean (Linear B) inflection (9/16 affixes above the bigram floor), supporting the interpretation that Linear A's short forms contribute materially to its lack of power. Across probes it returns differentiated, calibrated outcomes — null, no-power, supported-structural, circular, and contamination-sensitive. Results are consistent with identifiability, rather than corpus size, being the more fundamental present constraint. The June-2026 *301 "cracked" claim is a qualitative illustration of the failure modes the gate is designed to detect.