Modern natural-language systems determine the meaning of a word or a sentence statistically, by training large language models on extensive corpora and massive graphics-processor(GPU) resources. In the present work we pursue an altogether different route: we ask whetherthe meaning of written text is a physical quantity that can be registered directly by a torsion field detector, without any training of a statistical model. The study is carried outwithin the framework of torsimetry developed by Shkatov, in which the torsion field (TF)of physical objects and processes is recorded as the torsional contrast (TC) of the measuredobject with respect to the measuring instrument by means of single- and multi-coordinatetorsimeters. To suppress the spurious informational phantoms that otherwise accumulate inobject—instrument—operator systems and smear the readings of repeated measurements, weadopt the Method of Differential Test Assertions (MDTA), which composes the object undertest from pairs of textual assertions of opposite meaning and records their differential torsion response. We show that this differential response is a stable and reproducible signature of the semantic content of the text, and that the detector is able to discriminate the meanings of different words and sentences without any GPU-based training. In 19 independent measurement sessions, an impossible assertion (a total score greater than a physically unrealisable threshold) produced a consistently higher torsional contrast than the correspondingtrue assertion, with a positive sign in all 19 sessions and a mean difference of +112 arbitraryunits (p ≈ 2 × 10−6 ). The results support the hypothesis that the meaning of a word or asentence is an objective and stable property of the text, carried by its torsion field, and thatit can therefore be read directly from the instrumental signal rather than inferred from atraining corpus.