Abstract
<jats:p>When 52 laboratories were sent the same two reference materials and asked to assess cytotoxicity under ISO 10993-5, only 58% classified both correctly, and viability values reported for a single polyvinyl chloride sample spanned the entire possible range from 0 to 100%. That result is usually read as evidence that the standard permits too much latitude. This review argues it is better read as the terminal symptom of an error chain that begins several steps earlier and has never been examined as a whole. Two substantial literatures address parts of the problem and do not reference each other. The first documents chemical and optical interference in viability assays, establishing that carbon nanotubes, silver and iron oxide nanoparticles, manganese, polydopamine, bacterial nanocellulose, lignin, flavonoids and cyclodextrins all corrupt tetrazolium or resazurin signals, in some cases by more than 78%. The second addresses statistical estimation from dose-response data, covering model selection, asymptote constraint, outlier handling and the confounding of potency estimates by division rate. The first stops at the absorbance reading; the second assumes the input is clean. This review traces a single measurement through six sequential layers and shows how error introduced at each layer propagates forward: biological input variability, plate-level systematic bias, assay chemistry interference, normalisation choice, curve fitting, and threshold application. Two properties of the chain matter. Systematic biases confined to specific wells distort curve asymptotes rather than shifting the curve uniformly, so a 35% edge-well depression does not produce a 35% IC50 error but an unpredictable one that depends on plate layout. And a binary threshold at 70% viability converts continuous measurement error into a discrete regulatory outcome, with maximum sensitivity precisely where materials cluster. A Monte Carlo simulation using published effect magnitudes quantifies the consequence. Plate layout, not effect size, dominates: the same measured edge effect produced IC50 biases of 18%, 39% and 57% depending only on where the concentration series sat relative to the control wells, and the conventional mitigation of excluding perimeter wells from the series made matters worse by widening the gap between test and control positions. In the worst configuration a genuinely cytotoxic material was classified as non-cytotoxic in every one of 3,000 simulated experiments, while matching control position to test position reduced bias to 2.4% on the same plate. Growth rate inhibition metrics, the most widely adopted correction, address division rate alone and leave every upstream layer untouched. A minimum reporting framework is proposed.</jats:p>