Thirty years of kinetic studies on tRNA selection in the elongation cycle are reviewed, and confronted with results derived from various sources, including structural studies on the ribosome, genetic observations on ribosome and EF-Tu accuracy mutants, and codon-specific elongation rates. A coherent framework is proposed, which gives meaning to many puzzling effects. Ribosomal accuracy would be governed by a "double-trigger" principle, according to which the ribosome uses energy in the forward direction to create new configurations for tRNA selection, and energy in the backward direction to regain its initial configuration, in particular after a premature dissociation event. The conformation energy would come in part, in Hopfield's mode, from GTP cleavage on the ternary complex (TC). The reset energy would be provided in part, in the author's mode, from GTP cleavage on a binary EF-Tu.GTP complex (BC). There would be several paths for amino acid incorporation. The path of highest accuracy would involve TC binding followed by BC binding, followed either by GTP hydrolysis on the TC, or by TC dissociation and GTP hydrolysis on the BC. Codon-anticodon recognition would occur in at least three kinetically and geometrically distinct stages. In a first stage, there would be a very rapid sorting of the TCs with unstrained anticodons contacting a loosely held mRNA. This stage ends with the anchoring of the codon-anticodon complex by a cluster of three nucleotides of 16S RNA. The second stage would be the most discriminative one. It would operate on the 5 ms time scale and terminate with GTP cleavage on the TC. The third stage would provide a last, crude selection involving "naked" aa-tRNA, partially held back by steric hindrance. Streptomycin and most EF-Tu mutants as well as high accuracy ribosomal mutants would produce specific alterations at stage 2, which are mapped on the stage 2 kinetic mechanism. The ram ribosomal ambiguity mutants, and anticodon position 37 modifications could be markers of stages 1 and 3 selection. Dissociation events at stage 2 or stage 3, when they are not immediately followed by reset events create a leaky state favorable to shortcut incorporation events. These events are equivalent to an "error-prone codon-anticodon mismatch repair". From the recent evidence on ribosome structure, it is conjectured that the L7/L12 flexible stalk of the large ribosome subunit acts as a proofreading gate, and that the alternation of its GTPase activation center between "TCase" competence and "BCase" competence is a main factor in the control of accuracy.