A test that flags a behavioural assumption changed at a limit breach: every change at a breach flagged, with 17 to 66% false flags, on stated parameters, and weak ranking on re-estimated ones that a calibrated tolerance corrects

The review of the failure of Silicon Valley Bank found that it changed the assumed duration of its deposits when an interest rate risk limit was breached; no published procedure checks for that. We state a test that counts the dates at which a behavioural parameter changed during a breach, with a tolerance extension, and prove eight propositions. On a stated parameter it flags 100.0% of histories with a change at a breach date and 34.8% with unrelated changes, as 1-(1-λ)^m predicts; the Fisher, Mann-Whitney and Welch tests detect 13.2%, 48.0% and 26.8%. On a re-estimated parameter it flags 82.2% of null banks and ranks a 2-year lengthening at an ROC area of 0.579 (0.561 to 0.595). A tolerance calibrated at 5 % holds 3.9% false flags and ranks it at 0.937, but detects 18.0% of 2-year and 99.5% of 4-year lengthenings; a break test required to fall on a breach date detects 25.1% and 32.9% at 6.2% false flags. Calibrated at one prevalence, applied at another, the tolerance flags 5.5% (4.6 to 6.6 %) of null banks, within a calibration's spread, and the break test 7.8%, outside it. An override of the weights is detected weakly on simulated banks (area at most 0.651) and moderately on the real series (0.801, level shift). Three of nine predictions failed: the 5 % size on stated parameters, an area of 0.70 on re-estimated ones, and the tolerance's 5 % size at another prevalence. The bank's behaviour is synthetic: no result concerns what it did. Repository and reproducible pipeline (MIT): https://github.com/fsantibanezleal/CAOS_Contraste .