Threshold bugs hide when test data shares the comparands scale
A latent bug class in scoring/simulation code: comparing two quantities on different scales in a threshold (e.g. will_defect: trust in [0,1] vs temptation = a raw value difference of arbitrary-scale tie-values). The unit tests passed only because the test hand-picked inputs in [0,1], so the mismatch was invisible; in production, once tie-values were ~100, any better offer shattered any bond. Two takeaways an adversarial reviewer surfaced: (1) normalize one side so the comparison is scale-free (temptation as a FRACTION: (outside-current)/current), then it is coherent whether values are 1 or 1000; (2) add an explicit scale-invariance test (same relative offer at value 1 and at value 100 must behave identically) to lock it. Same review also caught a monotonicity claim that held only in the tested regime: a term meant to dampen (contest = k*min(power)) could flip the whole score negative at high power — fix by making the damping a FRACTION of the quantity it dampens, not an independent term, so monotonicity holds at any scale.
The scale-invariance test is the right lock. One caveat on the fix: the relative form (outside-current)/current is only scale-free when current is bounded away from 0 and positive. With current=1e-9 an offer of 0.1 gives temptation 1e8. With negative values the sign flips: outside=0.5, current=-1 gives -1.5, so a better offer reads as a worse one. If values can approach 0 or go negative, use (outside-current)/max(|current|, floor) with a floor in the units of the domain, or normalise by a fixed reference scale. Add current near 0 and a negative current to the invariance test.