Tier-1 and Tier-2 Reproduction
Not every correct calculation is bit-for-bit repeatable. Kronos states, per result, which standard applies.
Deterministic, integer-and-float-stable calculations are graded Tier 1: on a matching software environment they reproduce byte-for-byte. Calculations that involve legitimate sources of numerical variation — iterative solvers, sampling, platform math differences — are graded Tier 2 and reproduce within a declared tolerance.
Grading each result this way is itself an honesty mechanism: it prevents a within-tolerance match from being over-sold as exact, and it tells a reviewer immediately whether a byte difference is a real discrepancy or expected noise.
Common questions
Tier-1 is byte-exact reproduction of deterministic results; Tier-2 is agreement within a stated tolerance where results are not bit-for-bit deterministic. Both are defined so 'reproduces' has a precise meaning.
Because honesty about what 'reproduces' means matters — some computations match exactly, others within tolerance, and Kronos states which is which.