Limitations

Optional SciPy adapter

• Adaptive solve_ivp trajectories may vary within declared tolerances across SciPy/NumPy versions, compilers, operating systems, and architectures. Tolerance replay is not exact deterministic replay.

• A controlled t_eval grid standardizes retained comparison points but does not control or expose SciPy's internal adaptive step sequence.

• Event detection looks for sign changes across internal solver steps, so several zero crossings inside one step can be missed. PhaseProbe records event configuration and returned event states; it does not claim exhaustive root detection.

• PhaseProbe's portable state schema is real-valued. SciPy's method-specific complex-domain support is intentionally outside SolveIVPAdapter; split real and imaginary components.

• vectorized=True is a user assertion about the RHS contract and can help or hurt performance depending on method and system size.

• Solver success=True does not establish scientific accuracy. Users must declare invariant, convergence, endpoint, or qualitative policies appropriate to their model.

• A configured Python adapter module/factory executes trusted code during run and replay. PhaseProbe validates dotted names and never loads code during JSON parsing, but it is not a sandbox.

• Callable source is neither serialized nor treated as securely hashable. Model identity combines an explicit user identifier with serialized numerical configuration metadata.

• A reported bracket is finite-time, classifier-specific numerical evidence, not an exact bifurcation point.

• A smallest result is the smallest reproducible candidate found by the declared finite grid/refinement, not a proof of global minimality.

• The Lorenz metric is a finite-time divergence rate, not a Lyapunov exponent and not by itself proof of chaos.

• Fixed-step RK4 is transparent but not a substitute for adaptive solvers, stiffness detection, convergence studies, interval arithmetic, or rigorous numerics.

• Exact trace hashes can detect harmless floating-point differences across architectures or native math libraries. v0.1.0 has no configurable fuzzy replay mode.

• Only one scan parameter or one initial-state dimension is active per search. Non-monotonic multidimensional delta debugging is future work.

• Built-in models are small validation examples. Their success does not validate a downstream scientific model.

• Seeds are explicit, but an external adapter can still be nondeterministic through threads, native libraries, unordered data, clocks, file systems, or unrecorded external state.

• Trace retention is capped and may omit earlier behavior. Configuration must preserve a scientifically adequate observation window.

• Classification thresholds can create boundary ambiguity. The scan stops refinement when a midpoint matches neither stable endpoint class, but it cannot remove classifier bias.

• Replay fixtures can expose model configuration and state. Review them before committing if a downstream adapter contains sensitive data.

• Name research found no exact package collision on 2026-08-01, but it is not legal or trademark advice.