Non-power or irregular behavior. Purely oscillatory forms (e.g., sin(1/x)) may not fit power or power*log; prefer REG=OSC or a conservative REG=MULTI.
Alternating dominance. If subsequences flip which term leads, do not force a single class; set REG=MULTI and summarize per window/side.
Degenerate coefficients. If a fitted leading c hits 0, promote to the next nonvanishing term; if unresolved, REG=NOFIT or REG=MULTI.
Essential-type tails (rare). Use guarded essential lenses only when power and power*log fail quality; disagreeing orders -> prefer REG=MULTI.
Noise and finite precision. Use same-window samples for f and g, robust fits, bootstrap if available, and tail checks (monotone-up for INF, vanishing for ZERO, shrinking deviation for FINITE).
Sidedness and oscillation. Run left/right; if classes or DIR differ, REG=SIDED. Persistent tail variance/sign flips -> REG=OSC.
Near ties and thresholds. When |p_f - p_g| or |q_f - q_g| is near threshold, annotate with EDGE(p) / EDGE(q) and prefer conservative headline (ZERO < FINITE < INF).
Alignment is metadata. Always clamp before rapidities:a := clamp(a, -1+eps_a, +1-eps_a) with eps_a = 1e-6; alignment never changes headline class.
Not a 0/0 case. If p = 0 for either side with nonvanishing constant, handle via ordinary limits outside the 0/0 scheme.
Higher dimensions / paths. In R^n, path direction can change effective rates; treat by directional slices; divergent classes across directions -> regime note (SIDED-like).
Units / rescaling. Rescaling changes c but not (p,q); headline is invariant. Alignment bands can shift with windowing or fit quality—publish your manifest.
Navigation
Prev: Classifying 0over0 Limits — Applications & Integration (9)
Next: Classifying 0over0 Limits — Reproducibility Manifest (brief) (11)
SHUNYAYA 