Exact edit-distance baseline and source activation guards

This report uses RapidFuzz 3.14.6 in a fresh isolated Python 3.12 environment. It does not execute the repository's approximation algorithm. Runtime identity, installation wheel provenance, comparison data and seven source hashes, reproducible supporting script.

The literal source approximation branch requires ell >= 2^1000. From the selected integer definitions, the least length satisfying that guard alone is 2^(2^(2^999))-1. Other accuracy and bit-width conditions remain additional. The actual tower was never allocated; scaled boundaries were checked for exponents 1 through 4. At total length 10^12, ell=8. The source falls back to an exact answer below its physical guard. No source fallback timing was measured.

Generated existing-library comparisons

Symbols per string	Generated pair	Known exact distance	Median unrestricted call, ms	Cutoff 16 return	Cutoff call, ms
4,096	equal	0	0.001708	0	0.001583
4,096	sparse_substitutions	4	0.230833	4	0.007792
4,096	shifted_periodic	2	0.405666	2	0.008334
4,096	all_substitutions	4,096	0.338417	17	0.000250
16,384	equal	0	0.006375	0	0.006416
16,384	sparse_substitutions	16	5.640459	16	0.031250
16,384	shifted_periodic	2	6.308292	2	0.036000
16,384	all_substitutions	16,384	4.967416	17	0.000500
65,536	equal	0	0.022167	0	0.022166
65,536	sparse_substitutions	64	97.687792	17	0.034292
65,536	shifted_periodic	2	101.358500	2	0.130500
65,536	all_substitutions	65,536	81.294833	17	0.001333

Unrestricted calls have three sequential same-process timing samples; cutoff has one. No warmup/isolation protocol, cross-machine analysis or customer corpus was used. A cutoff return of 17 means distance > 16, not exact distance 17. A return at or below 16 is the exact distance. These different output requirements explain why comparing their timings is not a like-for-like full-distance speedup.

Every source string is ab repeated to the specified length. Equal targets are identical; sparse targets replace every 1,024th symbol with c; shifted targets are ba repeated; all-substitution targets are entirely c. Sparse distance equals the new-symbol count, which both lower-bounds edits and is achieved by substitutions. All-substitution distance equals length by the same argument. Shifted nonempty pairs admit two edits and cannot be one substitution because more than one position differs. Inputs are stored as exact UTF-8 files without trailing newlines, with hashes in the JSON record.

Scope and controls

The 2,023 passed controls combine all 961 pairs over binary strings of length at most four against an independent dynamic program (distance, four cutoff values, edit count and script application); 1,025 independently iterated integer-log checks; 12 toy guard-boundary checks; seven representative source-length guards; six representation/sentinel checks; and these 12 long cases. Validation record.

Precomposed é versus decomposed e plus combining accent has codepoint distance 2, UTF-8 byte distance 3 and NFC-normalized distance 0. The transformation changes the declared problem. Default case is preserved. No semantic or visual-equivalence score is inferred from literal distance.

The source's bad random executions need not preserve either side of its distance sandwich. Its result is a distance estimate rather than an edit script. The numerical source cutoff suffices for two guards only, not the entire physical-large-input predicate. Full mathematical proof, imported formal closure and source algorithm execution remain unaccepted. The baseline supplies no novel backend advantage and no commercial validation. Revised opportunity.