compression-distance detectors · surface similarity

Find near-duplicates and novelty, by compression distance.

at1 ncd uses normalized compression distance to measure how similar two files are without parsing them, a code-clone / near-duplicate scan and a clause/document novelty triage. It sees surface structure, not meaning.

Two detectors, one distance

Code-clone & near-duplicate scan

Scan a tree and surface files (or functions) that are near-copies of each other by compression distance, copy-paste clones, forked-and-diverged files, redundant assets, without a language parser or an AST.

Clause & document novelty triage

Given a corpus of prior documents, score how novel a new one is: which clauses or paragraphs closely echo something already on file, and which are genuinely new, a fast first pass before a human reads closely.

Parser-free, format-agnostic

NCD works on raw bytes, so the same detector runs on source, contracts, logs, or binaries with no per-format setup. It's a triage instrument that points a reviewer at the pairs worth looking at.

One command surface

# normalized compression distance: 0 = near-identical, 1 = unrelated
at1 ncd  distance  a.py  b.py            # -> 0.07   (near-duplicate)
at1 ncd  clones    ./src  --threshold 0.2  # find code-clone / near-dup clusters
at1 ncd  novelty   new.txt  --corpus ./prior/  # which clauses echo prior docs, which are new

Honest scope

NCD is a surface-similarityinstrument: it finds byte-level closeness, not semantic equivalence. Two files that say the same thing in different words read as far apart, and two unrelated files that share boilerplate read as close, so it’s a triage pass that points a human at the pairs worth reviewing, not a plagiarism or equivalence verdict. It is calibrated on synthetic and exemplar data, with broad real-corpus validation as the explicit graduation gate. For “what kind of file is this?” see at1 origin; for “how rule-governed is it?” see at1 determinism.