Move Detection

The core mechanism that assigns a persistent identifier to a block of text. When text moves from Section 2.1 to Section 5.3, the algorithm recognizes it as the same logical entity rather than a deletion followed by an unrelated insertion. This is achieved through clause-level hashing and fuzzy matching against a database of previously observed text segments. The system maintains a content-addressable store where each block's hash serves as its immutable fingerprint, enabling cross-location recognition even when surrounding context changes completely.

Distinguishes between a true move and a duplicate insertion. The algorithm employs n-gram similarity and vector embedding diff techniques to confirm that the source location's content has been removed and the target location's content is a match, not a copy. Key indicators include:

• Source depletion: The original location must show a deletion of the matched block
• Temporal ordering: The deletion timestamp must precede or coincide with the insertion
• Cardinality checks: Prevents false positives when boilerplate language appears in multiple locations

Enables the algorithm to track text movement across non-contiguous sections and hierarchical boundaries. A paragraph moved from a definitions section to an operative clause crosses structural contexts. The engine uses tree edit distance algorithms on the document's parsed AST to compute the minimal transformation cost, where a move operation is cheaper than a delete-plus-insert pair. This preserves the change provenance chain, showing reviewers that the obligation in Section 5 originated in the preamble without breaking the audit trail.

Move detection directly feeds into three-way merge and conflict resolution algorithms. When Branch A moves a clause and Branch B modifies the same clause in-place, a naive diff produces an unresolvable conflict. Move-aware differencing recognizes the relocation, applies Branch B's modification to the moved block at its new location, and generates a clean merge. This is critical for operational transformation in collaborative legal editing, where multiple attorneys simultaneously restructure and revise a document.

Traditional redline analysis floods reviewers with false positives when text is reorganized. Move detection collapses what would appear as a delete-and-insert pair into a single, annotated move operation. The redline output displays the block at its new location with a 'Moved from [source]' annotation, dramatically reducing cognitive load. This is especially valuable in obligation change detection, where a moved clause must not be misinterpreted as a new obligation requiring fresh scrutiny.

Move detection extends the classic Myers diff algorithm by adding a third operation to the edit graph: move. While Myers finds the shortest edit script using only insertions and deletions, move-aware algorithms search for longest common subsequences that appear in different positions. The computational complexity increases, but heuristic optimizations—such as locality-sensitive hashing to bucket candidate blocks—keep performance practical for large legal documents. The result is a minimal edit script that accurately reflects human editing intent.

A technique that identifies non-identical but similar strings or paragraphs across documents, crucial for aligning moved or reworded clauses that a strict text comparison would miss.

• How it works: Computes similarity scores using algorithms like Levenshtein distance or Jaccard index
• Threshold tuning: A similarity threshold determines when two blocks are considered the same entity despite textual differences
• Move detection dependency: Without fuzzy matching, a moved clause with minor edits would be treated as a deletion and a new insertion
• Example: A clause relocated from Section 2.1 to Section 5.3 with a single word change is matched at 98% similarity and flagged as a move, not a rewrite

A technique that generates a unique, fixed-size cryptographic fingerprint for an individual clause to efficiently detect any modification to its content across document versions.

• Mechanism: Applies a hash function to the normalized text of each clause
• Move detection role: Identical hashes in different locations immediately signal a pure relocation without any textual change
• Performance advantage: O(n) comparison time versus O(n²) for pairwise similarity computations
• Limitation: Fails to detect moved clauses that have undergone even minor edits; must be paired with fuzzy matching for robustness

A semantic comparison method that converts text chunks into high-dimensional mathematical vectors and measures the cosine distance between them to identify meaning-level changes.

• Representation: Each clause or paragraph is embedded into a dense vector space using a language model
• Move detection capability: A clause moved and substantially reworded while preserving its legal effect will have a high cosine similarity to its original form
• Threshold calibration: Cosine similarity above 0.95 typically indicates semantic equivalence despite surface-form differences
• Example: An indemnification clause rewritten from active to passive voice retains a 0.97 similarity score and is correctly tracked as a moved and reformulated obligation

A quantitative metric measuring the minimum number of single-character operations—insertions, deletions, and substitutions—required to transform one text string into another.

• Levenshtein distance: The most common variant, foundational to diff algorithms
• Normalization: Edit distance is often divided by the length of the longer string to produce a 0–1 similarity score
• Move detection context: Low edit distance between two blocks in different locations suggests a moved block with minor modifications
• Computational cost: O(n·m) dynamic programming complexity makes it expensive for long documents; typically applied only to candidate pairs identified by faster pre-filtering

A text comparison method that decomposes documents into contiguous sequences of 'n' words or characters and measures their overlap to detect paraphrased or reordered content.

• Tokenization: Text is split into overlapping n-grams; common choices are 3-grams or 5-grams
• Jaccard similarity: The ratio of shared n-grams to total unique n-grams provides a fast similarity score
• Move detection application: A clause moved and lightly edited will share a high proportion of its n-grams with the original, enabling rapid candidate identification
• Advantage over edit distance: Much faster to compute and robust to small insertions or deletions that would break exact sequence alignment