Human vs. AI – Diff-based line-level provenance for text under agentic editing

The core idea is that git already records who made each change to a file, and that authorship metadata can be mined to classify every line range as human or agent. The tool takes a file path and a flag identifying which commit authors count as human (--ours) or as agents (--theirs), then produces a listing of line ranges with a score between 0.0 and 1.0. A score of 1.0 means a range is fully human-authored, 0.0 means fully agent-authored, and intermediate values like 0.46 indicate a range originally written by a human but subsequently modified by an agent.

The algorithm goes beyond per-line tracking to identify coherent blocks of text, described as "islands" of human authorship inside a "sea" of machine-generated content. It factors in operations like joining, splitting, and dilution of authorship across edits, and is deliberately designed to avoid collapsing everything into either full human or full agent classification. The implementation is in Clojure, with behavior documented through a test file called caution_test.clj.

For a working engineer, the practical value is in mixed-authorship codebases where agentic tools have generated or modified significant portions of code. A developer could run the tool to identify which sections they personally wrote and then communicate boundaries to future agent sessions. The tool's output is a simple range listing, which could be consumed programmatically or reviewed manually. Installation requires bbin, and the tool runs as a CLI inside any git repository.

The limitations are significant and mostly unaddressed in the provided material. The tool depends entirely on git commit metadata being accurate, which breaks down if commits are squashed, rebased, or authored under shared identities. There is no discussion of performance on large repositories or files with extensive histories. The fractional scoring system is explained conceptually but the exact algorithm for computing dilution values is not detailed in the available material.

This event appeared on only one feed, Hacker News, with no headline variation or independent coverage. That means the tool's claims about its diffing approach, its handling of authorship dilution, and its practical utility are uncorroborated. An engineer evaluating this should treat it as an experimental project rather than a validated solution, and would need to test it against their own repositories to assess whether the provenance scores match their expectations.