Skip to content

CHANGELOG.md

Latest commit

 

History

History
112 lines (81 loc) · 5.38 KB

CHANGELOG.md

File metadata and controls

112 lines (81 loc) · 5.38 KB

Changelog

All notable changes to Likelihoodlum are documented here.

[1.1.0] — 2026-02-26

Three New Heuristics 🔬

Added comment density, commit time-of-day, and diff entropy analysis — bringing the total to twelve scoring signals.

Added

  • Commit Time-of-Day Analysis (0 to +5 pts) — Flags repos where >30% of commits happen between midnight–6am with suspicious velocity. Humans have circadian rhythms; LLMs don't sleep.
  • Comment Density Analysis (−3 to +5 pts) — Measures the ratio of comment lines to code lines in added content. LLMs over-explain (≥35% comments → +5); humans are lazier (<5% → −3).
  • Diff Entropy Analysis (−3 to +5 pts) — Computes Shannon entropy of diff content per commit. LLM diffs are more repetitive/formulaic (low entropy → +5); human diffs are varied and chaotic (high entropy → −3).
  • New report sections: 🕐 Commit time distribution, 💬 Comment density, 🔀 Diff entropy
  • Patch content extraction from commit details for comment and entropy analysis
  • comment_density, diff_entropy, and commit_times fields in JSON output
  • Comment line detection for Python, C/C++, Java, JS, Go, Rust, Swift, Dart, SQL, Lua, HTML

Changed

  • Total scoring signals increased from 9 to 12
  • Signal numbering updated (project-scale plausibility is now #11, generated file ratio is now #12)
  • Removed "diff complexity scoring" and "comment density analysis" from Contributing ideas (now implemented)

[1.0.0] — 2026-02-26

The Public Release 🎲

First public release with nine heuristic signals and a comprehensive scoring engine.

Scoring Signals

  • Code Velocity (−10 to +35 pts) — Lines/min between consecutive commits by the same author, with trimmed mean boost when heavy tail detected
  • Session Productivity (−5 to +20 pts) — Aggregate lines/min per coding session (>2hr gap = new session)
  • Commit Size Uniformity (−5 to +15 pts) — Coefficient of variation in commit sizes; LLMs are suspiciously uniform
  • Commit Message Patterns (0 to +15 pts) — Regex matching for LLM-typical phrasings, conventional commit verbosity, multi-scopes
  • Burst Detection (0 to +15 pts) — Rapid bursts (>300 lines in <30 min) and sustained high-velocity sessions
  • Multi-Author Discount (−10 to +5 pts) — More authors = less likely LLM; solo author = small bump
  • Extreme Per-Commit Velocity (0 to +10 pts) — Intervals exceeding 50 lines/min (~3,000 lines/hr)
  • Project-Scale Plausibility (−5 to +20 pts) — Total authored output vs repo creation date and active days
  • Generated File Ratio (informational) — Percentage of changes in vendor/generated files

Features

  • Zero dependencies — stdlib-only Python 3.10+
  • Concurrent commit detail fetching (up to 10 parallel requests)
  • Bot author filtering (dependabot, renovate, etc.) — excluded from analysis and API calls
  • Repo metadata fetch for true creation date
  • Generated/vendored file filtering (lockfiles, protobufs, Xcode, build artifacts, assets)
  • Bidirectional scoring — human patterns actively reduce the score
  • Full JSON output mode (--json)
  • .env file support with built-in fallback parser

Documentation

  • Wall of Truth leaderboard with 20+ tested repos
  • Anthropic Spotlight section
  • Contributing guide with heuristic development protocol
  • Issue templates for bug reports and feature requests
  • pip-installable via pyproject.toml

[0.3.0] — 2026-02-26

Added

  • Project-scale plausibility heuristic comparing total output against repo creation date
  • New 📈 Project-scale output section in report showing repo creation date, active days, and daily output
  • Daily output thresholds (300 / 800 / 2,000 / 5,000 lines per active day)
  • project_scale object in JSON output

[0.2.0] — 2026-02-26

Added

  • Conventional commit pattern detection (multi-scopes, verbose descriptions, feat(a, b):)
  • Bot author filtering — dependabot[bot] and similar excluded from velocity, sessions, and author counts
  • Trimmed mean velocity boost when trimmed mean >> median (heavy tail signal)
  • suspicious_pct fallback for interval percentage scoring
  • High-velocity session detection alongside duration-based bursts
  • Per-commit extreme velocity signal (>50 lines/min)
  • Clean-messages-but-high-velocity cross-signal
  • Expanded LLM message patterns (enhance X with Y, integrate/wire up, add X Y Z)

Changed

  • Burst detection threshold lowered from 500 to 300 authored lines, minimum commits from 3 to 2
  • Burst scoring bumped (3+ sessions: 8→10 pts)
  • Velocity scoring now considers trimmed mean for tier boosting (22→26 or 30 pts)

[0.1.0] — 2026-02-26

Added

  • Generated/vendored file filtering with authored_total vs generated_total breakdown
  • Negative scoring signals (human velocity, human session pace, high commit size variation)
  • Multi-author discount (−10 for 5+ authors, −5 for 3+)
  • Solo author bonus (+5)
  • Trimmed mean for velocity and session calculations
  • 📁 Line changes breakdown section in report
  • Seven heuristic signals in initial scoring engine

[0.0.1] — 2026-02-26

Added

  • Initial commit: basic LLM likelihood detector
  • Five heuristic signals: velocity, sessions, bursts, commit size uniformity, message patterns
  • GitHub API integration with token support
  • CLI with --token, --branch, --max-commits, --json flags
  • .env file support with python-dotenv fallback