Provenance Guard — AI Attribution Detection
A production-minded Flask system for creative platforms that combines LLM analysis and stylometric heuristics to explain whether submitted work is likely AI-generated, likely human-written, or uncertain.

signals
2
label tiers
3
rate limit
10 / hour
audit store
SQLite
The problem
Creative platforms need attribution context without pretending that AI detection is perfect. Provenance Guard gives readers a plain-language result and gives creators a path to contest a classification.
Multi-signal detection
The submission pipeline combines a Groq LLM signal for semantic and AI-like phrasing with local stylometric heuristics for vocabulary diversity, sentence structure, and uniformity. A weighted confidence score keeps the system from reducing a nuanced attribution question to a binary guess.
Production safety layer
Every decision is recorded in SQLite with UUIDs, content hashes, signal scores, confidence, and status. The API also applies a 10-requests-per-hour limit to submissions and supports creator appeals that move a decision to under_review while preserving the audit trail.
The engineering judgment
The original plan considered a flat JSON audit log, but SQLite was chosen for concurrent writes and structured queries. The system also exposes an uncertain label because a false positive can harm a creator more than an inconclusive result.