# Open-source detection components

This is an independent application, not an official Turnitin service or report.

## Local similarity

Noplag Engine, NoplagLabs, Apache License 2.0.
Reference: https://github.com/NoplagLabs/noplag-engine
Pinned source: aaf7839766460cf387e91b5a730aba8928db47d8
License: /licenses/noplag-apache-2.0.txt

The winnowing selection logic in lib/local-similarity.mjs is adapted to JavaScript. This application changes hashing to FNV32, normalizes Unicode grapheme clusters with original offsets, keeps the corpus in browser memory, and uses its own exact suffix-automaton alignment. It does not run the Noplag server or distribute its sample corpus. It does not inherit the upstream engine's benchmark results.

## Experimental AI classifier

Original training project: Copyright (c) 2024 Tuong Vy Vu, MIT.
https://github.com/vutuongvy101/multilingual-ai-human-text-detection
Checkpoint: bibbbu/multilingual-ai-human-detector_xlm-roberta-base
Upstream revision: b4aa2e06d8428b4d27f2ef625267044d74b2eeb8
License: /licenses/ai-upstream-mit.txt

Browser conversion: Copyright (c) 2026 MyToolster contributors, MIT.
https://github.com/mujian2026/multilingual-ai-text-detector
Model: https://huggingface.co/mujian2026/multilingual-ai-text-detector
Model revision: 60be618d23e169de6fe04b8e6135be7d8976abb7
Variant: q4, XLM-RoBERTa, human=0 / ai=1.
License: /licenses/ai-conversion-mit.txt

The application downloads the pinned public model and tokenizer directly to the user's browser. Text is processed locally. Segment scores are model outputs, not calibrated authorship probabilities. The token-weighted average is an application-defined aggregation, not a validated whole-document metric. Academic domain performance is unvalidated. The upstream dataset contains English, Simplified Chinese and Vietnamese QA responses, with AI responses from Qwen2.5-1.5B. Its 270-example test set is not evidence of general accuracy.

## Runtime

Transformers.js 3.8.1, Hugging Face, Apache-2.0.
https://github.com/huggingface/transformers.js
License: /licenses/transformers-apache-2.0.txt

ONNX Runtime Web, Microsoft, MIT.
https://github.com/microsoft/onnxruntime
License: /licenses/onnx-runtime-mit.txt

## PDF export

pdf-lib 1.17.1 and @pdf-lib/fontkit 1.1.1, MIT.
https://github.com/Hopding/pdf-lib
https://github.com/Hopding/fontkit

Noto Sans SC Regular, SIL Open Font License 1.1.
Static weight 400 instance generated from the Google Fonts variable TTF with fonttools.
https://github.com/google/fonts/tree/main/ofl/notosanssc
License: /licenses/noto-sans-cjk-ofl.txt

## Sentence similarity and comparison trial

Free-Turnitin-Plagiarism-Checker, cu-sanjay and contributors, Apache-2.0.
https://github.com/cu-sanjay/Free-Turnitin-Plagiarism-Checker
Pinned revision: 8302a5013ae9ee750965163d40fd53d337939c30
License: /licenses/free-turnitin-apache-2.0.txt

The comparison page includes the original calculateSimilarity and nGramSimilarity functions without changes. Network functions are omitted from the browser scoring module. Recorded web results were produced by the original server route in a controlled trial. The browser comparison uses supplied source text and does not run the upstream web search service.

The source-review pipeline also applies mathematically equivalent indexed cosine and five-word overlap scoring to each eligible sentence and each safely retrieved source. It preserves the upstream sentence boundary rule, minimum sentence length, source prefix of 5,000 characters, and rounded maximum-score aggregation. It processes all eligible sentences instead of only the first 20, uses the existing verified-source retrieval service, and records original text offsets and per-source scores. This additional Similarity Score is calculated before exclusions; it does not measure the proportion of copied text and is separate from exact text overlap and AI Writing.

## Cloud source PDF extraction

unpdf 1.8.1, Johann Schopplich, MIT.
https://github.com/unjs/unpdf
License: /licenses/unpdf-mit.txt
Uses the serverless PDF.js build for bounded text extraction from retrieved public PDF sources.