James Okafor

General 17 min

James Okafor • 1 week ago

How Code Similarity Detection Moved From Text to Trees

Five generations of code similarity detection have each been beaten by a student with a text editor. This is the technical history of how detectors moved from byte comparison to tokens, fingerprints, syntax trees and learned vectors, and what that arc tells you about choosing a tool that can still find a match when the code has been renamed, refactored, or generated by a model.

Read

General 10 min

James Okafor • 2 weeks ago

Token Fingerprinting vs AST Matching on 900 Refactored Submissions

Take 30 hand-written Java solutions, push them through five levels of mechanical refactoring, and you get a controlled ladder for asking a simple question: how much does each similarity engine actually see? The answers differ by nearly a factor of four, and the differences map cleanly onto how we should be designing assignments.

Read

General 11 min

James Okafor • 3 weeks ago

How Similarity Detection Learned to See Through Refactoring

Renaming variables and restructuring control flow used to be the cheat code for copying a classmate's submission. It hasn't been for years, and the techniques that closed that door are worth understanding before LLM rewrites move the goalposts again. A practical look at what each family of code similarity detection actually compares, and where all of them still fail.

Read

General 10 min

James Okafor • 1 month ago

Does Convergent AI Output Look Like Peer Plagiarism to a Detector?

Twenty-three of 412 submissions in a data structures course shared a Dijkstra implementation that differed by fewer than three tokens, and the peer similarity engine flagged all 23 as a single cluster. Nobody had copied anybody. Here is the mechanism behind convergent AI output, the identifiers and AST evidence that separated it from real collusion, and the three-pass workflow we used to keep the scores from contaminating each other.

Read

General 11 min

James Okafor • 1 month ago

What Cross-Language Code Plagiarism Detection Can and Cannot See

Cross-language code plagiarism detection compares normalized structure rather than raw text, which works when a translation was mechanical and fails when the student rewrote the algorithm. Here is what survives a Java-to-Python translation, what the token and IR approaches actually see, and how to run the check across a whole cohort without drowning in false positives.

Read

General 10 min

James Okafor • 1 month ago

Scanning 9,301 Python Files for Stack Overflow Copy-Paste

A practical, code-level guide to batch-scanning Python files for web-sourced code. We walk through token normalization, fingerprinting, uploading to Codequiry, interpreting web match URLs, and stacking an AI check on flagged files. Built for CS professors auditing assignments and engineering managers verifying contractor code.

Read

General 9 min

James Okafor • 2 months ago

A TA's Method for Refactoring-Resistant Plagiarism Checks

Line diffs collapse when students rename variables, reorder functions, and extract methods. Here's how token fingerprints and AST normalization catch refactored copies, and how to triage hundreds of similarity reports without drowning in false positives.

Read

General 7 min

James Okafor • 2 months ago

Teaching Web Code Plagiarism Detection With Real Student Cases

Web code plagiarism hides in plain sight when students copy from Stack Overflow, GitHub, or tutorials and rename a few variables. This post shows how to teach detection as a skill, design assignments that surface copied web code, and use a source-aware checker like Codequiry to see the evidence.

Read

General 8 min

James Okafor • 2 months ago

ChatGPT vs Copilot vs Gemini Code Detection Benchmarked

A head-to-head evaluation of AI code detection across ChatGPT-4o, GitHub Copilot, Claude 3.5 Sonnet, and Gemini 1.5 Pro. One pattern kept surfacing: text-only detectors miss refactored LLM code, while structural and multi-signal checks hold up.

Read

General 11 min

James Okafor • 2 months ago

Token Fingerprinting vs AST Matching on 1,000 Refactored Java Programs

When students rename variables, extract methods, and reorder statements to hide copied code, which detection algorithm actually holds up? A controlled experiment pits winnowing, token-based matching, and AST structural hashing against a ladder of refactoring transformations — and reveals why single-technique checkers miss the cases that academic-integrity panels care about most.

Read

General 9 min

James Okafor • 2 months ago

Finding Stack Overflow Code in Student Submissions With Fingerprints

A study of 5,000 Java assignments from three US universities found that nearly one in four contained code blocks directly traceable to Stack Overflow answers — yet traditional similarity checkers missed them all. We applied token-sequence fingerprinting and a web index of 1.2 million programming snippets to surface hidden web plagiarism at scale.

Read

General 11 min

James Okafor • 3 months ago

How Perplexity-Based AI Code Detectors Actually Work

Perplexity-based detectors aren’t magic — they measure how surprising a sequence of code tokens would be to a model trained on real human code. This report breaks open the inner math, real false-positive rates from Stanford and Edinburgh benchmarks, and why the strongest detectors stack statistical signals with AST fingerprinting and web-source checks.

Read