Code Coherence AI That Catches ChatGPT-Shaped Answers
Claude-powered analysis evaluates whether edit patterns, variable naming, and logical flow match authentic human coding. Transparent explanations — not a black-box verdict.
Powers the analysis
Not just final code
Every flag has reasoning
First — configurable retention
What authentic human code looks like
Iterative, messy, debuggy — and nothing like ChatGPT's first draft.
Claude-powered analysis
Every submission is analyzed by Claude for edit patterns, logical progression, and style consistency.
Edit pattern detection
Large clean pastes, lack of iteration, unnaturally consistent style — all flagged with evidence.
Logical flow scoring
Did the code evolve through debugging, or arrive fully formed? Authenticity shows in the progression.
Clipboard event correlation
Paste events correlated with suspiciously-large code chunks — a giveaway for LLM copy-paste.
Style consistency checks
Variable naming, comment patterns, and formatting consistency over time — humans drift, LLMs do not.
Transparent reasoning
Every flag comes with Claude's explanation — 'this block was pasted in one event', 'naming inconsistent with prior code', etc.
Integrated with integrity score
Code coherence is one of ten signals feeding the unified authenticity score per submission.
Full session replay
Replay every keystroke, paste, and run to verify coherence AI's findings manually.
Privacy-first
Code is sent to Claude for analysis per signed DPA. Self-hosted Ollama option available for regulated orgs.
Built for how modern teams hire
Humans debug. LLMs drop a finished solution.
Authentic coding shows iteration — a variable renamed, a function extracted, a bug chased. LLM-generated code arrives fully formed. Code coherence AI sees the difference and flags accordingly.
- Edit sequence analysis (type → test → refactor)
- Paste event correlation with code size
- Debugging pattern detection
Every flag has a reason
Black-box cheat detectors create false accusations. Claude's analysis includes a written explanation for every flag — so your team can verify before acting.
- Natural-language explanation per flag
- Evidence links to session replay
- Adjustable thresholds per role
Claude, OpenAI, or self-hosted
Regulated orgs that cannot send code to third parties can self-host the analysis pipeline. ClarityHire supports Anthropic Claude, OpenAI, and Ollama as analysis backends.
- Anthropic Claude (default) for nuanced reasoning
- OpenAI for breadth
- Self-hosted Ollama for sensitive data
How it works
Candidate codes
Monaco Editor captures every keystroke, paste, and edit throughout the session.
Session replay sent to AI
On submission, edit timeline plus final code is sent to Claude for coherence analysis.
AI returns explained flags
Each flag includes evidence and a natural-language explanation of why it was raised.
Your team reviews
Replay the session, read the explanation, confirm or dismiss — human in the loop always.
Frequently asked questions
Can code coherence AI catch ChatGPT-generated code?+
Yes, with high accuracy. LLM-generated code typically arrives as large clean pastes with consistent style and no iterative debugging pattern. Code coherence AI flags those patterns with explanations your team can verify.
How is this different from a plagiarism checker?+
Plagiarism checkers compare against a known corpus. Code coherence AI analyzes the edit pattern — how the code was written — not where it came from. It works even for freshly-generated LLM code that has never been seen before.
What if a candidate genuinely writes clean code quickly?+
Fast clean coding still shows iteration, debugging, and style variation. Code coherence AI looks at the full session trajectory, not a single snapshot. Confidence scores are adjustable per role to tune for senior candidates.
Is candidate code sent to a third party?+
Only if you choose Anthropic or OpenAI — both are vetted data processors with signed DPAs. For regulated data residency needs, use self-hosted Ollama and analysis runs entirely on your infrastructure.
Does the AI give false positives?+
False positives are always possible with any heuristic. ClarityHire combines code coherence AI with other signals (keystroke, paste, face) to reduce false flags and every flag comes with a human-readable explanation so your team can verify before acting.
Explore related features
ClarityHire is one platform. Every feature is built to work with the rest.
Cheat Detection
Ten signal types — face, keystroke, gaze, code coherence, paste events — analyzed in real time.
Keystroke Biometrics
XGBoost classifier detects when a different person takes over the keyboard.
Face Verification
InsightFace + ArcFace continuity checks confirm the same person throughout every session.
Coding Assessments
Live collaborative coding with Monaco, integrated execution, and AI-assisted grading.
Collaborative Code Editor
Monaco + Yjs CRDT lets interviewer and candidate co-edit code in real time.
Catch AI-generated code without false accusations
Turn on code coherence AI in one click. Explained flags, human in the loop, always.