@mantiz/core

Detection engine for Mantiz — AI lie detector for coding agents.

Pure TypeScript library with zero framework dependencies. Scans git diffs for AI agent cheating patterns.

Installation

npm install @mantiz/core

Usage

import { scanDiff } from '@mantiz/core'

const diff = `
diff --git a/test.js b/test.js
index abc..def 100644
--- a/test.js
+++ b/test.js
@@ -1,5 +1,5 @@
 describe('math', () => {
-  it('adds numbers', () => {
+  it.skip('adds numbers', () => {
     expect(1 + 1).toBe(2)
   })
 })
`

const result = scanDiff(diff)
console.log(`Trust Score: ${result.trustScore}/100`)
console.log(`Findings: ${result.findings.length}`)

API

scanDiff(rawDiff: string): ScanResult

Synchronously run all 11 detectors and return results.

scanDiffAsync(rawDiff: string): Promise<ScanResult>

Async version that also runs AI-assisted detection (via Fireworks/Groq).

ScanResult

{
  trustScore: number        // 0-100
  findings: Finding[]       // Detected cheating patterns
  summary: {                // Aggregate stats
    totalFindings: number
    highCount: number
    mediumCount: number
    lowCount: number
    filesScanned: number
  }
  fixInstructions: FixInstruction[]  // Auto-generated remediation
}

Detectors (11 patterns)

| # | Detector | What It Catches | |---|----------|----------------| | 1 | Disabled Assertion | .skip(), if(false), commented-out assertions | | 2 | Assertion Tampering | Changed expected values without source fix | | 3 | Mock-to-Avoid | Excessive mocking to bypass real errors | | 4 | Claim-Diff Mismatch | Commit msg doesn't match actual changes | | 5 | Silent Catch | Empty catch blocks that swallow errors | | 6 | Hallucinated Assertion | Unknown/non-existent assertion matchers | | 7 | AST Analyzer (Babel) | Parses JS/TS — detects trivial function bodies, async gutting, conditional wrapping | | 8 | Tree-sitter AST | Multi-language AST via WASM — Python, Go, Java, Ruby, Rust, PHP | | 9 | Historical Behavioral | Tracks author patterns — style changes, odd hours, score volatility | | 10 | Mutation Susceptibility | Detects fragile tests with low assertion density | | 11 | AI-Assisted | LLM-powered semantic analysis (Fireworks + Groq) |

File Importance Scoring

Findings are weighted by file type:

| File Type | Multiplier | Examples | |-----------|:----------:|---------| | core / test / source | 1.0x | *.ts, *.test.ts, engine code | | config | 0.5x | package.json, tsconfig.json | | docs | 0.3x | README.md, CHANGELOG.md | | artifact | 0.05x | Agent tool dirs (.kuma/, .claude/), gitignored paths |

Claim-Diff Mismatch Nuance

When scanning via PR URL (with title + author context), Mantiz downgrades findings for:

• Bot authors (renovate[bot], dependabot[bot], etc.) — LOW confidence
• Honest docs-only PRs — LOW confidence

Manual diffs (without context) always get full severity.

ScanResult

{
  trustScore: number        // 0-100, weighted by file importance
  findings: Finding[]       // Detected cheating patterns
  summary: {
    totalFindings: number
    highCount: number
    mediumCount: number
    lowCount: number
    filesScanned: number
  }
  fixInstructions: FixInstruction[]  // Auto-generated remediation
}