Skip to content

NeftalemMG/Trace

Repository files navigation

Trace

A real-time platform that audits, scores, and fixes how AI systems represent Canadian organizations, programs, and opportunities. Every tool makes live API calls.


Setup

git clone https://github.com/NeftalemMG/Trace.git
cd TRACE
npm install
cp .env.local.example .env.local
# add your API keys to .env.local
npm run dev

Open http://localhost:3000


API Keys

Key Where to get it Cost
ANTHROPIC_API_KEY console.anthropic.com/settings/keys Pay per use
PERPLEXITY_API_KEY perplexity.ai/settings/api Free credits available
GEMINI_API_KEY aistudio.google.com/app/apikey Completely free
OPENAI_API_KEY platform.openai.com/api-keys Pay per use

Only ANTHROPIC_API_KEY is required. The other three unlock their respective platforms in the Multi-Platform tool.


Pages

Home: Landing page with live telemetry feed, platform accuracy comparison, and equity visibility bars.

Query Engine (/query-engine): Enter any student-facing query about Canadian opportunities. Runs a dual-layer analysis: a naive pass showing what AI tells the student, then a verification pass that identifies hallucination patterns, ghost entities, and equity gaps. Six research presets included.

Audit Tool (/audit): Enter an organization's details to receive an AiEO Visibility Score across six dimensions: Accuracy, Completeness, Hallucination Resistance, Equity Reach, Data Freshness, and Regional Specificity. Returns benchmark comparisons, key issues, likely hallucination patterns, and queries the org is probably invisible to.

Multi-Platform (/multi-platform): Run the same query simultaneously across Claude, Perplexity, Gemini, and ChatGPT GPT-4o. Returns side-by-side responses with per-platform accuracy estimates, hallucination risk scores, cross-platform entity verification, and consensus/contested claim analysis.

Hallucination Registry (/registry): Browse documented hallucinations from the research study. Submit new ones by pasting any AI response — Claude classifies the error type, risk level, community impact, and harm scenario automatically. Pre-loaded with all five documented hallucinations from the paper.

Equity Map (/equity-map): Select a community (Indigenous-led tech, Black founders, Halifax ocean/bio, etc.) and get a real analysis of which organizations serve them, how accurately AI represents each one, and what structural gaps exist. Visualizes the 54-point gap between Indigenous-led tech (28/100) and major urban ecosystems (82/100).

AiEO Guide (/guide): Enter an organization's details and receive a full implementation guide: ready-to-copy JSON-LD schema markup, a directory submission list with URLs, content strategy tactics with rationale, equity-specific visibility recommendations, FAQ schema questions, and a phased implementation timeline.


API Routes

All routes are in src/pages/api/.

  • query-engine.ts - Dual-layer Claude analysis (naive response + critical verifier)
  • audit.ts - 6-dimension AiEO score generation
  • multi-platform.ts - Parallel calls to Claude, Perplexity, Gemini, and OpenAI GPT-4o with cross-platform analysis
  • registry.ts - Hallucination classification (error type, risk level, harm scenario)
  • equity-map.ts - Community ecosystem analysis
  • guide.ts - Full AiEO implementation guide generation

Research Background

We tested GPT-4o, Gemini, Perplexity, and Claude across five student queries targeting niche Canadian ecosystems. Every response was cross-referenced against official sources. Key findings:

  • Gemini: 40% accuracy, 95% confidence - highest hallucination risk
  • ChatGPT: 55% accuracy, 90% confidence - fabricated entities presented with full confidence
  • Perplexity: 65% accuracy, 75% confidence - best accuracy via live search
  • Claude: 70% accuracy, 45% confidence - only platform to flag uncertainty

About

Topics

Resources

Stars

Watchers

Forks

Releases

No releases published

Packages

 
 
 

Contributors