RealityCheck – AI Cognitive Manipulation Detection Engine

• 

  Critical manipulation detected with explainable score, confidence, tactic heatmap, and evidence integrity panel

• 

  Cognitive Security Lab showing validation metrics, fairness controls, and real-time streaming ingestion workflow.

• 

  History Log with score drift analytics and audit-ready records of past analyses.

• 

  RealityCheck landing page introducing the cognitive cybersecurity mission and core platform strengths.

• 

  Executive dashboard summarizing risk, ablation gain, inter-rater agreement, and fairness drift status.

• 

  Analysis Workbench for source selection, policy presets, and live manipulation scanning of input text.

Inspiration

RealityCheck was inspired by a clear blind spot in content safety: many systems detect toxicity or factual errors, but psychological manipulation tactics often pass through. We wanted to detect how language pressures cognition while staying neutral on ideology.

---

What it does

RealityCheck analyzes digital text (news, social, speeches, marketing) and detects:

• fear amplification
• false urgency
• authority manipulation
• polarization framing
• evidence absence

It returns:

• Manipulation score (0 \rightarrow 100)
• Risk band
• Tactic-level breakdown
• Phrase-level rationale
• Confidence + uncertainty
• Fairness guardrail status
• Audit-ready exports (JSON/PDF)

The scoring intuition is:

[ \text{ManipulationScore} = \sum_i w_i \cdot s_i + \text{concentration_bonus} - \text{confidence_penalty} - \text{language_penalty} ]

where (s_i) are tactic/category scores and (w_i) are calibrated weights.

---

How we built it

• Frontend: React + TypeScript + Vite
  
• Backend: FastAPI + Pydantic + SQLAlchemy
  
• Inference: Hybrid rules + compact transformer ensemble
  
• Validation: Human-eval pack, ablation, threshold tuning, stress/fairness checks
  
• Product surfaces: Dashboard, Analysis Workbench, Lab, History, Transparency, Extension MVP

Core ensemble logic:

[ \hat{y}{tactic} = \alpha \cdot y{\text{rules}} + (1-\alpha)\cdot y_{\text{model}} ]

with per-tactic (\alpha) tuned from validation curves.

---

Challenges we ran into

1. Precision vs recall tradeoff
   Avoiding false positives on expressive but legitimate writing.
2. Explainability depth
   Making outputs understandable to judges and non-ML stakeholders.
3. Neutrality + fairness constraints
   Ensuring no political stance classification and adding identity-swap drift checks.
4. Hackathon runtime constraints
   Keeping the system laptop-runnable while preserving technical depth.

---

Accomplishments that we're proud of

• Built an end-to-end cognitive cybersecurity platform, not a toy classifier.
• Added phrase-level rationale, confidence calibration, and governance metadata.
• Implemented fairness guardrails and stress-test workflows.
• Delivered stream triage with review/escalation actions.
• Produced submission-grade evidence artifacts and reproducible evaluation scripts.

---

What we learned

• In high-stakes AI, governance and explainability are first-class requirements.
• Manipulation detection needs multi-signal structure, not keyword matching.
• Clear boundary definitions increase trust:
  • what the model does
  • what it does not do
• Operational workflows (triage, audit export, transparency) are critical for adoption.

---

What's next for RealityCheck – AI Cognitive Manipulation Detection Engine

• Expand from 5 to 10 tactic classes.
• Improve multilingual robustness beyond English-first scope.
• Deploy compact ONNX model as default runtime path.
• Add enterprise controls (team workflows, RBAC, policy presets).
• Strengthen active-learning loops using drift + error-analysis feedback.
• Launch production browser extension backed by the full API stack.

Built With

• docker
• fastapi
• github
• html/css
• huggingface
• javascript
• onnx
• pydantic
• pytest
• python
• react
• recharts
• sqlalchemy
• sqlite
• typescript
• uvicorn
• vite