EDURIGHTS

Building an Anti-Sybil Protocol for Education Sponsorship — Solo, with AI Agents

I built a protocol that verifies learning before releasing sponsorship funds. Here's the architecture, the anti-fraud stack, and how I did it with zero human developers.

The Problem

Education sponsorship is broken in two ways.

First, Sybil farming. ShapeShift's OP Rewards lost 78% of incentive payouts to automated accounts. Steemit was drained within months. Any system that rewards "learning" without verifying cognitive expenditure gets farmed to zero.

Second, credential inflation. Traditional education verifies attendance, not learning. The Education Gini coefficient correlates with income inequality at r > 0.7. A bachelor's degree now gets you what a high school diploma used to.

251 million children are out of school (UNESCO 2024). 70% of 10-year-olds in low-income countries can't read (World Bank). The projected cost: $21 trillion in lost future income.

The Protocol: PoUC (Proof of Useful Contribution)

The core loop is simple:

Watch → Quiz → Pass → Points → Redeem

Each video is chunked into ~90-second steps. The player sends signed heartbeat attestations every 15 seconds. No skipping, no tab-switching, no speed-up.

After each video, a 4-question MCQ. ≥80% to pass. 3 attempts max, then 1-hour cooldown.

Pass → +15 points + 0.1 GPA

GPA is 0–4.0 scale. Must stay above 3.0 for rewards. Points are non-transferable and redeemable for physical gift cards.

The Anti-Sybil Stack

The interesting part is how we prevent gaming without KYC.

Layer 0 — Economic (already shipping)

const MAX_DAILY_USD = 10;
const MAX_DAILY_ORDERS = 3;
const LEARNING_THRESHOLD = 1; // courses completed

A $10 daily cap makes farming uneconomical at scale. A Sybil operator needs hundreds of accounts, each requiring hours of genuine learning time. The math doesn't work.

Layer 1 — Temporal (already shipping)

Quiz: 3 attempts max → 1-hour cooldown
Redemption: 24-48h admin review window

Time becomes the bottleneck, not code. You can't automate your way through a 24-hour delay.

Layer 2 — Behavioral (coming soon)

Mouse entropy, scroll patterns, micro-timing analysis. Headless browsers produce zero-variance patterns. Canvas/WebGL/audio fingerprint fusion to detect incognito rotation.

Layer 3 — Reputation (coming soon)

Collateralized staking for high-value redemptions. Users stake USDC to unlock higher limits. Sybil linkage detection triggers slashing.

The Tech Stack

Frontend: React 18 + TypeScript + Tailwind CSS
API: Cloudflare Pages Functions
Auth: Supabase (PostgreSQL)
Payments: Whop (third-party)
Email: Resend API
Deploy: Wrangler CLI → Cloudflare Pages

~5,000 lines of TypeScript. No blockchain, no tokens. The economics are intentionally boring — B2B sponsorship, merchant deposits, and a top-up spread.

The Unusual Part: Built by AI Agents

I can't code. I learned 3 months ago and realized I'm faster at orchestrating than typing.

The team:

Agent	Role	Model
CTO	Architecture, code generation	DeepSeek-V4 Flash
Art	UI/UX design, component code	Kimi K3
QA	Testing, bug detection	DeepSeek-V4 Flash
Legal	Compliance review	DeepSeek-V4 Flash
CMO	Market positioning	DeepSeek-V4 Flash
CFO	Economic model	DeepSeek-V4 Flash

I describe what I want. The agents discuss, decide, and execute. I review and ship.

This is the part that gets the most interesting reactions from other developers — some are fascinated, some are horrified. Both reactions are valid.

The Questions I'm Still Stuck On

1. Verification without LLM-gaming. Current MCQ is trivial for GPT-4o. Temporal forgetting curves (24h delay between content and quiz) help but kill UX. What actually works?

2. Identity without exclusion. KYC stops Sybil attacks but excludes refugees, undocumented learners, and children — the exact users this is built for. How do you solve one without sacrificing the other?

3. Cold start. The loop works technically. But to create real value, I need sponsors. To attract sponsors, I need learners. Bootstrap a two-sided market with zero of either?

The Vision

The education system isn't broken — it's working exactly as designed. It filters, credentials, and reproduces class structures.

PoUC is an attempt to build something parallel: a system where what you can actually do matters more than where you came from.

---

Stack: React, TypeScript, Tailwind, Cloudflare Pages, Supabase GitHub: Coming soon (open-sourcing as protocol stabilizes) Live: https://edurights.org

Built With

• anti-fraud
• anti-sybil
• cryptography
• design
• edtech
• game
• mechanism
• social
• theory