Every day, a massive volume of complex AI research (papers and codebases) is published. However, simply reading an abstract, skimming a GitHub repo, or watching a YouTube explainer isn't enough to truly understand the mechanics of new models. The core problem is that the gap between a paper being "published" and actually being "understood" by the community is widening.

What "Papers with Claude Code" Solves

• Course Creation Friction: It eliminates the weeks of manual work usually required to build a technical curriculum. By analyzing an arXiv URL and a GitHub repo, Claude Code generates a grounded, hallucination-free course in under a minute.

• Broken Platform Economics: It replaces expensive monthly subscriptions and greedy platform middlemen with a native x402 micropayment system. Learners pay fractions of a cent in USDC (on Base/Kite Chain) only for the exact concepts they unlock. By using ERC-8021 builder codes, it automatically routes on-chain royalties directly to the original paper authors whenever their work is taught.

• Isolated Learning: It moves education out of private silos. Progress is recorded on the AIN blockchain, creating a verifiable learning passport for users and mapping a real-time, global Knowledge Graph of what the community actually understands.

0. Deterministic Course Generation from Papers

6-step headless pipeline: parse paper → extract 15-30 concept nodes → topological sort by prerequisites → generate lessons with quizzes → scaffold files → publish to AIN with x402 pricing. Deterministic slugs ensure the same paper always produces the same course ID regardless of input URL format (arXiv, GitHub, HuggingFace all normalize to canonical title). Paper text treated as DATA only — defense against prompt injection in adversarial papers.

1. P256 Passkey as Blockchain Wallet

WebAuthn mandates P256 (ES256), but P256 lacks reliable public key recovery from signatures — unlike secp256k1's ecrecover. We designed a new 98-byte signature format (0x02 + compressedPubKey(33) + r(32) + s(32)) that embeds the public key directly, making verification deterministic on AIN blockchain while remaining backward-compatible with existing 66-byte secp256k1 signatures. The passkey IS the wallet — no seed phrases, no private key management.

2. Dual-Chain Architecture

The knowledge graph needs fast, append-only writes. Payments need EVM compatibility and stablecoins. No single chain does both. We split: AIN blockchain stores the knowledge graph (explorations, topics, frontier maps), Kite/Base chain handles USDC micropayments (ERC-4337 AA, ERC-8004 agent identity, ERC-8021 attribution). A Cogito Node writes an exploration to AIN, tags it with a price, and a learner on Base pays to access it — x402 bridges the two chains.

3. x402 Micropayment Protocol

At ~$0.001 per stage, traditional payment rails (Stripe minimum fees, subscriptions) don't work. We adopted HTTP 402: server returns payment details, client signs and retries with X-Payment header. An x402-aware fetch wrapper handles the 402 → sign → retry flow transparently — application code never sees payment logic. Adapter pattern switches between mock (dev) and real (Kite chain) via env var.

4. Ephemeral K8s Pods for Per-Learner AI Tutors

Each learner gets a dedicated Claude Code instance in a dynamically created Kubernetes pod. Perfect isolation — no shared state, no context leakage. Dynamic creation over pre-allocated pool: security outweighs the 5-10s startup latency. WebSocket over SSH: no port exposure, works through standard proxies. Three cleanup paths (unmount, HMR, tab close) required to prevent zombie session leaks.

5. Knowledge Graph as Collective Intelligence Ledger

Multiple autonomous Cogito Nodes independently explore papers and write to the same AIN blockchain graph — no central coordinator. Each exploration creates a graph node with typed edges (extends, related, prerequisite). Frontier maps aggregate depth across all nodes, enabling cross-node discovery: Node A sees "transformers explored to depth 5" by Node B and fills gaps at depth 6. Throughput limited by consensus — solved with polling intervals and batch writes.

6. Self-Sustaining Cogito Node Economics

Local LLM (Qwen3-32B-AWQ on A6000) costs ~$0.004/inference vs. cloud API at $0.01-0.05. At $0.001/stage micropayment prices, cloud APIs would never break even. Local LLM becomes self-sustaining at ~180 queries/day. The SUSTAIN phase tracks P&L and shifts exploration toward higher-demand topics visible on frontier maps.

7. ERC-8004 + ERC-8021: Agent Identity and Author Attribution

Each Cogito Node registers as an ERC-8004 autonomous agent on Base — not a plain NFT, but full identity: setAgentURI(), metadata registry, wallet binding. Every x402 payment transaction includes ERC-8021 builder codes encoding arxiv_id + github_repo + first_author — immutable on-chain attribution of original paper authors whenever their work is monetized.