Engineers mint the fixes they already figured out. Agents pay $0.50 to one-shot the same wall. Everyone gets paid in USDC, on-chain.
Join the beta Read the design docTwo weeks ago I burned three hours and ~$25 of RunPod time trying to serve Qwen 3.5 on an AMD MI300X. Docs said supported. Claude said supported. Both lied — vLLM's pip wheels are CUDA-only, and the ROCm image wasn't on the pod template. I found that in a stale GitHub issue and a Discord thread.
No frontier model was going to fix this. Every frontier model's weights froze before that hardware-stack combination stabilized. The knowledge existed — in someone else's head, buried in someone else's chat log. A $0.50 block from the engineer who'd already burned those three hours would have saved me everything. That's the $100B problem frozen-weight models structurally can't solve.
Install kp. Drop in a Claude Code, Cursor, or ChatGPT export. The six-stage pipeline — ingest, PII-scrub, semantic chunk, structured extraction, quality score, creator review — runs on your machine. Only the clean block hash touches our infra.
Every block ships a deterministic test bundle. Staked validators re-run it in a sandboxed container and sign attestations. Blocks earn a tier badge (T1 gold → T4 unvalidated). Buyers pay for trust, not hidden content.
Mint the block as an IP asset on Story Protocol. Agents discover it via MCP, pay over HTTP 402, inject it, complete the task. You earn 70% of every sale. Royalties flow automatically across derivative blocks.
50 technical debugging tasks across GPU serving, ML training, framework deadlocks, infra/deploy, and generic debugging. Four arms, 200 completions, LLM-as-judge grading on discrete MET/NOT MET criteria.
| Arm | Accuracy |
|---|---|
| Gemma 3 12B, no block | 43.0% |
| Gemma 3 12B + knowledge block | 90.5% |
| Claude Sonnet 4.6, no block | 81.5% |
| Claude Sonnet 4.6 + knowledge block | 94.0% |
A $0 open-weight model plus a $0.50 block outperforms a frontier model operating blind by 9 percentage points. Blocks don't just close the gap — they invert the hierarchy.
Error-loop detected → MCP auto-invokes search_blocks → three candidates → HTTP 402 → agent signs from Privy wallet → block decrypts → task completes. Sub-$2 auto-approves; above $2 prompts a human. Creator royalty lands in the same tx.
You already burned the three hours. You already wrote the Slack post nobody saw. Export the session, run the local SDK, mint the fix. Every time another agent hits the same wall, you earn 70% of the sale in USDC — directly, on-chain, no aggregator.
{"mcpServers": {"knowledge-protocol": {"command": "npx", "args": ["@kaeva/knowledge-protocol-mcp"]}}}
Four tools: search_blocks, preview_block, purchase_block, inject_block. Auto-trigger on error-loop. Prefund a Privy wallet with $10 USDC — first 20 stuck-points covered. Works with any model taking system-prompt injection (Opus, Sonnet, GPT-4o, Gemini, local Llama). HTTP 402 native — agents sign, not users. Every T2+ block ships a runnable test_bundle/.
Yes. SDK runs locally — only the scrubbed block hash touches our infra. Encrypted payloads on IPFS via Pinata. On-chain hash is a severable pointer (GDPR-compliant).
$0.50 (semantic) to $50+ (T1 gold). Agents pay from prefunded USDC wallets via HTTP 402. Split: 70% creator / 25% validator / 5% protocol.
Tier badges. T4 is unvalidated. T2 means a validator ran the test_bundle/ in sandbox and attested the output matches. T1 adds TEE attestation.
Yes, with care. Validator comp is per-task piece-rate (service model), not pool share. Never framed as appreciating assets. Delaware C-corp, crypto-literate counsel.
Creators pay own extraction compute (~$0.01–0.05); downside capped. Tiered dispute oracle: T1 needs specific counterexample + declared-env match. Sub-$1 griefing rate-limited. V1 manual; V2 automated sandbox re-run.