website (still working on it) https://capture.fun/
Give feedback to the team
Your message is private to you and the Capture team.
Create an account to send feedbackAfter signup, you can write a private message with this update attached.
website (still working on it) https://capture.fun/
Your message is private to you and the Capture team.
Create an account to send feedbackAfter signup, you can write a private message with this update attached.
About this project
AI / ML Models
Capture is a decentralized Reinforcement Learning Human Feedback (RLHF) platform. Contributors get paid for evaluating AI decisions - Step 1 stake a USDC bond to prove human, Step 2 judge a vision model's outputs (Choose A or B), and Step 3 get paid in USDC when their answer forms a consensus with other contributors. Make too many bad judgments and it may cost losing the bond. The aggregated preferences form an RLHF dataset with on-chain provenance, all settled on Solana.
View projectLiveness — made it actually verify motion (no Google Cloud) - App now sends real motion evidence (face-box track + spoof live-score) instead of a hash; server verifies a face genuinely moved. Bare digests like "11"×32 now fail. - Removed the Google Cloud / Play Integrity requi…
- Upgrading the slashing conditions to dynamic: floor = 0.45 + max(0, 0.70 − network_agreement), capped at 0.90 — so the floor rises (gets stricter) when the whole network's agreement sags - Ran an independent adversarial review of that logic, fixed the findings, and added re…
Hardening + store prep this week: - Release-signed the Android build (real signing cert, not debug) - Turned on R8 minify: 44.6 MB → 27.8 MB - Added real liveness gating, demo bypass off for paid work - Split signing: the API holds no private keys, only the background worker s…