What's actually shipped so far, in order — and what's next.
Every claim here traces to real commits and real, verified test runs, not intentions.
Phase 0 — testbed validated
The core idea proven end to end: a task goes out, a volunteer's GPU runs real inference, the output hash meets a difficulty target, a block gets minted. Genesis block minted in 6 seconds.
Security hardening, round 1
Every submission now carries a real Ed25519 wallet signature and the model name used, cross-checked against what's on file for that volunteer.
Phase 0→1 bridge
First block minted via the real AI product's task-routing path, not just a manual test. Canary fidelity-check pool expanded, rate limiting added, wallets encrypted at rest.
Security hardening, round 2
A second, independent review (including an outside AI agent doing an adversarial pass) found and closed a reward-inflation bug and a signature-bypass edge case before either could be exploited for real.
Phase 1 — internal testers
The dashboard, the live public tunnel, and the fidelity-check ("canary") system all validated against two different real hardware tiers.
Phase 1.5 — reward and quality improvements
Volunteers now get credited for total real work done, not just the one winning attempt, and the AI product receives the best available output rather than whichever happened to win the hash race.
5-tester public rollout
Opened up to real outside testers for the first time. Two of five hit real setup bugs — all found, fixed, and verified live same-session. The one-script install, automatic Ollama setup, and self-updating client all came out of this round.
Phase 2 — protocol hardening
Memory-hard proof-of-work closing the cheapest real cheat (faking output without running real inference). Three real economic bugs found and fixed during that work. Fidelity-check prompts made gaming-resistant against a sophisticated volunteer. A real installer with auto-start, clean uninstall, and a diagnostics bundle. Benchmark verification closing a literal loophole where any submission would have passed. All of it shipped with zero disruption to already-connected volunteers.
Reputation tiers & fixing a real "broken volunteer keeps winning" bug
Found live: a volunteer that had gone unresponsive could still keep winning task races ahead of healthy ones. Closed with a proper eligibility check plus a tiered reputation system that rewards consistently-completing volunteers over time (the reputation tiers themselves are live-tracking today, not yet changing task routing, while that gets validated further).
Phase 3 — mobile groundwork
The coordinator now understands mobile/on-device inference as a distinct volunteer tier. The technical spikes for an iOS app (wallet cryptography compatibility, on-device model inference) both passed real end-to-end tests.
This page, and the redesigned front door
You're reading the result — built specifically to explain this project honestly to someone who's never heard of it before.
Bringing Soft Yeti and J-Claw both to a real, complete V1
Both halves of this system need to mature past "actively-tested project" before layering commercial complexity on top of either. This is the current focus, ahead of everything below it.
Opening up to more real volunteers
Deliberately growing past a small, known group of testers — the only way to find out where today's assumptions actually hold up under real, unpredictable use.
Closing the model-substitution gap
Today's fidelity checks catch a volunteer that isn't running real inference at all. They don't yet catch someone quietly swapping in a cheaper model than they claim. This is the next honest gap to close.
Difficulty that adjusts automatically
Right now the mining difficulty is fixed. As the volunteer count and hardware mix grows, this needs to adjust on its own to keep block timing fair and predictable.
On-chain model registry & reputation
Tracking which models are actually in use across the network and rewarding volunteers running well-regarded models.
Proof of Inference Depth
A planned redesign of the core mining mechanism itself — chained self-refinement passes instead of a pure hash lottery, so every bit of compute spent counts toward the reward rather than being thrown away on failed attempts.
Finishing the iOS app
The hard technical questions are answered; what's left is the actual app — wallet, mining engine, background handling, and its own dashboard.
Subscription economy going live
The other half of the loop: a real AI product people and businesses subscribe to using YETI. Deliberately placed after both projects reach a mature V1, not before — launching a subscription product on top of an immature system would be premature. See the "Coming soon" section on the front page.
Longer-term protocol work
Multi-validator consensus, peer-based quality scoring, and cryptographic (zkML) proof that a specific model produced a specific output — the furthest-out, most ambitious items on the list.