Lab
138 explorable, interactive pieces — every concept we write about, made tangible.
Pipeline-parallel decentralised post-training, split across four nodes. Click any internal stage to place the compromised node — the task-vector injection indicator and attack metrics update to show 94% ASR collapsing alignment from 80% to 6%.
An on-chain agent is told to pay an invoice, but one MCP tool hides an instruction in its description that redirects the transfer to an attacker. Pick a defense layer — model, tool, client, policy, custody — and run it; alignment alone refused under 3% of these in MCPTox.
Four guardrail tiers for on-chain AI agents: probabilistic classifiers through Lean 4 theorem provers. Select an attack scenario to see which tier catches it — and which let it through.
Every route to trustworthy AI inference, priced on one log axis: TEE attestation at 1.07×, optimistic re-execution at 2–4×, zkML and FHE at 10³–10⁵×, verifiable FHE beyond. Hover, tap, or arrow through the bars to see what each overhead actually buys — and whom you still trust.
A data market can't pay you for your data's value AND protect it with one privacy budget ε. Drag ε: the worst-case leakage bound hits ≈100% by ε=8 — exactly where from-scratch models finally become usable. Real attacks sit far below the bound, and that gap is your only cover.
Gas cost vs. number of AI inference proofs under three aggregation strategies: naive per-proof (linear), SnarkPack O(log N), and ZK recursion (flat). Drag the slider to find the crossover where each strategy wins.
Seven ZK proof configurations mapped on prover time vs. on-chain verification gas. Click any system to see proof size, setup requirements, and zkML viability. The annotated wrapper band shows the constant-gas region only reachable by STARK→SNARK wrapping.
Seven years of the Elliptic AML benchmark in four switchable charts: the 2019 table everyone forgot, the 2026 leakage-free re-ranking, one GraphSAGE trained on four different graphs, and the F1 cliff at time step 43. Hover, tap, or tab across the bars for exact precision and recall.
ZKLoRA proof times across six base models: the prover bears 31–74 s of computation per LoRA module while the verifier spends 1–2 s. Toggle per-module vs full adapter set; tap a model to see total prover and verifier work — the asymmetry is the deal.
MSM and NTT throughput across four ZK prover hardware tiers — CPU, GPU H100, TPU v6e8, ZK ASIC — normalized to CPU on a log scale. Toggle SNARK (MSM-heavy) or STARK (NTT-heavy) to see how the proof system's operation mix determines end-to-end speedup. Tap a row to inspect.
Log-scale chart comparing physical qubit counts in today's best quantum hardware against the ~500,000-qubit threshold needed to break secp256k1. Tap a chip to see how far it falls short.
Three-column diagram of multi-agent LLM concurrency anomalies — stale-generation, phantom-tool, and causal-cascade. Tap a column to inspect the anomaly; toggle ORI Protection to see how Observable-Read Isolation prevents each failure mode.