Writing from the Daedalus workshop
Deep-dives on protocols we're shipping, frameworks we're building, and research we're publishing — organized by product line.
FCMP++ Explained: Full-Chain Membership Proofs, Designed for Innova
FCMP++ (Full-Chain Membership Proofs with Curves) proves a spent output is one member of the entire UTXO set without revealing which — replacing classical ring signatures with whole-chain anonymity. Innova's construction uses dual-curve Merkle trees (secp256k1 + Ed25519), arity 256, max depth 8. Activation is scheduled at block 7,320,000; not yet live on mainnet.
Innova's Five-Layer Privacy Stack, Explained Layer by Layer
Innova stacks five independent privacy primitives. Layer 1 hides amounts (Pedersen + Bulletproofs). Layer 2 hides the sender (Lelantus, with FCMP++ as the evolution). Layer 3 hides the receiver (stealth and silent-payment addresses). Layer 4 hides the network origin (Dandelion++). Layer 5 hides staked amounts (NullStake).
NullStake: Zero-Knowledge Private Staking Designed for Innova
NullStake is Innova's designed zero-knowledge private staking protocol, with two generations scheduled for activation. V1 (Sigma protocol, ~1.8 KB proofs) activates at block 9,500,000 via IIP-0009. V2 (Poseidon2 + BPAC, ~720 bytes) activates at block 10,000,000 via IIP-0010. Until then, Innova uses transparent staking.
DEBATE · PAIR · PLAN→EXECUTE · VERIFY: Dual-Model AI Coding
JaCode is designed around four dual-model conversation modes. DEBATE pits two models against each other to surface blind spots. PAIR trades driver/navigator roles like pair programming. PLAN→EXECUTE splits strategy from tactics across two models. VERIFY has a second model audit the first's work before commit. JaCode is pre-release — this post describes the design, not shipping capabilities.
Why CPU Side-Channel Attacks Need a Hardware Fix
CPU timing side-channel attacks let co-resident processes extract cryptographic keys by watching micro-architectural timing. A decade of microcode patches has shown this is not a problem CPU vendors will fully fix. Lavalamp's approach is to mitigate at the hardware boundary with an FPGA — OS-transparent, vendor-independent, no application changes.
Logos vs OpenAI vs Together vs Replicate: Fine-Tuning Buyer's Guide
Pick OpenAI if you want a hosted GPT endpoint and don't need the weights. Pick Together or Fireworks if you want hosted inference on open models. Pick Replicate if you're already running everything through Replicate's API. Pick Logos if you want to own the weights and run inference yourself.
45 Open-Weight Models Ranked for Fine-Tuning in 2026
Pick Llama 3.3 70B or Qwen 2.5 72B for the safest 70B-class production bet. Pick Phi-4 14B or Mistral Small 24B for efficient reasoning at small size. Pick Gemma 4 26B MoE or 31B Dense for multimodal + long context. Pick Hermes 3 for function-calling pre-training. Reach for GLM-5.1, Kimi K2.5, or MiniMax M1 only when frontier agentic capability is the job.
What is Logos? The Fine-Tuning Service Where You Keep the Weights
Logos is DDG's fine-tuning-as-a-service. You choose an open-weight model, we train it on your data, you get the raw weights as .safetensors plus a LoRA adapter — no hosted inference, no API, no vendor lock-in. Four tiers from $349 to $3,499, plus a Custom quote.
Annuities On-Chain: How Lucrum's Annuitas Module Works
Annuitas is Lucrum's on-chain annuity protocol. A singleton contract stores each annuity as a uint256 ID. Three V1 tiers: Variable (principal preserved, pays harvested yield), Fixed-Term (target payment via yield→buffer→principal waterfall), Guaranteed (Fixed-Term with a Guard-backed floor in V2). Payments route anywhere; a permissionless dead-man's-switch handles beneficiary fallback.
Lucrum Architecture: FOLLIS, FRS, Vestibulum, Tribunus Explained
Lucrum is one integrated DeFi protocol — open-source contracts, DAO, and financial tools in a single monorepo. FOLLIS is the single ERC-20 governance token; FRS is an asset class where each launch (FRS-wBTC, FRS-USDC, etc.) is its own ERC-20 paying yield in the denominated asset. Vestibulum is the onboarding gateway; Tribunus is the one-address-one-vote DAO.
Content-Addressed Steps: How opentine Guarantees Bit-Exact Replays
opentine hashes every agent step from its inputs — prompt, tools, model choice, parent step — into a stable ID. Two steps with the same inputs always produce the same ID. That makes replay bit-exact, caching automatic, and forking a cheap pointer operation rather than an expensive re-execution.
opentine vs LangGraph vs Temporal vs DSPy: Choosing an Agent Runtime
LangGraph fits LangChain users who don't need forkable runs. Temporal fits durable-workflow problems where agents are incidental. DSPy fits prompt-compilation research. opentine fits when run forkability, bit-exact replay, and model-agnosticism are the primitive you're building on.
opentine: Why the Fork Primitive Matters
opentine treats every agent step as a content-addressed node in a DAG so you can fork, replay, and diff runs the way git lets you branch and rebase code.
Sub-100ms Finality via DAGKNIGHT Cascade Voting, Explained
DAGKNIGHT is a parameterless block DAG consensus protocol. Instead of a fixed-parameter security / latency tradeoff, it adapts confirmation criteria dynamically based on observed network conditions. Solidus uses DAGKNIGHT cascade voting to reach ~100ms finality while preserving decentralized security guarantees.
An Open-Source Alternative to Hyperliquid: The Case for Solidus
Hyperliquid is a closed-source, centrally-operated L1 that proves on-chain perps can match CEX latency. Solidus is an open-source, permissionless alternative: 16-chain DAG architecture, DAGKNIGHT finality at ~100ms, POEM multi-algorithm mining, non-custodial asset security, and the ForumCore perps engine.