
YuhoLens turns Japan's most opaque corporate disclosures into auditable investment memos. Yuho reports, the annual securities filings every TSE-listed company submits to EDINET run hundreds of pages of dense Japanese, and analysts spend days hunting for the single paragraph that justifies a price target. We fine-tuned a 14B open-weight LLM on a two-pass extraction of the EDINET-Bench corpus. KG-1 captures structural sections (segments, risks, MD&A); KG-2 captures the reasoning links between them. On top of the base model, a 4-agent LangGraph pipeline (Drafter → Critic → Citer → Editor) composes equity memos where every factual claim carries a (ref:filing#section) anchor back to the source filing. Training used ORPO on synthetic preference pairs gated by a canonical citation-rate regex, so the reward signal cannot be gamed by paraphrase. We ran a BO5 selection across checkpoints and picked the variant with the highest gated win-rate on a held-out memo set. The release ships as HuggingFace safetensors plus a GGUF quantization tuned for AMD ROCm inference. A public reproducibility ledger, surfaced on yuholens.site, lets anyone re-run the evaluation that selected the shipped checkpoint, inspect every (filing, memo, citation) triple, and verify that the model's claims survive grep against the underlying yuho.
10 May 2026

Praxis is a regime-adaptive cryptocurrency trading agent built around a hierarchical risk governor that the LLM can never override. Five deterministic signal agents (trend, volatility, momentum, mean-reversion, swing structure) vote on direction via plurality consensus; an ADX-based regime classifier modulates which agents are active; and a half-Kelly position sizer draws its win/loss ratio empirically from trailing trade history. The system executes on two surfaces simultaneously. Kraken CLI paper trades run alongside EIP-712 signed TradeIntents posted to the ERC-8004 Risk Router contract on Sepolia, so every validated signal is recorded both off-chain and on-chain. A deterministic risk governor enforces six hard kill criteria — stale data, daily loss cap, max drawdown, consecutive losses, spread-too-wide, volatility shock — plus two signal-quality gates, before any trade reaches execution. The heart of the submission is a 24-page methodology paper. The backtest framework covers BTC/USD since December 2013 and ETH/USD since October 2015 (12.4 and 10.5 years), models realistic Kraken fees and volatility-scaled slippage, and includes a full statistical validation stack: three-fold walk-forward, Probabilistic Sharpe Ratio, Deflated Sharpe correction for multiple-trial bias, Jobson-Korkie-Memmel two-sample Sharpe test, Monte Carlo permutation tests, per-agent and per-feature ablations, and cost sensitivity across four Kraken fee tiers. Out-of-sample (2023-2026, 88 trades), Praxis delivers a Sharpe of 1.24 with max drawdown 8.1%, beating BTC buy-and-hold, ETH buy-and-hold, 50/50 equal-weight, and a cost-free EMA crossover baseline on risk-adjusted return. Every number in the paper is reproducible from a single command, with fixed random seeds and a CI-level reproducibility assertion guarding the reference configuration.
12 Apr 2026