Findings & Insights

Every finding traces to a real run or probe. We publish what DOESN'T work too — honesty is the product. The discipline that reports a failed edge is worth more than the edge.

The full report — all 19 findings, each with its source file or measured run — lives in one place: Findings/FINDINGS.md at the root of this repository. This page only summarizes it, so the two never drift.

In short

• Smart money & whales — on Rhea's wNEAR pools, trade direction comes from the token pair, not the kind label; the whale threshold decides whether there is any signal at all, and NEAR flow is concentrated.
• Alpha & strategies — on NEAR the edge lab tested funding-fade (−28.52% OOS) and price-momentum (−35.04% OOS) after costs. Neither passed, the validated registry is empty, and the firm abstains. The regime classifier beats persistence (78.9% vs 76.1%), but accuracy is not profit.
• Execution & risk — gated tickets (R:R ≥ 2, a real invalidation), aggression earned by data, and a token-security veto that fails closed.
• NEAR & verifiability — limited function-call keys, sealed desk outputs, anchored decisions, an escrow vault, and an oracle relay with an honest freshness trade-off.
• Method — honesty by design: deterministic tools compute, LLMs debate, and a firm of twelve analyst voices has its decisions anchored on NEAR.