Pre-registration: LLM trading agents in a live crypto paper-trading arena (15-day run, Claude Haiku, Sonnet and Opus)

Pre-registered design and exact code of a 15-day live paper-trading study of large language model (LLM) trading agents on cryptocurrency markets, deposited during the run and before any outcome analysis. Setting. 24 paper accounts (US$ 197 each) trade BTC, ETH and SOL on live Pionex prices (spot and 2x perpetuals) from 2026-11-01 21:01:25 UTC to 2026-11-16 21:01:25 UTC. AI arms use Claude Haiku 5.5, Claude Sonnet 5.5 and a Claude Opus 5.5 agent with read-only market tools, all through strict JSON outputs; code sizes every position and places every stop. They run against rule-based arms, mechanical playbook twins, a random-entry null, buy-and-hold, and three forced-direction arms that must take long or short on every asset every hour. Confirmatory analysis. Gate G0 (pooled decision information coefficient at 1 h of six AI streams, stationary block bootstrap), then Holm across C1–C4; a separate Holm family F1–F3 for the forced-direction arms; an exploratory tier under Benjamini–Hochberg (q = 0.10). Duration fixed at 15 days in advance, with no early stop or extension. Contents. PREREGISTRATION.md (plan and four amendments), the arena source code and test suite at commit 2a384ffc6327758a8a5f32bc5c6a6bbd4a92ac9c, Dockerfile, compose.yaml and SHA256 checksums.