Apex2 Design Study — Design skill for Claude Code

Apex 2 — design study for a BTC perpetual-futures bot (Stage 0)

**One-line result:** across 150 market signals tested at six horizons on 6.4 years of Bybit BTCUSDT data, nothing predicts BTC on its own beyond noise; a trend-filtered, volatility-targeted long halves the worst drawdown (79% → 33%) but does not beat holding BTC statistically; one fragile lead survives (scale exposure by Binance long/short account crowding), sitting exactly at its family-level significance hurdle. The return band these edges support is about 2–5% a month with 25–50% drawdowns; 10% a month is not reachable at any leverage.

Everything here was produced on 2026-09-04 by Claude (Fable 5.1) working with the architect, Rafael Vargas, using multi-agent workflows: 51 agents (feature builders, validators, blind implementation pairs, reconcilers, blind verifiers, judges, critics, writers) plus the main session. It is published as-is so the method, the numbers and the mistakes can be inspected.

Read this first

• docs/STAGE0_REPORT.md — the verdict, the ladder, the fee map, the census, judges, honesty section, Stage 1 proposal, sign-off block, addendum (PT-BR summary at the end).
• docs/SPEC.md — the approved study design (stages, kill rules, simulator requirements, roadmap).
• docs/APEX2_SIM_DESIGN.md — the simulator design chosen by a 3-architect / 3-judge panel, with an 11-step implementation plan.

Layout

path	what
contracts/	the written contracts every agent worked from (common rules, feature schema, ten feature families, blind ladder pair, fee map, prober, validator, verifier, judges, critic, report, simulator design panel)
harness/	the census instrument (data loaders, block-bootstrap statistics, ladder, census runner with registered pass rules, calibration on synthetic features, sensitivity, summaries)
features//	each family's build script, manifest (definitions, availability lags, registered expected signs) and validation report