Methodology

Data ingests news, stats, and filings with semantic dedup. Research maintains platform-aware retrieval (SmartSearch over RootMem rules + corpus). Simulation stress-tests strategies as paper trades under enforced platform rules. Orchestration hill-climbs the weakest platform slice. No stage places live orders.

Each platform's execution mechanics — settlement, position caps, payout structures, margin rules — live as structured registry entries with cited evidence (config/market-platforms.json). The simulator refuses trades that violate them, and every trade log records the rule ids it applied. A metric that can't cite its rules doesn't publish.

CPU-only, LLM-free recall over platform rules and corpus: weighted term parse (proper nouns > nouns > verbs), substring recall, reciprocal-rank fusion, and a score-adaptive cutoff. Deterministic by construction — the same query against the same corpus always retrieves the same context.

Strategies are markdown skills, optimized in text space: roll out paper trades, reflect on failures, propose bounded add/delete/replace edits, and gate every edit on held-out validation Sharpe. Rejected edits enter a buffer so the loop never re-proposes a known failure.