What (One-liner)

Canvas LLM-in-the-loop strategy generation: regime + symbols → Fin-R1 generates Python pair-trading code + rationale → optional Python simulation. Tier 5 integrity-check track (not Tier 4 numerical tolerance).

How to use

macro_factor_v1 (regime=expansion) → casa_strategy_v1 (SPY+TLT)
                                      → strategy_code (momentum pair trade)
                                      → [run-sandbox Docker execution]
                                      → risk_management_v1

Core formulas

See frontmatter. Flow: generate-strategy LLM → extract Python code → optional dual-agent-simulate → combined return.

Assumptions & applicability

Assumptions: Fin-R1 / cloud LLM available + code block extraction works + Python sim tail already golden.

Applies to: macro-regime-driven auto-generated pair strategies, scenario analysis, LLM-finance demo.

Does not apply to: live trading, HFT, byte-deterministic regulator demos.

Known limitations

See frontmatter. Six items: LLM non-determinism / prompt drift / mock fallback / independent Python sim / LLM latency / sandbox independence.

References

Honest status (2026-08-03, YG decision): this operator has NO academic anchor. The regime→weight table is hand-set heuristics (not estimated from any mechanism or data), and the LLM lane is a capability demonstration. CASA is positioned as a dual-agent mechanism DEMO / teaching vessel, not an analysis tool. For user-authored strategies with real provenance (determinism badge, sandbox, fingerprints), use the Script node (Xlang / Python lanes).

If the planned academicization lands (board item C-y), the canonical literature to anchor against — cited here as the roadmap, NOT as current implementation basis:

• Gatev, Goetzmann & Rouwenhorst (2006) — Pairs Trading: Performance of a Relative-Value Arbitrage Rule
• Jegadeesh & Titman (1993) — Returns to Buying Winners and Selling Losers (momentum)
• Hamilton (1989) — A New Approach to the Economic Analysis of Nonstationary Time Series (regime switching)
• Brock & Hommes (1998) — Heterogeneous beliefs and routes to chaos (bull/bear heterogeneous agents)

Golden Test

Tier 5 integrity-check track. Python tail (simulate_dual_agent) covered by tests/golden/fixtures/tier4/casa_dual_agent/. LLM part requires: prompt_cid + model pin + automated integrity-check roundss (after Phase B framework lands).

Changelog

• 1.0.0 (2026-04-21) — First Active (Tier 5 integrity-check track)