Invest AI ships a production dashboard with macro consensus. Decision Engine owns multi-timeframe technical scoring and backtests. Two repos on purpose — ADR-002 on the Decision Engine side, project comparison on Invest AI.

Integration needs a contract both sides can enforce. That contract is daily_scores.v1.json.

Layer 1 vs Layer 2

Layer	Owner	Job
Layer 1 — Technical	Decision Engine	M/W/D/K composite, hard gates, StockType caps, regime block
Layer 2 — Macro overlay	Invest AI	FRED, event calendar, combo modifiers, narrative

Decision Engine must not import FRED or LLM explanation layers. Invest AI must not reimplement scoring formulas once import is enabled (ADR-026).

Export contract (Decision Engine ADR-011)

Canonical artifact: daily_scores.v1.json

Key fields:

• version, rules_version, as_of_date (≤ T−1 for day T)
• regime.level — technical regime
• scores[] — per-ticker composite, setup, action, confidence

Production gate: export does not flow to Invest AI production cache until ADR-009 four-regime backtests pass for the referenced rules_version.

Import pipeline (Invest AI ADR-026)

Decision Engine
  → daily_scores.v1.json
       ↓
stock-ai-worker (decision_engine_importer.py)
  → validate schema + rules_version + as_of_date
  → merge into build_payload()
  → macro_radar overlay (caps/discounts only)
  → SQLite v2:dashboard:today
       ↓
FastAPI (read only — ADR-012 unchanged)

Hard rules:

1. FastAPI never fetches Decision Engine at request time.
2. Worker does not recompute composite from raw OHLCV once import is on.
3. Import failure → keep last good cache; set decision_engine_import.status: degraded.

Regime mapping

Decision Engine regime.level	Invest AI macro_radar.regime_gate
RISK_ON	risk_on
NORMAL	neutral
CAUTION	caution
DEFENSIVE	defensive

Technical regime is informational in payload field decision_engine.regime. Macro gate remains authoritative for combo modifiers.

Phase 2 boundaries (Decision Engine ADR-012)

Before coding indicators and scoring modules:

• engine/ writes export; scheduler/ runs 4:30pm ET job.
• api/ is read-only — no on-demand scoring over HTTP.
• See ADR-012.

Why not merge repos?

• Different release cadence (quant rules vs product UX).
• Backtest gate blocks bad math before users see it.
• Invest AI cache boundary (ADR-012) stays intact.

Status today: Phase 1 data layer in Decision Engine; ADRs and schema defined; worker import path planned in ADR-026 implementation checklist.