Upload alpha_factory/personas/gatekeeper.py with huggingface_hub

alpha_factory/personas/gatekeeper.py

@@ -0,0 +1,86 @@
+"""
+Production Gatekeeper — Persona 6 (Bigfish, run SPARINGLY)
+Final go/no-go decision on promoted alphas.
+Only invoked for alphas that survived all upstream gates.
+"""
+from ..infra.llm_client import LLMClient
+from ..schemas import BrainMetrics, GatekeeperMemo, Blueprint
+
+
+SYSTEM_PROMPT = """You are a senior portfolio manager with 20 years of experience.
+You are making the FINAL decision on whether to add this alpha to a production portfolio.
+
+You have seen the full pipeline output: the hypothesis, the expression, the BRAIN metrics,
+the crowd scout's novelty assessment, and the performance surgeon's diagnosis.
+
+Your memo must address:
+1. STRENGTHS — what makes this alpha worth running?
+2. WEAKNESSES — what could go wrong?
+3. RISKS — regime fragility, crowding, capacity constraints
+4. RECOMMENDATION — go (add to portfolio) or no-go (reject)
+
+Standards for go:
+- Sharpe OS ≥ 1.25
+- No single-year catastrophic loss (> -15%)
+- Turnover reasonable for the signal (not churning)
+- Genuinely orthogonal to existing book (corr < 0.65)
+- Academic rationale is defensible (not curve-fitted)
+- Would you bet your own money on this? If no → no-go.
+
+You must be SKEPTICAL by default. The bar for "go" is HIGH.
+Most alphas should get no-go. That's correct behavior.
+"""
+
+
+async def gate_alpha(
+    llm: LLMClient,
+    blueprint: Blueprint,
+    metrics: BrainMetrics,
+    max_corr: float,
+    fitness_score: float,
+    model: str | None = None,
+) -> GatekeeperMemo:
+    """
+    Final production gate. Invoked ONLY for alphas that passed all prior gates.
+    Uses Bigfish model (72B) for maximum reasoning quality.
+    """
+    user_prompt = f"""PRODUCTION GATE REVIEW
+
+## Alpha Blueprint
+- Theme: {blueprint.theme}
+- Anomaly: {blueprint.anomaly_tag.value}
+- Archetype: {blueprint.archetype}
+- Neutralization: {blueprint.neutralization.value}
+- Decay: {blueprint.decay}
+- Components: {len(blueprint.components)}
+- Academic anchor: {blueprint.academic_anchor or 'none cited'}
+- Novelty claim: {blueprint.novelty_claim}
+
+## BRAIN Metrics
+- Sharpe (full): {metrics.sharpe_full:.3f}
+- Sharpe (IS): {metrics.sharpe_is:.3f}
+- Sharpe (OS): {metrics.sharpe_os:.3f}
+- Fitness: {metrics.fitness:.3f}
+- Turnover: {metrics.turnover:.3f}
+- Max Drawdown: {metrics.max_drawdown:.3f}
+- Returns: {metrics.returns:.3f}
+- Yearly Sharpe: {metrics.yearly_sharpe}
+
+## Novelty Assessment
+- Max correlation to library: {max_corr:.3f}
+- Fitness score (composite): {fitness_score:.3f}
+
+## Decision Required
+Write a 1-page production memo. Be SKEPTICAL. 
+Would you bet your own capital on this alpha? 
+Output your go/no-go decision with confidence level (0.0 to 1.0)."""
+
+    memo = await llm.generate_json(
+        prompt=user_prompt,
+        schema=GatekeeperMemo,
+        model=model or llm.config.bigfish_model,
+        temperature=0.3,
+        system_prompt=SYSTEM_PROMPT,
+    )
+
+    return memo