Panel Ruling: NBA Player Props Research & Intelligence Pipeline

Date: 2026-03-31 Panel: Opus 4.6, Sonnet 4.6, Gemini 3.1 Pro, Grok 3, GPT-4.1 (via OpenRouter) Topic: How should research/intel work for the NBA Player Props desk Winner: Opus (Advisor B) — scored 1st by multiple reviewers

Consensus Decisions (all 5 advisors agree)

1. Separate Pipeline

Build a separate player props research pipeline, not an extension of the team module. Player props research has different timing, different queries, and different output format than team-level research.

2. Multi-Pass Timing

Run multiple passes, not a single pre-game sweep:

3. Hybrid Auto-Adjust with Confidence Gating

• AUTO-ADJUST (no human review): Confirmed OUT status, official minutes restrictions, verified lineup changes
• FLAG for review: Ambiguous injury language ("questionable," "game-time decision"), narrative/rumor intel, coaching hints
• Cap: No single auto-adjustment exceeds 20% on any parameter

4. Minutes-Based Scaling (sqrt rule)

When a player's projected minutes change:

• Mean scales proportionally: new_mean = base_mean * (new_minutes / base_minutes)
• Stdev scales by square root: new_stdev = base_stdev * sqrt(new_minutes / base_minutes)
• This is mathematically correct — variance in counting stats scales linearly with opportunity, so stdev scales by sqrt. Linear stdev scaling (which some advisors proposed) understates uncertainty for minutes-restricted players.

5. Usage Cascade Model

When a teammate is confirmed OUT:

• Redistribution factor: Only ~60% of a missing player's usage gets redistributed (not 100%)
• Stat-specific sensitivity multipliers: Points 1.0x, Assists 0.7x, Rebounds 0.4x (validates against existing 3012-row game log)
• Rebound redistribution is NOT even: 70% to direct replacement, 20% other frontcourt, 10% guards
• Formula: adjusted_stat = base_stat + (missing_player_stat * absorption_rate * redistribution_factor * sensitivity_multiplier)

6. Key Data Sources

7. Output Format

Structured JSON extraction from LLM (not free-text parsing):

interface IntelAdjustment {
  player_name: string;
  stat_type: string;
  adjustment_type: 'mean_shift' | 'variance_shift' | 'minutes_cap';
  magnitude: number;
  confidence: 'HIGH' | 'MEDIUM' | 'LOW';
  source: string;
  timestamp: string;
}

Key Unique Insights

Opus (winner):

• Shared-runner/pluggable-profile design that scales to all 13 desks
• Sigma scales with sqrt of minutes ratio (mathematically correct)
• 0.6 redistribution factor for absent player usage
• "Kill switch" Telegram command (freeze props) to instantly disable all auto-adjustments
• Blowout risk should both reduce minutes AND increase sigma by 1.15x

Sonnet:

• Stat-specific sensitivity multipliers (0.4 reb, 0.7 ast) with regression validation against game logs
• Milestone effects as mixture distribution (90% base, 10% "milestone push" with mean+3, halved stdev)
• Confidence gating with clear HIGH/MEDIUM/LOW action rules

Gemini:

• Response was truncated (cut off) — limited evaluation possible

Grok:

• Solid foundational coverage, correct on hybrid auto-adjust and 20% cap
• DFS ownership as sentiment overlay

GPT-4.1:

• Event-driven PIM should subscribe to team module output as trigger
• Coach pre-game media transcripts as specific query target

Blind Spots (all 5 missed)

• Live in-game research for second-half props or live betting adjustments — all advisors focused exclusively on pre-game research

Raw council files: /home/ubuntu/edgeclaw/data/councils/2026-03-31/player-props-research/