DIME
Get the picksJoin

PROJECT · findings note

project_debate_persona_lockdown

Multi-persona debate (Koerner/Levitan/Raybon + judge) now reasons about FIX 1-5 lockdown flags + line_edges per pick. Closes the picker-facing gap end-to-end. Shipped 2026-05-21 via sport-agents e7be3cf + debate-v2 d7525b2.

← All findingssource · memory/project_debate_persona_lockdown.md

The actual production pick chain (corrected mental model after the user pointed out my earlier description was wrong):

sport-data-service (data SoT)
   ↑                    ↑
   │                    │
sports-scout-service   data-hydrator
(own collectors;       (assembler + FIX 1-5 lockdown)
 ML pick generation)    │
   │                    │
   └────────────────┐  │
                    ↓  ↓
            snapshot.scout.player_props
            (picks from scout, surfaced by hydrator)
                    │
                    ↓
            debate-service-v2 orchestrator
              - extractPropsFromSnapshot
              - lockdown pre-filter (drops Out/Doubtful/Questionable)
              - AnalyzeProps → sport-agents /analyze
                    │
                    ↓
            sport-agents /api/v1/analyze (multi-persona)
              - Koerner  (Projection Modeler)  ─┐
              - Levitan  (Sharp Process)        ├─ each reads
              - Raybon   (Stat Hunter)          │  Lockdown signals
              - (judge synthesizer)            ─┘  per pick
                    │
                    ↓
            verdicts → UI

Two corrections to the earlier mental model:

  1. sports-scout-service does NOT call data-hydrator. They are PARALLEL SDS consumers. sports-scout-service has its own collectors (sport_data_service_client.py, injury_collector, roster_collector, etc.) and runs its own ML pipeline. data-hydrator pulls sports-scout-service's outputs (likely via ml-scout-service) and includes them as snap.scout.player_props.

  2. sport-agents /analyze is MULTI-PERSONA. Three analyst personas (Koerner: projection modeler / Levitan: sharp process / Raybon: stat hunter) + judge synthesizer. Each persona has its own system prompt + methodology and emits AGREE/DISAGREE + confidence

    • reasoning. Judge synthesizes into final verdict.

What shipped (2 commits, end-to-end)

sport-agents commit e7be3cf

  • api/server.py: Pick.prop_diagnostic: Optional[dict] field.
  • core/debate_engine.py: per-pick prompt block renders a **Lockdown signals** section showing:
    • active_status (informational — drops happen upstream in debate-v2)
    • recent_vs_baseline_flag when not stable (with season_avg)
    • role_anchor_adjusted when true (with current vs historical minutes)
    • minutes_source_confidence when not stable (especially diverge_low_confidence)
    • line_edges entry for the pick's exact line (quality + z-score) Each signal only rendered when its value is non-default so noise stays low for clean players.

debate-v2 commit d7525b2

  • internal/sportagents/client.go:
    • Pick.PropDiagnostic (json.RawMessage).
    • AnalyzeProps signature gains propDiagnostics map[string]json.RawMessage (player_id → raw diagnostic, same pattern as playerMemory).
    • Per-pick attachment in the conversion loop.
  • internal/orchestrator/orchestrator.go:
    • buildPropDiagnosticMap(snap) walks snap.Players.Players[*].PropDiagnostic into the map.
    • Phase 5 call passes the map.

Why each persona interprets differently (sport-agents prompt design)

Each persona's existing system prompt reflects their methodology. The Lockdown signals are presented as DATA (not commandments), so each persona reads them through their lens:

  • Koerner (proprietary projection models, regression analysis, historical baselines): heavily weights role_anchor_adjusted + recent_vs_baseline_flag — these touch projection drift.
  • Levitan (sharp process, market edge, line analysis): weighs line_edges.quality and line_z_score — these speak to his mean-relative-edge thinking. He'll DISAGREE strongly on a no_edge quality line.
  • Raybon (situational splits, rate-based metrics): weighs minutes_source_confidence + recency signals — these align with his split-aware analysis.

The judge sees all persona verdicts AND has the same lockdown context (via the picks_dicts passed through to run_judge), so it can arbitrate based on which persona's reasoning best matches the lockdown signals.

What's still NOT in this chain

  • sports-scout-service still doesn't read prop_diagnostic. Its ML models can emit picks that violate lockdown gates (the picks reaching debate-v2 are filtered, but the ML pipeline itself is unchanged). Making sports-scout-service lockdown-aware would reduce wasted compute (don't generate picks that get dropped) but isn't required for production-pick correctness — the filter + debate gates handle it.
  • line_edges no_edge quality is currently INFORMATIONAL in the debate prompt (personas see it; can ignore it). debate-v2's pre-filter only drops active-status violations. A separate follow-up could also drop no_edge-only picks before debate.

See also: project_methodology_lockdown (FIX 1-4 origin), feedback_active_tonight_gate (FIX 5), project_wnba_lockdown_live (data-layer end-to-end for WNBA), project_wnba_sport_agents_playbook (WNBA picker adapter).