polly-gamba 1.0.42 → 1.0.43

package/dist/claude-trader.js

@@ -103,7 +103,7 @@ class ClaudeTrader {
                     role: 'user',
                     content: `You are a Polymarket ${PAPER_MODE ? 'PAPER TRADING (no real money)' : 'paper'} trader running a high-volume moneyball strategy. Your job is to place paper trades on EVERY market where you have ANY opinion on fair value — even slight.
 ${PAPER_MODE ? '\n⚠️ PAPER MODE: This is a test environment. Be AGGRESSIVE — trade more, skip less. Lower your threshold to 3% edge for any market resolving within 6 months.\n' : ''}
-TOOLS: place_order, skip_all, get_budget_status, get_positions
+TOOLS: place_order, skip_all, get_budget_status, get_positions, scan_smart_money, get_wallet_positions, get_late_money, get_market_trades
 RULES:
 - Output ONLY tool calls. Zero prose.
 - For EVERY market in the list: if current price differs from your estimated fair probability by more than ${PAPER_MODE ? '3%' : '5%'}, place a trade.
@@ -116,8 +116,16 @@ RULES:
 ## POSITION DISCIPLINE:
 - Max $100 per market (20% of $500 budget). The MCP enforces this — don't fight it.
 - Same-direction re-entry is NEVER allowed while a position is open. You CANNOT average down into an existing open position. The MCP enforces this hard block. If you still believe in the thesis after a price drop, close the existing position first, then re-enter fresh next scan.
+- CORRELATED PAIRS: Holding YES on one election outcome + NO on the complementary outcome is one correlated exposure, not two independent bets. Count the combined notional against your per-theme cap ($20 total). They cannot both win.
 - exit_trigger is required on every trade. Be specific: "Exit when price hits 0.X" or "Exit when [specific news event]" — not "when narrative converges."
-- Call get_budget_status AND get_positions at the start of each scan to know available capital and your current open positions. get_positions returns {"positions": [...], "closed_markets": [...]}. You MUST review BOTH before trading to apply concentration AND closed-market discipline correctly.`
+- Call get_budget_status AND get_positions at the start of each scan to know available capital and your current open positions. get_positions returns {"positions": [...], "closed_markets": [...]}. You MUST review BOTH before trading to apply concentration AND closed-market discipline correctly.
+
+## SMART MONEY (poly-scout tools — use when evaluating edge):
+- scan_smart_money(): Returns top Polymarket wallets by ROI + their consensus positions. If 3+ smart wallets are on the same side of a market you're evaluating, treat it as +5pp edge boost on that side.
+- get_late_money(market_id): Detects unusual volume buildup in last 48h vs baseline. A strong late-money signal (confidence > 0.7) on a market you're already considering = stronger conviction. Do NOT trade solely on late money — use it as a corroborating signal alongside your own fair value estimate.
+- get_wallet_positions(address): Use to verify specific wallet's current positions when scan_smart_money flags them.
+- get_market_trades(token_id): Use to inspect raw trade flow on a specific market when you want to verify smart money claims.
+- Smart money is a signal amplifier, not a replacement for thesis. Always form your own fair value estimate first.`
                 }
             }));
             this.ready = true;
@@ -227,6 +235,35 @@ ${m.description ? `- Description: ${m.description.slice(0, 200)}` : ''}`).join('
 - Anchor fair value on base rates and world knowledge first; use sportsbooks to adjust by ±5pp only. If you cannot defend the estimate without the sportsbook line, the edge is too speculative.
 - Exception: if both Polymarket AND sportsbooks agree with your base-rate analysis, the convergence strengthens the thesis.
 
+## ELECTION MARKET DISCIPLINE (apply when any market involves a political election, referendum, or leadership transition):
+
+### Turnout Shock Screening
+Before sizing any election bet, assess turnout risk. If 2+ of these signals are present, apply a 40% DISCOUNT to any structural incumbency advantage in your fair-value estimate:
+- Pre-election voter registration spike >15% above the prior comparable election cycle
+- Opposition rally sizes or enthusiasm signals markedly above historical norms
+- "Change election" or "historic election" framing dominant across independent media
+- Predicted or early-voting turnout above the country's historical 90th percentile
+- Incumbent has been in power >12 years (fatigue compounds at this threshold)
+Rationale: Structural advantages (gerrymandering, media control, state resources) are calibrated for normal-turnout elections. When mobilization is historic, marginal voters break disproportionately for the opposition, inverting the structural model.
+
+### Historical Analog Quality
+When citing a historical precedent to anchor fair value (e.g. "in 2022, X happened"), you MUST assign it an explicit structural similarity score (0–100%) and state the key differences:
+- Score ≥80%: strong anchor, use as primary prior
+- Score 50–79%: weak anchor, adjust by ±5pp only, state the differences
+- Score <50%: discard the analogy, use base rates instead
+Example of a BAD analogy: Hungary 2026 ← Hungary 2022. Similarity ~50%: 2022 had 6-party unified opposition (fragmented, easy to split); 2026 had a single-party challenger (Magyar/TISZA, clean brand). Different structural conditions. Do not use 2022 to anchor Orbán's survival probability in 2026.
+
+### Regime-Fatigue Flag
+If an incumbent has been in power >10 years AND visible economic pain or corruption signals exist (inflation, stagnant wages, high-profile corruption exposed in mainstream coverage): flag this as a REGIME-FATIGUE ELECTION. In regime-fatigue elections, reduce structural incumbency advantage estimates by 30% before computing fair value. This reflects the empirical pattern that structural advantages erode faster than polling captures when voter frustration crosses a threshold.
+
+### Correlated Hedge Pair Sizing
+If you hold YES on one outcome AND NO on the complementary outcome in the same election (e.g., "Orbán becomes PM" YES + "Magyar becomes PM" NO), treat both positions TOGETHER as one correlated exposure unit:
+- Combined notional = sum of both position sizes
+- Both positions cannot simultaneously win — the upside and downside are correlated, not independent
+- Cap the combined notional at $20 (2 positions × $10), not $20 per leg
+- In the concentration table, label these as "HEDGE PAIR — one correlated exposure" and count them as 2/2 (theme FULL)
+- Do NOT add a third position to the same election even if the third market looks mispriced — you are already fully exposed to the outcome
+
 STEP 1: Call get_positions. Then output a THEME CONCENTRATION TABLE — list every theme with ≥1 open position, showing count/max (e.g. "2028 US election: 3/2 FULL", "GTA VI: 2/2 FULL", "Hungary 2026 election: 1/2"). Any theme showing FULL must be skipped entirely. You MUST output this table before evaluating any market.
 STEP 2: For each market, check your theme concentration table first. If the market's theme is FULL, skip it immediately. Otherwise apply horizon discipline, closed market discipline, and edge check.
 STEP 3: If price differs from your fair probability by the required edge (accounting for sportsbook haircut if applicable), place a trade ($10 USDC). Call skip_all only if you have zero opinion on all markets.`;

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "polly-gamba",
-  "version": "1.0.42",
+  "version": "1.0.43",
   "description": "Coinbase price signal → Claude brain → Polymarket CLOB execution",
   "main": "dist/index.js",
   "bin": {

package/settings.json

@@ -7,6 +7,13 @@
         "REDIS_URL": "redis://localhost:6379",
         "POLLY_REDIS_PREFIX": "polly"
       }
+    },
+    "poly-scout": {
+      "command": "npx",
+      "args": ["--yes", "-y", "@gonzih/poly-scout"],
+      "env": {
+        "REDIS_URL": "redis://localhost:6379"
+      }
     }
   }
 }

package/src/claude-trader.ts

@@ -120,7 +120,7 @@ export class ClaudeTrader {
           role: 'user',
           content: `You are a Polymarket ${PAPER_MODE ? 'PAPER TRADING (no real money)' : 'paper'} trader running a high-volume moneyball strategy. Your job is to place paper trades on EVERY market where you have ANY opinion on fair value — even slight.
 ${PAPER_MODE ? '\n⚠️ PAPER MODE: This is a test environment. Be AGGRESSIVE — trade more, skip less. Lower your threshold to 3% edge for any market resolving within 6 months.\n' : ''}
-TOOLS: place_order, skip_all, get_budget_status, get_positions
+TOOLS: place_order, skip_all, get_budget_status, get_positions, scan_smart_money, get_wallet_positions, get_late_money, get_market_trades
 RULES:
 - Output ONLY tool calls. Zero prose.
 - For EVERY market in the list: if current price differs from your estimated fair probability by more than ${PAPER_MODE ? '3%' : '5%'}, place a trade.
@@ -133,8 +133,16 @@ RULES:
 ## POSITION DISCIPLINE:
 - Max $100 per market (20% of $500 budget). The MCP enforces this — don't fight it.
 - Same-direction re-entry is NEVER allowed while a position is open. You CANNOT average down into an existing open position. The MCP enforces this hard block. If you still believe in the thesis after a price drop, close the existing position first, then re-enter fresh next scan.
+- CORRELATED PAIRS: Holding YES on one election outcome + NO on the complementary outcome is one correlated exposure, not two independent bets. Count the combined notional against your per-theme cap ($20 total). They cannot both win.
 - exit_trigger is required on every trade. Be specific: "Exit when price hits 0.X" or "Exit when [specific news event]" — not "when narrative converges."
-- Call get_budget_status AND get_positions at the start of each scan to know available capital and your current open positions. get_positions returns {"positions": [...], "closed_markets": [...]}. You MUST review BOTH before trading to apply concentration AND closed-market discipline correctly.`
+- Call get_budget_status AND get_positions at the start of each scan to know available capital and your current open positions. get_positions returns {"positions": [...], "closed_markets": [...]}. You MUST review BOTH before trading to apply concentration AND closed-market discipline correctly.
+
+## SMART MONEY (poly-scout tools — use when evaluating edge):
+- scan_smart_money(): Returns top Polymarket wallets by ROI + their consensus positions. If 3+ smart wallets are on the same side of a market you're evaluating, treat it as +5pp edge boost on that side.
+- get_late_money(market_id): Detects unusual volume buildup in last 48h vs baseline. A strong late-money signal (confidence > 0.7) on a market you're already considering = stronger conviction. Do NOT trade solely on late money — use it as a corroborating signal alongside your own fair value estimate.
+- get_wallet_positions(address): Use to verify specific wallet's current positions when scan_smart_money flags them.
+- get_market_trades(token_id): Use to inspect raw trade flow on a specific market when you want to verify smart money claims.
+- Smart money is a signal amplifier, not a replacement for thesis. Always form your own fair value estimate first.`
         }
       }))
       this.ready = true
@@ -255,6 +263,35 @@ ${m.description ? `- Description: ${m.description.slice(0, 200)}` : ''}`).join('
 - Anchor fair value on base rates and world knowledge first; use sportsbooks to adjust by ±5pp only. If you cannot defend the estimate without the sportsbook line, the edge is too speculative.
 - Exception: if both Polymarket AND sportsbooks agree with your base-rate analysis, the convergence strengthens the thesis.
 
+## ELECTION MARKET DISCIPLINE (apply when any market involves a political election, referendum, or leadership transition):
+
+### Turnout Shock Screening
+Before sizing any election bet, assess turnout risk. If 2+ of these signals are present, apply a 40% DISCOUNT to any structural incumbency advantage in your fair-value estimate:
+- Pre-election voter registration spike >15% above the prior comparable election cycle
+- Opposition rally sizes or enthusiasm signals markedly above historical norms
+- "Change election" or "historic election" framing dominant across independent media
+- Predicted or early-voting turnout above the country's historical 90th percentile
+- Incumbent has been in power >12 years (fatigue compounds at this threshold)
+Rationale: Structural advantages (gerrymandering, media control, state resources) are calibrated for normal-turnout elections. When mobilization is historic, marginal voters break disproportionately for the opposition, inverting the structural model.
+
+### Historical Analog Quality
+When citing a historical precedent to anchor fair value (e.g. "in 2022, X happened"), you MUST assign it an explicit structural similarity score (0–100%) and state the key differences:
+- Score ≥80%: strong anchor, use as primary prior
+- Score 50–79%: weak anchor, adjust by ±5pp only, state the differences
+- Score <50%: discard the analogy, use base rates instead
+Example of a BAD analogy: Hungary 2026 ← Hungary 2022. Similarity ~50%: 2022 had 6-party unified opposition (fragmented, easy to split); 2026 had a single-party challenger (Magyar/TISZA, clean brand). Different structural conditions. Do not use 2022 to anchor Orbán's survival probability in 2026.
+
+### Regime-Fatigue Flag
+If an incumbent has been in power >10 years AND visible economic pain or corruption signals exist (inflation, stagnant wages, high-profile corruption exposed in mainstream coverage): flag this as a REGIME-FATIGUE ELECTION. In regime-fatigue elections, reduce structural incumbency advantage estimates by 30% before computing fair value. This reflects the empirical pattern that structural advantages erode faster than polling captures when voter frustration crosses a threshold.
+
+### Correlated Hedge Pair Sizing
+If you hold YES on one outcome AND NO on the complementary outcome in the same election (e.g., "Orbán becomes PM" YES + "Magyar becomes PM" NO), treat both positions TOGETHER as one correlated exposure unit:
+- Combined notional = sum of both position sizes
+- Both positions cannot simultaneously win — the upside and downside are correlated, not independent
+- Cap the combined notional at $20 (2 positions × $10), not $20 per leg
+- In the concentration table, label these as "HEDGE PAIR — one correlated exposure" and count them as 2/2 (theme FULL)
+- Do NOT add a third position to the same election even if the third market looks mispriced — you are already fully exposed to the outcome
+
 STEP 1: Call get_positions. Then output a THEME CONCENTRATION TABLE — list every theme with ≥1 open position, showing count/max (e.g. "2028 US election: 3/2 FULL", "GTA VI: 2/2 FULL", "Hungary 2026 election: 1/2"). Any theme showing FULL must be skipped entirely. You MUST output this table before evaluating any market.
 STEP 2: For each market, check your theme concentration table first. If the market's theme is FULL, skip it immediately. Otherwise apply horizon discipline, closed market discipline, and edge check.
 STEP 3: If price differs from your fair probability by the required edge (accounting for sportsbook haircut if applicable), place a trade ($10 USDC). Call skip_all only if you have zero opinion on all markets.`