Gate adaptation on repeatable evidence

README.md

@@ -15,7 +15,7 @@ https://polymarket-site-eta.vercel.app/personal.html
 
 The site fetches live Polymarket markets, generates agent suggestions, lets you
 run frequent paper cycles, and syncs the shared arena state through Neon or
-Vercel Blob. Engine v39 also installs an offline app shell and caches timestamped
+Vercel Blob. Engine v40 also installs an offline app shell and caches timestamped
 market snapshots. During an outage, cycles continue locally; cached entries are
 allowed for 90 minutes, older snapshots become mark-only, and all cached data
 expires after 24 hours.
@@ -26,7 +26,7 @@ small samples toward neutral, caps sizing changes to 0.68x-1.30x, and reserves
 15% of candidates for deterministic exploration so a stale regime cannot become
 permanent.
 
-Engine v39 treats each binary stake as capable of falling to zero even when the
+Engine v40 treats each binary stake as capable of falling to zero even when the
 18% stop cannot fill. New core positions are capped at 2.5%-4% of equity and
 aggressive positions at 3%-5%, with lower limits for near-term, extreme-price,
 reversal, and fast-moving setups. Oversized positions inherited from older
@@ -55,7 +55,7 @@ evidence. Set `EVAL_MARKETS`, `EVAL_CONCURRENCY`, `EVAL_HORIZONS`, or
 `EVAL_COST_CENTS` to change the audit.
 The first 80-market audit found that reversal signals lost 4.34% on average in
 both chronological partitions, while crypto and longshot samples were also
-negative overall. Engine v39 therefore blocks reversal and sports-trend entries outside the fixed
+negative overall. Engine v40 therefore blocks reversal and sports-trend entries outside the fixed
 15% exploration lane and applies modest sizing penalties to crypto and longshots.
 It does not boost any rule from this audit because no positive rule was robust
 across the chronological split.
@@ -66,16 +66,22 @@ segment and averaged -4.13%. Sports trends were negative in train and test and
 averaged -3.53% at 72 hours. Politics trends were the sole cohort with positive
 row-level returns in all three 72-hour segments, but its market-cluster interval
 still crossed zero; that supports a longer hold test, not a larger entry bet.
-Engine v39 gives Politics trend positions that 72-hour observation window before
+Engine v40 gives Politics trend positions that 72-hour observation window before
 ordinary signal exits. Stops, profit locks, settlement handling, and risk-budget
 reductions remain immediate.
 
-Engine v39 also subtracts a half-cent round-trip cost when grading each live
+Engine v40 also subtracts a half-cent round-trip cost when grading each live
 walk-forward signal. Confidence uses the largest independent matching bucket,
 not the sum of five overlapping feature buckets, and evidence from older engine
 versions is down-weighted. This prevents a handful of duplicated observations
 from authorizing larger positions or hiding a modest negative regime.
 
+Engine v40 adds uncertainty-aware promotion and demotion. A matching setup must
+accumulate at least eight effective observations and agree across at least two
+feature views before repeatable positive evidence can increase size or repeatable
+negative evidence can block a new entry. Mixed evidence stays close to neutral
+instead of being mistaken for an edge.
+
 Run `npm run evaluate:settlements` to evaluate fixed decisions made 1, 3, 7,
 14, 30, and 90 days before known binary settlements. The audit uses one
 observation per resolved market and horizon, includes losing contracts at zero,

index.html

@@ -341,7 +341,7 @@ footer{margin-top:34px;padding-top:22px;border-top:1px solid var(--border);color
   <nav class="topnav">
     <div class="brand">
       <div class="logo">🏆</div>
-      <div><div class="brand-name">Polymarket Arena</div><div class="brand-sub">10 agents · 5 core + 5 aggressive</div><div class="build-badge">Adaptive offline engine · v39</div></div>
+      <div><div class="brand-name">Polymarket Arena</div><div class="brand-sub">10 agents · 5 core + 5 aggressive</div><div class="build-badge">Adaptive offline engine · v40</div></div>
     </div>
     <div class="tabs" id="tabs">
       <button class="tab" data-tab="overview">Overview</button>
@@ -363,7 +363,7 @@ footer{margin-top:34px;padding-top:22px;border-top:1px solid var(--border);color
   <div class="personal-banner" id="personalBanner">
     <b>Personal research mode.</b> This copy is for your own analysis, paper tracking, and manual trade research only. It does not pool money, onboard investors, custody funds, bypass eligibility rules, or place orders without your manual approval.
   </div>
-  <div class="live-build-banner"><b>Build v39 active:</b> walk-forward learning now grades returns after estimated trading cost and counts independent observations once when estimating confidence. Older engines are down-weighted, weak live regimes are contained sooner, and the 197-market historical risk priors remain active. This remains paper trading; profits are not guaranteed.</div>
+  <div class="live-build-banner"><b>Build v40 active:</b> adaptive sizing now requires repeatable net-positive evidence with uncertainty bounds. Repeatable net-negative cohorts are blocked, mixed cohorts stay near neutral while learning, and older engines remain down-weighted. This remains paper trading; profits are not guaranteed.</div>
 
   <!-- ============ OVERVIEW ============ -->
   <section class="tabpanel" data-tab="overview">
@@ -425,7 +425,7 @@ footer{margin-top:34px;padding-top:22px;border-top:1px solid var(--border);color
       <div class="card-h"><h3>Returns — all agents</h3><div><span class="small muted">$10,000 start each</span><div class="rangebar" data-chart-ranges></div></div></div>
       <div class="chart-wrap"><svg id="chart2" viewBox="0 0 960 320" preserveAspectRatio="xMidYMid meet"></svg></div>
       <div class="legend" id="comboLegend2"></div>
-      <div class="small muted" style="margin-top:10px">The initial week is an approximate historical replay using prices available on each day and current liquidity as a proxy. The v39 return starts from live cycles only.</div>
+      <div class="small muted" style="margin-top:10px">The initial week is an approximate historical replay using prices available on each day and current liquidity as a proxy. The v40 return starts from live cycles only.</div>
     </div>
   </section>
 
@@ -745,7 +745,7 @@ footer{margin-top:34px;padding-top:22px;border-top:1px solid var(--border);color
   </section>
 
   <footer>
-    Build adaptive-offline-v39 · Paper trading only · Live prices from Polymarket's public Gamma API · Not financial advice ·
+    Build adaptive-offline-v40 · Paper trading only · Live prices from Polymarket's public Gamma API · Not financial advice ·
     <a class="market-link" href="https://github.com/theodore-song/polymarket-analyst" target="_blank" rel="noopener">Source on GitHub</a>
   </footer>
 </div>
@@ -774,7 +774,7 @@ const POLITICS_TREND_MIN_HOLD_HOURS = 72;
 const EXIT_CONFIRM_HOURS = 6;
 const AGENTS_KEY = "pma_agents_v2";
 const SUG_KEY = "pma_suggestions_v5";
-const SUGGESTION_ENGINE_VERSION = 39;
+const SUGGESTION_ENGINE_VERSION = 40;
 const FOCUS_KEY = "pma_focus_v1";
 const VIEW_KEY = "pma_view_v1";
 const PF_SORT_KEY = "pma_portfolio_sort_v1";
@@ -1568,6 +1568,18 @@ function learningFeatures(trade){
   return [`signal:${trade.signal_type||"unknown"}`,`quality:${trade.quality||"unknown"}`,
     `category:${trade.category||"Other"}`,`side:${trade.side||"unknown"}`,`price:${entryBand(trade.entry_price)}`];
 }
+function summarizeLearningBucket(bucket,shrinkage){
+  const weight=Math.max(0,Number(bucket&&bucket.weight||0));
+  const raw=weight?Number(bucket.sum||0)/weight:0;
+  const variance=weight?Math.max(0,Number(bucket.sumSq||0)/weight-raw*raw):0;
+  const priorWeight=4,priorVariance=0.04;
+  const pooledVariance=(priorWeight*priorVariance+weight*variance)/(priorWeight+weight||1);
+  const stderr=Math.sqrt(pooledVariance/Math.max(1,weight));
+  const score=Number(bucket&&bucket.sum||0)/(weight+shrinkage);
+  return {samples:Number(bucket&&bucket.count||0),weight:+weight.toFixed(2),raw:+raw.toFixed(4),
+    score:+score.toFixed(4),stderr:+stderr.toFixed(4),lower_bound:+(score-1.28*stderr).toFixed(4),
+    upper_bound:+(score+1.28*stderr).toFixed(4),win_rate:weight?Number(bucket.wins||0)/weight:0};
+}
 function defaultSignalLedger(){return {pending:[],outcomes:[]};}
 function signalPrice(m,side){return side==="YES"?Number(m&&m.yes_price):Number(m&&m.no_price);}
 function updateSignalLedger(st,markets,suggestions){
@@ -1603,24 +1615,31 @@ function buildSignalCalibration(ledger){
     const versionWeight=version===SUGGESTION_ENGINE_VERSION?1:(version===SUGGESTION_ENGINE_VERSION-1?0.55:0.25);
     const weight=Math.exp(-age/(30*86400000))*versionWeight;
     const ret=clamp(Number(outcome.return||0),-1,2);if(!Number.isFinite(ret))continue;
-    learningFeatures(outcome).forEach(key=>{const b=buckets[key]||(buckets[key]={weight:0,sum:0,wins:0,count:0});
+    learningFeatures(outcome).forEach(key=>{const b=buckets[key]||(buckets[key]={weight:0,sum:0,sumSq:0,wins:0,count:0});
-      b.weight+=weight;b.sum+=ret*weight;b.wins+=(ret>0?weight:0);b.count++;});
+      b.weight+=weight;b.sum+=ret*weight;b.sumSq+=ret*ret*weight;b.wins+=(ret>0?weight:0);b.count++;});
   }
-  const learned=Object.fromEntries(Object.entries(buckets).map(([key,b])=>[key,{samples:b.count,weight:+b.weight.toFixed(2),
+  const learned=Object.fromEntries(Object.entries(buckets).map(([key,b])=>[key,summarizeLearningBucket(b,12)]));
-    score:+(b.sum/(b.weight+12)).toFixed(4),win_rate:b.weight?b.wins/b.weight:0}]));
+  const learnedRows=Object.values(learned);
   return {version:SUGGESTION_ENGINE_VERSION,samples:((ledger&&ledger.outcomes)||[]).length,
     current_samples:((ledger&&ledger.outcomes)||[]).filter(x=>Number(x.strategy_version||0)===SUGGESTION_ENGINE_VERSION).length,buckets:learned,
+    promoted_buckets:learnedRows.filter(r=>r.weight>=8&&r.lower_bound>0.003).length,
+    demoted_buckets:learnedRows.filter(r=>r.weight>=8&&r.upper_bound<-0.003).length,
     pending:((ledger&&ledger.pending)||[]).length};
 }
 function calibratedOpportunity(s,calibration){
   const rows=learningFeatures(s).map(k=>calibration&&calibration.buckets&&calibration.buckets[k]).filter(Boolean);
   const independentWeight=rows.length?Math.max(...rows.map(r=>Number(r.weight||0))):0;
   const score=rows.length?rows.reduce((sum,r)=>sum+r.score,0)/rows.length:0,confidence=independentWeight/(independentWeight+20);
-  const samples=Number(calibration&&calibration.samples||0);
+  const positiveRows=rows.filter(r=>Number(r.weight||0)>=8&&Number(r.lower_bound||0)>0.003);
-  const blocked=(samples>=12&&confidence>=0.20&&score<-0.012)||(samples>=30&&confidence>=0.30&&score<-0.006);
+  const negativeRows=rows.filter(r=>Number(r.weight||0)>=8&&Number(r.upper_bound||0)<-0.003);
-  const trustedScore=score*clamp(confidence*2.5,0.25,1);
+  const promoted=positiveRows.length>=2&&negativeRows.length===0;
-  return {score:+score.toFixed(4),confidence:+confidence.toFixed(3),
+  const demoted=negativeRows.length>=2&&positiveRows.length===0;
-    multiplier:+clamp(1+trustedScore*1.8,0.82,1.15).toFixed(3),allowed:!blocked};
+  const trustedScore=promoted?positiveRows.reduce((sum,r)=>sum+Number(r.lower_bound||0),0)/positiveRows.length
+    :(demoted?negativeRows.reduce((sum,r)=>sum+Number(r.upper_bound||0),0)/negativeRows.length:score*0.10);
+  const state=promoted?"promoted":(demoted?"demoted":"observing");
+  return {score:+score.toFixed(4),confidence:+confidence.toFixed(3),state,promoted,demoted,
+    supporting_features:promoted?positiveRows.length:negativeRows.length,
+    multiplier:+clamp(1+trustedScore*1.8,0.82,1.15).toFixed(3),allowed:!demoted};
 }
 function tradeReturnForLearning(trade,closed){
   const basis=Math.max(1,Number(trade.original_cost||trade.cost||0));
@@ -1640,15 +1659,13 @@ function buildAdaptiveProfile(p){
     observations.push({ret,weight,closed,version});
     if(closed&&version===SUGGESTION_ENGINE_VERSION){currentClosedObservations.push({ret,weight});currentClosedReturns.push(ret);}
     learningFeatures(trade).forEach(key=>{
-      const b=buckets[key]||(buckets[key]={weight:0,sum:0,wins:0,count:0});
+      const b=buckets[key]||(buckets[key]={weight:0,sum:0,sumSq:0,wins:0,count:0});
-      b.weight+=weight;b.sum+=ret*weight;b.wins+=(ret>0?weight:0);b.count++;
+      b.weight+=weight;b.sum+=ret*weight;b.sumSq+=ret*ret*weight;b.wins+=(ret>0?weight:0);b.count++;
     });
   };
   (p.closed||[]).forEach(t=>add(t,true));
   (p.positions||[]).filter(t=>daysHeld(t)>=1).forEach(t=>add(t,false));
-  const summarize=b=>({samples:b.count,weight:+b.weight.toFixed(2),raw:b.weight?b.sum/b.weight:0,
+  const learned=Object.fromEntries(Object.entries(buckets).map(([k,b])=>[k,summarizeLearningBucket(b,6)]));
-    score:+(b.sum/(b.weight+6)).toFixed(4),win_rate:b.weight?b.wins/b.weight:0});
-  const learned=Object.fromEntries(Object.entries(buckets).map(([k,b])=>[k,summarize(b)]));
   const totalWeight=observations.reduce((s,x)=>s+x.weight,0),totalSum=observations.reduce((s,x)=>s+x.ret*x.weight,0);
   const globalScore=totalSum/(totalWeight+10);
   const currentClosedWeight=currentClosedObservations.reduce((s,x)=>s+x.weight,0);
@@ -1689,12 +1706,18 @@ function learnedOpportunity(cfg,p,s,profile=null,calibration=null){
   const confidence=independentWeight/(independentWeight+12),exploration=stableExploration(cfg.id,s.market_id);
   const market=calibratedOpportunity(s,calibration||{samples:0,buckets:{}});
   const historical=historicalOpportunityPrior(s);
-  const personalBlocked=(model.samples>=6&&confidence>=0.20&&score<-0.025)||(model.samples>=12&&confidence>=0.30&&score<-0.015);
+  const positiveRows=rows.filter(r=>Number(r.weight||0)>=6&&Number(r.lower_bound||0)>0.005);
+  const negativeRows=rows.filter(r=>Number(r.weight||0)>=6&&Number(r.upper_bound||0)<-0.01);
+  const personalPromoted=positiveRows.length>=2&&negativeRows.length===0;
+  const personalBlocked=negativeRows.length>=2&&positiveRows.length===0;
+  const personalEvidence=personalPromoted?positiveRows.reduce((sum,r)=>sum+Number(r.lower_bound||0),0)/positiveRows.length
+    :(personalBlocked?negativeRows.reduce((sum,r)=>sum+Number(r.upper_bound||0),0)/negativeRows.length:score*0.10);
   const blockedBy=historical.blocked?"historical":(personalBlocked?"personal":(!market.allowed?"walk-forward":null));
   const blocked=Boolean(blockedBy)&&!exploration;
   return {score:+score.toFixed(4),confidence:+confidence.toFixed(3),market_score:market.score,market_confidence:market.confidence,
+    personal_state:personalPromoted?"promoted":(personalBlocked?"demoted":"observing"),market_state:market.state,
     historical_score:historical.score,historical_confidence:historical.confidence,historical_features:historical.features,
-    multiplier:+clamp((1+score*2.4)*market.multiplier*historical.multiplier,0.65,1.30).toFixed(3),
+    multiplier:+clamp((1+personalEvidence*2.4)*market.multiplier*historical.multiplier,0.65,1.30).toFixed(3),
     exploration,allowed:!blocked,blocked_by:blocked?blockedBy:null,features};
 }
 function emotionalState(ret,trail,trend,rank){
@@ -2129,6 +2152,7 @@ function openPositions(p,cfg,rankedSugs,focus,decision,avoidMarketIds,peerStats=
     const peerAdjusted=peerAdjustedSuggestion(s,peerStats),learning=learnedOpportunity(cfg,p,peerAdjusted,learningProfile,d.marketLearning);
     return Object.assign({},peerAdjusted,{learning_score:learning.score,learning_confidence:learning.confidence,
       market_learning_score:learning.market_score,market_learning_confidence:learning.market_confidence,
+      learning_state:learning.personal_state,market_learning_state:learning.market_state,
       historical_prior_score:learning.historical_score,historical_prior_confidence:learning.historical_confidence,
       historical_prior_features:learning.historical_features,learning_block_reason:learning.blocked_by,
       learning_multiplier:learning.multiplier,learning_exploration:learning.exploration,learning_allowed:learning.allowed});
@@ -2186,12 +2210,13 @@ function openPositions(p,cfg,rankedSugs,focus,decision,avoidMarketIds,peerStats=
       strategy_version:SUGGESTION_ENGINE_VERSION,
       momentum_strength:s.momentum_strength,signal_strength:s.signal_strength,signal_confidence:s.signal_confidence,signal_type:s.signal_type,price_change_1d:s.price_change_1d,price_change_1w:s.price_change_1w,
       learning_score:s.learning_score,learning_confidence:s.learning_confidence,market_learning_score:s.market_learning_score,market_learning_confidence:s.market_learning_confidence,
+      learning_state:s.learning_state,market_learning_state:s.market_learning_state,
       historical_prior_score:s.historical_prior_score,historical_prior_confidence:s.historical_prior_confidence,historical_prior_features:s.historical_prior_features,
       learning_multiplier:s.learning_multiplier,learning_exploration:s.learning_exploration,
       risk_budget_pct:+(riskBudgetPct*100).toFixed(2),
       peak_price:+entry.toFixed(4),gain_stops:{},stop_losses:{}});
     p.history.push({date:logDay(),action:"OPEN",question:s.question,side:s.side,
-      detail:`${decision?decision.mode+" mode — ":""}Bought ${shares} ${s.side} '${s.question.slice(0,40)}' @ ${pct(entry)} for ${fmtUSD(cost)} · max binary loss budget ${(riskBudgetPct*100).toFixed(1)}% · signal margin ${((Math.abs(s.net_edge!=null?s.net_edge:s.edge))*100).toFixed(1)}c · learned weight ${Number(s.learning_multiplier||1).toFixed(2)}x${(s.historical_prior_features||[]).length?` · history prior ${(s.historical_prior_features||[]).join("+")}`:""} · evidence ${Math.round((s.evidence_score||0)*100)}${s.peer_note?` (${s.peer_note})`:""}`});
+      detail:`${decision?decision.mode+" mode — ":""}Bought ${shares} ${s.side} '${s.question.slice(0,40)}' @ ${pct(entry)} for ${fmtUSD(cost)} · max binary loss budget ${(riskBudgetPct*100).toFixed(1)}% · signal margin ${((Math.abs(s.net_edge!=null?s.net_edge:s.edge))*100).toFixed(1)}c · learned weight ${Number(s.learning_multiplier||1).toFixed(2)}x (${s.learning_state||"observing"}/${s.market_learning_state||"observing"})${(s.historical_prior_features||[]).length?` · history prior ${(s.historical_prior_features||[]).join("+")}`:""} · evidence ${Math.round((s.evidence_score||0)*100)}${s.peer_note?` (${s.peer_note})`:""}`});
     opened++;openedIds.push(String(s.market_id));
   }
   const eqAfter=equity(p);
@@ -2560,7 +2585,7 @@ function decisionSummary(p){
   const blockerRows=Object.entries(d.rejectionCounts||{}).filter(([,count])=>count>0).sort((a,b)=>b[1]-a[1]);
   const blockers=blockerRows.length?` Blocks: ${blockerRows.slice(0,4).map(([key,count])=>`${blockerLabels[key]||key} ${count}`).join(", ")}.`:"";
   const learning=d.learning?` Learning: ${d.learning.samples} completed trades retained with older engines down-weighted, ${(d.learning.global_score*100).toFixed(2)}% shrunk expectancy; ${d.learning.current_samples||0} completed under v${SUGGESTION_ENGINE_VERSION}${d.learning.best?`; strongest ${d.learning.best.feature.replace(":"," ")}`:""}${d.learning.worst?`; weakest ${d.learning.worst.feature.replace(":"," ")}`:""}.`:"";
-  const calibration=d.marketLearning?` Walk-forward calibration: ${d.marketLearning.samples||0} net-of-cost graded signals (${d.marketLearning.current_samples||0} under v${SUGGESTION_ENGINE_VERSION}), ${d.marketLearning.pending||0} awaiting a future price. Confidence counts independent outcomes once. Historical prior: reversal and sports-trend entries require the fixed 15% exploration lane; crypto, longshots, and YES entries are sized down, not forbidden. Politics trends receive 72 hours before ordinary signal exits.`:"";
+  const calibration=d.marketLearning?` Walk-forward calibration: ${d.marketLearning.samples||0} net-of-cost graded signals (${d.marketLearning.current_samples||0} under v${SUGGESTION_ENGINE_VERSION}), ${d.marketLearning.pending||0} awaiting a future price; ${d.marketLearning.promoted_buckets||0} feature cohorts promoted and ${d.marketLearning.demoted_buckets||0} demoted. Confidence counts independent outcomes once and uncertainty gates sizing. Historical prior: reversal and sports-trend entries require the fixed 15% exploration lane; crypto, longshots, and YES entries are sized down, not forbidden. Politics trends receive 72 hours before ordinary signal exits.`:"";
   return `${d.mode} mode: ${d.reason}${emotion} Limits now: ${d.maxNew} new trade${d.maxNew===1?"":"s"}, max ${(d.maxFrac*100).toFixed(1)}% per position${d.minConv?`, conviction ${d.minConv}+`:""}.${learning}${calibration}${exposure}${allocation}${candidates}${blockers}`;
 }
 function renderAgentBrief(cfg,p,st){
@@ -4131,7 +4156,7 @@ window.PMA_ENGINE_DIAGNOSTICS=Object.freeze({
     coreTradeLossPct:MAX_CORE_TRADE_LOSS_PCT*100,aggressiveTradeLossPct:MAX_AGGRESSIVE_TRADE_LOSS_PCT*100,
     coreGapTradeLossPct:MAX_CORE_GAP_TRADE_LOSS_PCT*100,aggressiveGapTradeLossPct:MAX_AGGRESSIVE_GAP_TRADE_LOSS_PCT*100,
     offlineEntryMaxAgeMinutes:OFFLINE_ENTRY_MAX_AGE_MS/60000,offlineCacheMaxAgeHours:OFFLINE_CACHE_MAX_AGE_MS/3600000,explorationPct:15,
-    signalRoundTripCostCents:SIGNAL_ROUND_TRIP_COST*100,independentConfidence:true,
+    signalRoundTripCostCents:SIGNAL_ROUND_TRIP_COST*100,independentConfidence:true,uncertaintyGatedCalibration:true,
     historicalPrior:"reversal and sports trends blocked outside exploration; crypto, longshots, and YES sized down"}),
 });
 function runEngineSelfTest(){
@@ -4177,6 +4202,13 @@ function runEngineSelfTest(){
   const ledgerState={signal_ledger:{pending:[{key:"ledger-test",market_id:"ledger-test",observed_at:hoursAgo(13),side:"YES",entry_price:0.40,
     signal_type:"trend",quality:"confirmed",category:"Politics"}],outcomes:[]}};
   updateSignalLedger(ledgerState,[market({id:"ledger-test",yes_price:0.50,no_price:0.50})],[]);
+  const calibrationCandidate={signal_type:"trend",quality:"confirmed",category:"Politics",side:"YES",entry_price:0.42};
+  const calibrationFromReturns=returns=>buildSignalCalibration({pending:[],outcomes:returns.map(ret=>Object.assign({},calibrationCandidate,
+    {strategy_version:SUGGESTION_ENGINE_VERSION,return:ret,evaluated_at:closedAt}))});
+  const singleCalibration=calibratedOpportunity(calibrationCandidate,calibrationFromReturns([0.10]));
+  const stablePositiveCalibration=calibratedOpportunity(calibrationCandidate,calibrationFromReturns(Array(24).fill(0.12)));
+  const stableNegativeCalibration=calibratedOpportunity(calibrationCandidate,calibrationFromReturns(Array(24).fill(-0.12)));
+  const noisyCalibration=calibratedOpportunity(calibrationCandidate,calibrationFromReturns(Array.from({length:24},(_,i)=>i%2?0.12:-0.12)));
   const lossLearner=defaultPortfolio();
   for(let i=0;i<6;i++)lossLearner.closed.push({strategy_version:SUGGESTION_ENGINE_VERSION,signal_type:"trend",quality:"confirmed",category:"Politics",side:"YES",entry_price:0.42,original_cost:100,
     realized_pnl:-50,opened_at:hoursAgo(72+i),closed_at:closedAt});
@@ -4225,7 +4257,11 @@ function runEngineSelfTest(){
       historicalPriorAllowsPoliticsTrend:!priorPoliticsTrend.blocked,
       ledgerMaturesWithoutLookahead:ledgerState.signal_ledger.pending.length===0&&ledgerState.signal_ledger.outcomes.length===1&&ledgerState.signal_ledger.outcomes[0].return===0.2375,
       ledgerIsNetOfCosts:ledgerState.signal_ledger.outcomes[0].gross_return===0.25&&ledgerState.signal_ledger.outcomes[0].estimated_cost_return===0.0125,
-      independentCalibrationConfidence:calibratedOpportunity({signal_type:"trend",quality:"confirmed",category:"Politics",side:"YES",entry_price:0.42},buildSignalCalibration({pending:[],outcomes:[{strategy_version:SUGGESTION_ENGINE_VERSION,signal_type:"trend",quality:"confirmed",category:"Politics",side:"YES",entry_price:0.42,return:0.10,evaluated_at:closedAt}]})).confidence<0.06,
+      independentCalibrationConfidence:singleCalibration.confidence<0.06,
+      singleOutcomeStaysObserving:singleCalibration.state==="observing"&&singleCalibration.multiplier<=1.01,
+      stablePositiveIsPromoted:stablePositiveCalibration.state==="promoted"&&stablePositiveCalibration.multiplier>1.02&&stablePositiveCalibration.allowed,
+      stableNegativeIsBlocked:stableNegativeCalibration.state==="demoted"&&!stableNegativeCalibration.allowed,
+      noisyEvidenceStaysNeutral:noisyCalibration.state==="observing"&&noisyCalibration.multiplier>=0.99&&noisyCalibration.multiplier<=1.01,
       blocksConfirmedLosingRegime:!lossOpportunity.allowed,containmentMode:lossDecision.mode,containmentMaxNew:lossDecision.maxNew,
       legacyLossDoesNotFreezeCurrentEngine:legacyLossDecision.mode!=="Loss Regime Containment"},
     riskBudget:{core:coreRiskBudget,aggressiveGap:aggressiveGapBudget,boundedStake:boundedStakeForRisk(10000,1000,riskCfg,{entry_price:0.82,days_to_resolution:5,price_change_1d:0.06}),

sw.js

@@ -1,4 +1,4 @@
-const CACHE_NAME = "polymarket-arena-v39";
+const CACHE_NAME = "polymarket-arena-v40";
 const APP_SHELL = ["/", "/index.html", "/personal.html", "/cycle-worker.js"];
 
 self.addEventListener("install", event => {