Replace decorative overload policy with real serving pipeline and dedicated Serving page

The old overload policy had dead controls (maxQueueDepth, rateLimitPerCustomer never read)
and trivial flat penalties. This replaces it with a full serving pipeline where deployed
models form a fleet, requests route through priority/degradation logic, and policy choices
create meaningful strategic tradeoffs.

New serving pipeline: fleet building from deployed models (size/quant/MoE multipliers),
demand categorization by 5 priority tiers, enterprise capacity reservation, priority-ordered
serving with overflow behaviors (queue/reject/degrade), auto-degradation to faster models
under load, and Batch API to fill idle capacity at discounted rates.

4 new research nodes gate features progressively: Intelligent Request Routing, Priority
Queue System, Request Batching, and Auto-Scaling. New dedicated Serving page with pipeline
metrics, model fleet utilization, and research-gated policy controls.

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>

packages/game-engine/src/data/techTree.ts

@@ -433,6 +433,48 @@ export const TECH_TREE: ResearchNode[] = [
     effects: [{ type: 'unlock_product_line', target: 'agents-platform', value: 1 }],
   },
 
+  // === SERVING INFRASTRUCTURE ===
+  {
+    id: 'request-routing',
+    name: 'Intelligent Request Routing',
+    description: 'Route requests to optimal model size/variant. Unlocks routing strategy and per-tier rate limits.',
+    era: 'scaleup',
+    category: 'efficiency',
+    prerequisites: ['inference-optimization'],
+    cost: { researchPoints: 2, compute: 25, ticks: 150 },
+    effects: [{ type: 'unlock_feature', target: 'request-routing', value: 1 }],
+  },
+  {
+    id: 'priority-queues',
+    name: 'Priority Queue System',
+    description: 'SLA-aware scheduling with granular priority controls. Unlocks priority ordering and overflow policies.',
+    era: 'scaleup',
+    category: 'efficiency',
+    prerequisites: ['request-routing'],
+    cost: { researchPoints: 3, compute: 30, ticks: 180 },
+    effects: [{ type: 'unlock_feature', target: 'priority-queues', value: 1 }],
+  },
+  {
+    id: 'request-batching',
+    name: 'Request Batching',
+    description: 'Group inference requests for higher throughput. Unlocks Batch API product line at 50% discount.',
+    era: 'scaleup',
+    category: 'efficiency',
+    prerequisites: ['inference-optimization'],
+    cost: { researchPoints: 2, compute: 20, ticks: 120 },
+    effects: [{ type: 'unlock_feature', target: 'request-batching', value: 1 }],
+  },
+  {
+    id: 'auto-scaling',
+    name: 'Auto-Scaling Infrastructure',
+    description: 'Dynamically reallocate compute during demand spikes. +20% effective capacity headroom.',
+    era: 'bigtech',
+    category: 'efficiency',
+    prerequisites: ['request-routing'],
+    cost: { researchPoints: 4, compute: 60, ticks: 300 },
+    effects: [{ type: 'efficiency_boost', target: 'auto_scaling', value: 0.2 }],
+  },
+
   // === DATA ===
   {
     id: 'data-pipeline',

packages/game-engine/src/systems/market/apiTierSystem.ts

@@ -1,9 +1,10 @@
-import type { ApiTierState, ApiTierId, DeveloperEcosystem } from '@ai-tycoon/shared';
+import type { ApiTierState, ApiTierId, DeveloperEcosystem, TierServingMetrics } from '@ai-tycoon/shared';
 import {
   API_TIER_ORDER,
   API_CONVERSION_RATES,
   API_TIER_CHURN_RATES,
   API_TOKENS_PER_DEVELOPER_PER_TICK,
+  REJECTION_CHURN_MULTIPLIER,
 } from '@ai-tycoon/shared';
 
 export interface ApiTickResult {
@@ -18,6 +19,8 @@ export function processApiTiers(
   modelQuality: number,
   seasonalApiMultiplier: number,
   ecosystem: DeveloperEcosystem,
+  apiPaidMetrics: TierServingMetrics,
+  apiFreeMetrics: TierServingMetrics,
 ): ApiTickResult {
   const updated: ApiTierState = {
     tiers: { ...tiers.tiers },
@@ -89,6 +92,23 @@ export function processApiTiers(
   updated.totalDevelopers = totalDevelopers;
   updated.totalTokensPerTick = totalTokens;
 
+  const freeRejectRate = apiFreeMetrics.demandTokens > 0
+    ? apiFreeMetrics.rejectedTokens / apiFreeMetrics.demandTokens : 0;
+  if (freeRejectRate > 0) {
+    const extraChurn = updated.tiers.free.developerCount * freeRejectRate * 0.01 * REJECTION_CHURN_MULTIPLIER;
+    updated.tiers.free.developerCount = Math.max(0, updated.tiers.free.developerCount - extraChurn);
+  }
+
+  const paidRejectRate = apiPaidMetrics.demandTokens > 0
+    ? apiPaidMetrics.rejectedTokens / apiPaidMetrics.demandTokens : 0;
+  if (paidRejectRate > 0) {
+    for (const id of API_TIER_ORDER) {
+      if (id === 'free') continue;
+      const extraChurn = updated.tiers[id].developerCount * paidRejectRate * 0.005 * REJECTION_CHURN_MULTIPLIER;
+      updated.tiers[id].developerCount = Math.max(0, updated.tiers[id].developerCount - extraChurn);
+    }
+  }
+
   return {
     apiTiers: updated,
     apiRevenue: Math.max(0, apiRevenue),

packages/game-engine/src/systems/market/consumerTierSystem.ts

@@ -1,12 +1,13 @@
-import type { ConsumerTierState, ConsumerTierId } from '@ai-tycoon/shared';
+import type { ConsumerTierState, ConsumerTierId, TierServingMetrics } from '@ai-tycoon/shared';
 import {
   CONSUMER_TIER_ORDER,
   CONVERSION_RATES,
   TIER_CHURN_RATES,
   FREE_TIER_ADOPTION_RATE,
   CONSUMER_TOKENS_PER_SUBSCRIBER,
-  OVERLOAD_PENALTY_EXPONENT,
   NETWORK_DEGRADATION,
+  REJECTION_CHURN_MULTIPLIER,
+  QUEUE_CHURN_MULTIPLIER,
 } from '@ai-tycoon/shared';
 
 export interface ConsumerTickResult {
@@ -20,9 +21,9 @@ export function processConsumerTiers(
   playerConsumerCustomers: number,
   modelQuality: number,
   seasonalConsumerMultiplier: number,
-  demandCapacityRatio: number,
   networkLatencyPenalty: number,
-  overloadPolicy: { degradeQualityUnderLoad: boolean; prioritizeEnterprise: boolean },
+  consumerPaidMetrics: TierServingMetrics,
+  consumerFreeMetrics: TierServingMetrics,
 ): ConsumerTickResult {
   const updated = {
     tiers: { ...tiers.tiers },
@@ -97,26 +98,64 @@ export function processConsumerTiers(
 
   updated.totalUsers = totalUsers;
 
+  const paidDemand = consumerPaidMetrics.demandTokens;
+  const freeDemand = consumerFreeMetrics.demandTokens;
+  const totalDemand = paidDemand + freeDemand;
+
+  let servingPenalty = 0;
+  if (totalDemand > 0) {
+    const totalRejected = consumerPaidMetrics.rejectedTokens + consumerFreeMetrics.rejectedTokens;
+    const totalQueued = consumerPaidMetrics.queuedTokens + consumerFreeMetrics.queuedTokens;
+    const rejectedFraction = totalRejected / totalDemand;
+    const queuedFraction = totalQueued / totalDemand;
+
+    servingPenalty = rejectedFraction * 1.5 + queuedFraction * 0.5;
+
+    const avgQuality = totalDemand > 0
+      ? (consumerPaidMetrics.avgQualityDelivered * paidDemand + consumerFreeMetrics.avgQualityDelivered * freeDemand) / totalDemand
+      : modelQuality;
+    const qualityGap = Math.max(0, modelQuality - avgQuality);
+    servingPenalty += qualityGap * 0.8;
+
+    if (consumerFreeMetrics.rejectedTokens > 0 && freeDemand > 0) {
+      const freeRejectRate = consumerFreeMetrics.rejectedTokens / freeDemand;
+      const extraChurn = updated.tiers.free.userCount * freeRejectRate * 0.01 * REJECTION_CHURN_MULTIPLIER;
+      updated.tiers.free.userCount = Math.max(0, updated.tiers.free.userCount - extraChurn);
+    }
+
+    if (consumerPaidMetrics.rejectedTokens > 0 && paidDemand > 0) {
+      const paidRejectRate = consumerPaidMetrics.rejectedTokens / paidDemand;
+      for (const id of CONSUMER_TIER_ORDER) {
+        if (id === 'free') continue;
+        const extraChurn = updated.tiers[id].userCount * paidRejectRate * 0.005 * REJECTION_CHURN_MULTIPLIER;
+        updated.tiers[id].userCount = Math.max(0, updated.tiers[id].userCount - extraChurn);
+      }
+    }
+
+    if (totalQueued > 0) {
+      for (const id of CONSUMER_TIER_ORDER) {
+        const extraChurn = updated.tiers[id].userCount * queuedFraction * 0.002 * QUEUE_CHURN_MULTIPLIER;
+        updated.tiers[id].userCount = Math.max(0, updated.tiers[id].userCount - extraChurn);
+      }
+    }
+  }
+
   let headroomBonus = 0;
-  let overloadPenalty = 0;
-  if (demandCapacityRatio <= 1) {
-    headroomBonus = (1 - demandCapacityRatio) * 0.2;
+  if (totalDemand > 0) {
+    const totalServed = consumerPaidMetrics.servedTokens + consumerFreeMetrics.servedTokens;
+    const servedFraction = totalServed / totalDemand;
+    if (servedFraction > 0.95) {
+      headroomBonus = (servedFraction - 0.95) * 4;
+    }
   } else {
-    overloadPenalty = Math.min(1, Math.pow(demandCapacityRatio - 1, OVERLOAD_PENALTY_EXPONENT));
+    headroomBonus = 0.1;
   }
 
   const netLatencyPenalty = networkLatencyPenalty * NETWORK_DEGRADATION.satisfactionPenaltyPerLatency;
   updated.satisfaction = Math.min(1, Math.max(0,
-    0.3 + modelQuality * 0.5 + headroomBonus - overloadPenalty - netLatencyPenalty,
+    0.3 + modelQuality * 0.5 + headroomBonus - servingPenalty - netLatencyPenalty,
   ));
 
-  if (overloadPolicy.degradeQualityUnderLoad && demandCapacityRatio > 0.85) {
-    updated.satisfaction = Math.max(0, updated.satisfaction - 0.02);
-  }
-  if (overloadPolicy.prioritizeEnterprise && demandCapacityRatio > 0.9) {
-    updated.satisfaction = Math.max(0, updated.satisfaction - 0.01);
-  }
-
   updated.viralCoefficient = modelQuality > 0.5 ? 1 + (modelQuality - 0.5) * 2 : 0;
 
   return {

packages/game-engine/src/systems/market/enterprisePipeline.ts

@@ -5,6 +5,7 @@ import type {
   EnterpriseSegment,
   EnterprisePipelineStage,
   DeveloperEcosystem,
+  TierServingMetrics,
 } from '@ai-tycoon/shared';
 import {
   BASE_LEAD_RATE,
@@ -17,6 +18,7 @@ import {
   ENTERPRISE_SLA_REQUIREMENTS,
   ENTERPRISE_CAPABILITY_REQUIREMENTS,
   ENTERPRISE_TOKENS_PER_TICK,
+  ENTERPRISE_REJECTION_SLA_MULTIPLIER,
 } from '@ai-tycoon/shared';
 import { ENTERPRISE_NAMES } from '../../data/enterpriseNames';
 
@@ -62,7 +64,7 @@ export function processEnterprisePipeline(
   devEcosystem: DeveloperEcosystem,
   seasonalEntMultiplier: number,
   currentTick: number,
-  demandCapacityRatio: number,
+  enterpriseServingMetrics: TierServingMetrics,
 ): EnterprisePipelineResult {
   const pipeline = [...ent.pipeline];
   const activeContracts = [...ent.activeContracts];
@@ -129,7 +131,10 @@ export function processEnterprisePipeline(
     if (lead.stage === 'qualification') {
       transitionProb *= modelCapability >= lead.requiredCapability ? 1 : 0.1;
     } else if (lead.stage === 'poc') {
-      transitionProb *= Math.max(0.2, 1 - Math.max(0, demandCapacityRatio - 0.9) * 5);
+      const entDemand = enterpriseServingMetrics.demandTokens;
+      const entRejected = enterpriseServingMetrics.rejectedTokens;
+      const rejectRate = entDemand > 0 ? entRejected / entDemand : 0;
+      transitionProb *= Math.max(0.2, 1 - rejectRate * 5);
     } else if (lead.stage === 'negotiation') {
       transitionProb *= Math.max(0.3, 1 - (lead.dealValue / 10_000_000) * 0.5);
     }
@@ -181,14 +186,22 @@ export function processEnterprisePipeline(
     const updated = { ...contract };
     updated.totalTicks++;
 
-    if (demandCapacityRatio <= (1 / updated.slaUptime)) {
+    const entDemand = enterpriseServingMetrics.demandTokens;
+    const entServed = enterpriseServingMetrics.servedTokens;
+    const entRejected = enterpriseServingMetrics.rejectedTokens;
+    const servedFraction = entDemand > 0 ? entServed / entDemand : 1;
+    const wasRejected = entRejected > 0;
+    const qualityMet = enterpriseServingMetrics.avgQualityDelivered >= 0.85;
+
+    if (servedFraction >= updated.slaUptime && qualityMet && !wasRejected) {
       updated.uptimeTicks++;
     } else {
       updated.slaViolations++;
-      const penalty = updated.pricePerMToken * (updated.tokensPerTick / 1_000_000) * SLA_PENALTY_FRACTION;
+      const severityMultiplier = wasRejected ? ENTERPRISE_REJECTION_SLA_MULTIPLIER : 1.0;
+      const penalty = updated.pricePerMToken * (updated.tokensPerTick / 1_000_000) * SLA_PENALTY_FRACTION * severityMultiplier;
       slaPenalties += penalty;
       updated.slaPenaltiesPaid += penalty;
-      updated.satisfaction = Math.max(0, updated.satisfaction - 0.005);
+      updated.satisfaction = Math.max(0, updated.satisfaction - (wasRejected ? 0.01 : 0.005));
     }
 
     if (updated.totalTicks > 0 && updated.slaViolations === 0) {

packages/game-engine/src/systems/market/index.ts

@@ -1,5 +1,6 @@
-import type { GameState, MarketState, BenchmarkResult, Competitor } from '@ai-tycoon/shared';
-import { CONSUMER_TOKENS_PER_SUBSCRIBER } from '@ai-tycoon/shared';
+import type { GameState, MarketState, BenchmarkResult } from '@ai-tycoon/shared';
+import { CONSUMER_TOKENS_PER_SUBSCRIBER, API_TOKENS_PER_DEVELOPER_PER_TICK, BATCH_API_DEMAND_PER_DEV, makeInitialServingMetrics } from '@ai-tycoon/shared';
+import type { TrafficPriority, TierServingMetrics } from '@ai-tycoon/shared';
 import { BENCHMARKS } from '../../data/benchmarks';
 import { computeSeasonal } from './seasonalSystem';
 import { updateObsolescence } from './obsolescenceSystem';
@@ -9,6 +10,9 @@ import { processApiTiers } from './apiTierSystem';
 import { processProductLines } from './productLines';
 import { processDeveloperEcosystem } from './developerEcosystem';
 import { processEnterprisePipeline } from './enterprisePipeline';
+import { processServingPipeline } from './servingPipeline';
+import type { DemandByTier } from './servingPipeline';
+import type { ResearchBonuses } from '../researchBonuses';
 
 export interface MarketTickResult {
   marketState: MarketState;
@@ -44,24 +48,26 @@ function getSegmentQuality(
   return weightedSum / totalWeight;
 }
 
-export function processMarketV2(state: GameState, currentTickCapacity: number): MarketTickResult {
+export function processMarketV2(
+  state: GameState,
+  currentTickCapacity: number,
+  effectiveInferenceFlops?: number,
+  researchBonuses?: ResearchBonuses,
+): MarketTickResult {
   const consumerQuality = getSegmentQuality('consumer', state.models.benchmarkResults, state.models.bestDeployedModelScore);
   const enterpriseQuality = getSegmentQuality('enterprise', state.models.benchmarkResults, state.models.bestDeployedModelScore);
   const modelQuality = state.models.benchmarkResults.length > 0
     ? (consumerQuality + enterpriseQuality) / 2
     : state.models.bestDeployedModelScore / 100;
 
-  // --- Seasonal ---
   const seasonal = computeSeasonal(state.meta.tickCount);
 
-  // --- Obsolescence ---
   const obsolescence = updateObsolescence(
     state.market.obsolescence,
     state.meta.currentEra,
     state.meta.tickCount,
   );
 
-  // --- Developer Ecosystem ---
   const freeApiDevs = state.market.apiTiers.tiers.free.developerCount;
   const totalApiDevs = state.market.apiTiers.totalDevelopers;
   const engineeringCount = state.talent.departments.engineering.headcount;
@@ -75,7 +81,6 @@ export function processMarketV2(state: GameState, currentTickCapacity: number):
     state.meta.currentEra,
   );
 
-  // --- TAM & Market Shares ---
   const chatProduct = state.models.productLines.find(p => p.type === 'chat-product');
   const textApi = state.models.productLines.find(p => p.type === 'text-api');
 
@@ -106,32 +111,7 @@ export function processMarketV2(state: GameState, currentTickCapacity: number):
   const playerDevCustomers = tam.segments.developer.shares.find(s => s.playerId === 'player')?.customers ?? 0;
   const playerEntCustomers = tam.segments.enterprise.shares.find(s => s.playerId === 'player')?.customers ?? 0;
 
-  // --- Consumer Tiers ---
-  const consumerDemandEstimate = state.market.consumerTiers.totalUsers * CONSUMER_TOKENS_PER_SUBSCRIBER;
-  const demandCapacityRatio = currentTickCapacity > 0
-    ? consumerDemandEstimate / currentTickCapacity
-    : consumerDemandEstimate > 0 ? 10 : 0;
-
-  const consumerResult = processConsumerTiers(
-    state.market.consumerTiers,
-    playerConsumerCustomers,
-    modelQuality,
-    seasonal.multipliers.consumer,
-    demandCapacityRatio,
-    state.infrastructure.networkLatencyPenalty,
-    state.market.overloadPolicy,
-  );
-
-  // --- API Tiers ---
-  const apiResult = processApiTiers(
-    state.market.apiTiers,
-    playerDevCustomers,
-    modelQuality,
-    seasonal.multipliers.api,
-    devEcosystem,
-  );
-
-  // --- Product Lines ---
+  // --- Product Lines (compute first to get token demand) ---
   const productResult = processProductLines(
     state.market.codeAssistant,
     state.market.agentsPlatform,
@@ -142,22 +122,103 @@ export function processMarketV2(state: GameState, currentTickCapacity: number):
     seasonal.multipliers.enterprise,
   );
 
-  // --- Enterprise Pipeline ---
+  // --- Pre-compute demand estimates by tier for serving pipeline ---
+  const consumerTiers = state.market.consumerTiers;
+  const apiTiers = state.market.apiTiers;
+  const enterprise = state.market.enterprise;
+
+  const consumerPaidTokens = (consumerTiers.tiers.plus.userCount + consumerTiers.tiers.pro.userCount + consumerTiers.tiers.team.userCount) * CONSUMER_TOKENS_PER_SUBSCRIBER;
+  const consumerFreeTokens = consumerTiers.tiers.free.userCount * CONSUMER_TOKENS_PER_SUBSCRIBER;
+
+  const apiPaidTokens =
+    apiTiers.tiers.payg.developerCount * API_TOKENS_PER_DEVELOPER_PER_TICK.payg
+    + apiTiers.tiers.scale.developerCount * API_TOKENS_PER_DEVELOPER_PER_TICK.scale
+    + apiTiers.tiers['enterprise-api'].developerCount * API_TOKENS_PER_DEVELOPER_PER_TICK['enterprise-api']
+    + productResult.codeAssistantTokenDemand;
+  const apiFreeTokens = apiTiers.tiers.free.developerCount * API_TOKENS_PER_DEVELOPER_PER_TICK.free;
+
+  let enterpriseTokens = 0;
+  for (const contract of enterprise.activeContracts) {
+    enterpriseTokens += contract.tokensPerTick;
+  }
+  enterpriseTokens += productResult.agentsPlatformTokenDemand;
+
+  const demandByTier: DemandByTier = {
+    'enterprise': enterpriseTokens,
+    'api-paid': apiPaidTokens,
+    'consumer-paid': consumerPaidTokens,
+    'api-free': apiFreeTokens,
+    'consumer-free': consumerFreeTokens,
+  };
+
+  // --- Batch API demand ---
+  let batchDemand = 0;
+  if (state.market.overloadPolicy.batchApiEnabled) {
+    for (const id of ['free', 'payg', 'scale', 'enterprise-api'] as const) {
+      batchDemand += apiTiers.tiers[id].developerCount * (BATCH_API_DEMAND_PER_DEV[id] ?? 0);
+    }
+    batchDemand *= Math.max(0.1, modelQuality);
+  }
+
+  const completedResearch = state.research?.completedResearch ?? [];
+
+  // --- Serving Pipeline ---
+  const servingResult = processServingPipeline({
+    modelsState: state.models,
+    effectiveInferenceFlops: effectiveInferenceFlops ?? currentTickCapacity,
+    overloadPolicy: state.market.overloadPolicy,
+    demandByTier,
+    batchApi: {
+      ...state.market.batchApi,
+      totalBatchDemand: batchDemand,
+    },
+    modelQuality,
+    researchUnlocks: {
+      servingRoutingUnlocked: completedResearch.includes('request-routing'),
+      priorityQueuesUnlocked: completedResearch.includes('priority-queues'),
+      batchApiUnlocked: completedResearch.includes('request-batching'),
+      autoScalingBonus: completedResearch.includes('auto-scaling') ? 0.2 : 0,
+    },
+  });
+
+  const sm = servingResult.servingMetrics;
+
+  // --- Consumer Tiers (now with serving metrics) ---
+  const consumerResult = processConsumerTiers(
+    state.market.consumerTiers,
+    playerConsumerCustomers,
+    modelQuality,
+    seasonal.multipliers.consumer,
+    state.infrastructure.networkLatencyPenalty,
+    sm.tierMetrics['consumer-paid'],
+    sm.tierMetrics['consumer-free'],
+  );
+
+  // --- API Tiers (now with serving metrics) ---
+  const apiResult = processApiTiers(
+    state.market.apiTiers,
+    playerDevCustomers,
+    modelQuality,
+    seasonal.multipliers.api,
+    devEcosystem,
+    sm.tierMetrics['api-paid'],
+    sm.tierMetrics['api-free'],
+  );
+
+  // --- Enterprise Pipeline (now with serving metrics) ---
   const salesDept = state.talent.departments.sales;
-  const salesHeadcount = salesDept.headcount;
-  const salesEffectiveness = salesDept.effectiveness;
 
   const enterpriseResult = processEnterprisePipeline(
     state.market.enterprise,
     state.reputation.score,
     state.models.bestDeployedModelScore,
     state.models.bestDeployedSafetyScore,
-    salesHeadcount,
-    salesEffectiveness,
+    salesDept.headcount,
+    salesDept.effectiveness,
     devEcosystem,
     seasonal.multipliers.enterprise,
     state.meta.tickCount,
-    demandCapacityRatio,
+    sm.tierMetrics['enterprise'],
   );
 
   // --- Aggregate revenue ---
@@ -165,9 +226,10 @@ export function processMarketV2(state: GameState, currentTickCapacity: number):
     + productResult.codeAssistantRevenue
     + productResult.agentsPlatformRevenue;
 
-  const apiRevenue = apiResult.apiRevenue
+  let apiRevenue = apiResult.apiRevenue
     + enterpriseResult.contractRevenue
-    - enterpriseResult.slaPenalties;
+    - enterpriseResult.slaPenalties
+    + servingResult.batchRevenue;
 
   const totalTokenDemand = consumerResult.totalConsumerTokenDemand
     + apiResult.totalApiTokenDemand
@@ -186,26 +248,7 @@ export function processMarketV2(state: GameState, currentTickCapacity: number):
   const openSourceCount = state.market.openSourcedModels.length;
   if (openSourceCount > 0) {
     const revenueReduction = openSourceCount * 0.10 * 0.3;
-    const adjustedApiRevenue = apiRevenue * (1 - revenueReduction);
-    return {
-      marketState: {
-        ...state.market,
-        tam,
-        consumerTiers: consumerResult.consumerTiers,
-        apiTiers: apiResult.apiTiers,
-        codeAssistant: productResult.codeAssistant,
-        agentsPlatform: productResult.agentsPlatform,
-        enterprise: enterpriseResult.enterprise,
-        developerEcosystem: devEcosystem,
-        seasonalPhase: seasonal.phase,
-        seasonalMultiplier: seasonal.multipliers.consumer,
-        obsolescence,
-        subscriberHistory,
-      },
-      apiRevenue: Math.max(0, adjustedApiRevenue),
-      subscriptionRevenue,
-      totalTokenDemand,
-    };
+    apiRevenue = apiRevenue * (1 - revenueReduction);
   }
 
   return {
@@ -221,6 +264,8 @@ export function processMarketV2(state: GameState, currentTickCapacity: number):
       seasonalPhase: seasonal.phase,
       seasonalMultiplier: seasonal.multipliers.consumer,
       obsolescence,
+      servingMetrics: sm,
+      batchApi: servingResult.batchApi,
       subscriberHistory,
     },
     apiRevenue: Math.max(0, apiRevenue),

packages/game-engine/src/systems/market/servingPipeline.ts

@@ -0,0 +1,462 @@
+import type {
+  OverloadPolicy,
+  TrafficPriority,
+  TierServingMetrics,
+  ServingMetrics,
+  ModelUtilizationEntry,
+  BatchApiState,
+} from '@ai-tycoon/shared';
+import type { BaseModel, ModelVariant, ModelFamily, ModelsState, SizeTier } from '@ai-tycoon/shared';
+import {
+  MODEL_SIZE_THROUGHPUT_SCALER,
+  MOE_SPEED_MULTIPLIER,
+  FLOPS_TO_TOKENS_MULTIPLIER,
+  QUANTIZATION_CONFIGS,
+  REJECTION_SATISFACTION_PENALTY,
+  QUEUE_SATISFACTION_PENALTY,
+  DEGRADATION_SATISFACTION_PENALTY,
+  BASE_LATENCY_MS,
+  QUEUE_LATENCY_MS_PER_PERCENT,
+  BATCH_API_MAX_PENDING,
+} from '@ai-tycoon/shared';
+import { makeInitialServingMetrics } from '@ai-tycoon/shared';
+
+export interface ModelServingSlot {
+  modelId: string;
+  modelName: string;
+  sizeTier: SizeTier;
+  isVariant: boolean;
+  quantization: string | null;
+  qualityScore: number;
+  speedMultiplier: number;
+  throughputCapacity: number;
+  isMoE: boolean;
+}
+
+export interface DemandByTier {
+  enterprise: number;
+  'api-paid': number;
+  'consumer-paid': number;
+  'api-free': number;
+  'consumer-free': number;
+}
+
+export interface ServingPipelineInput {
+  modelsState: ModelsState;
+  effectiveInferenceFlops: number;
+  overloadPolicy: OverloadPolicy;
+  demandByTier: DemandByTier;
+  batchApi: BatchApiState;
+  modelQuality: number;
+  researchUnlocks: {
+    servingRoutingUnlocked: boolean;
+    priorityQueuesUnlocked: boolean;
+    batchApiUnlocked: boolean;
+    autoScalingBonus: number;
+  };
+}
+
+export interface ServingPipelineResult {
+  servingMetrics: ServingMetrics;
+  batchApi: BatchApiState;
+  batchRevenue: number;
+}
+
+function buildModelFleet(
+  modelsState: ModelsState,
+  effectiveInferenceFlops: number,
+): ModelServingSlot[] {
+  const slots: ModelServingSlot[] = [];
+
+  const deployedBases = modelsState.baseModels.filter(m => m.isDeployed);
+  const deployedVariants: { variant: ModelVariant; baseModel: BaseModel }[] = [];
+
+  for (const family of modelsState.families) {
+    for (const variant of family.variants) {
+      if (!variant.isDeployed) continue;
+      const base = modelsState.baseModels.find(m => m.id === variant.baseModelId);
+      if (base) deployedVariants.push({ variant, baseModel: base });
+    }
+  }
+
+  const totalDeployed = deployedBases.length + deployedVariants.length;
+  if (totalDeployed === 0 || effectiveInferenceFlops <= 0) return slots;
+
+  const flopsPerModel = effectiveInferenceFlops / totalDeployed;
+
+  for (const model of deployedBases) {
+    const sizeFactor = MODEL_SIZE_THROUGHPUT_SCALER[model.sizeTier] ?? 1.0;
+    const moeFactor = model.architecture.type === 'moe' ? MOE_SPEED_MULTIPLIER : 1.0;
+    const throughput = flopsPerModel * FLOPS_TO_TOKENS_MULTIPLIER * sizeFactor * moeFactor;
+
+    slots.push({
+      modelId: model.id,
+      modelName: model.name,
+      sizeTier: model.sizeTier,
+      isVariant: false,
+      quantization: null,
+      qualityScore: model.rawCapability / 100,
+      speedMultiplier: moeFactor,
+      throughputCapacity: throughput,
+      isMoE: model.architecture.type === 'moe',
+    });
+  }
+
+  for (const { variant, baseModel } of deployedVariants) {
+    const sizeFactor = MODEL_SIZE_THROUGHPUT_SCALER[baseModel.sizeTier] ?? 1.0;
+    const moeFactor = variant.architecture.type === 'moe' ? MOE_SPEED_MULTIPLIER : 1.0;
+    const quantConfig = variant.quantization ? QUANTIZATION_CONFIGS[variant.quantization] : null;
+    const quantSpeedFactor = quantConfig?.speedMultiplier ?? 1.0;
+    const qualityRetention = quantConfig?.qualityRetention ?? 1.0;
+    const throughput = flopsPerModel * FLOPS_TO_TOKENS_MULTIPLIER * sizeFactor * moeFactor * quantSpeedFactor;
+
+    slots.push({
+      modelId: variant.id,
+      modelName: variant.name,
+      sizeTier: baseModel.sizeTier,
+      isVariant: true,
+      quantization: variant.quantization ?? null,
+      qualityScore: (baseModel.rawCapability / 100) * qualityRetention,
+      speedMultiplier: moeFactor * quantSpeedFactor,
+      throughputCapacity: throughput,
+      isMoE: variant.architecture.type === 'moe',
+    });
+  }
+
+  return slots;
+}
+
+function sortFleetByStrategy(
+  fleet: ModelServingSlot[],
+  strategy: string,
+  overallUtilization: number,
+): ModelServingSlot[] {
+  const sorted = [...fleet];
+  switch (strategy) {
+    case 'quality-first':
+      sorted.sort((a, b) => b.qualityScore - a.qualityScore);
+      break;
+    case 'speed-first':
+      sorted.sort((a, b) => b.throughputCapacity - a.throughputCapacity);
+      break;
+    case 'balanced':
+    default:
+      if (overallUtilization > 0.8) {
+        sorted.sort((a, b) => b.throughputCapacity - a.throughputCapacity);
+      } else {
+        sorted.sort((a, b) => b.qualityScore - a.qualityScore);
+      }
+      break;
+  }
+  return sorted;
+}
+
+interface FleetState {
+  remaining: Map<string, number>;
+  used: Map<string, number>;
+}
+
+function serveFromFleet(
+  demand: number,
+  fleet: ModelServingSlot[],
+  fleetState: FleetState,
+  policy: OverloadPolicy,
+  tier: TrafficPriority,
+  overallUtilization: number,
+): TierServingMetrics {
+  if (demand <= 0) {
+    return { demandTokens: 0, servedTokens: 0, queuedTokens: 0, rejectedTokens: 0, degradedTokens: 0, avgQualityDelivered: 1 };
+  }
+
+  let remaining = demand;
+  let served = 0;
+  let degraded = 0;
+  let qualityWeightedSum = 0;
+
+  const bestQuality = fleet.length > 0 ? Math.max(...fleet.map(s => s.qualityScore)) : 1;
+  const degradationActive = policy.autoDegradation.enabled && overallUtilization > policy.autoDegradation.triggerThreshold;
+
+  for (const slot of fleet) {
+    if (remaining <= 0) break;
+
+    const isDegraded = slot.qualityScore < bestQuality * 0.95;
+    if (isDegraded && !degradationActive) continue;
+    if (isDegraded && slot.qualityScore < policy.autoDegradation.minQualityFloor) continue;
+
+    const available = fleetState.remaining.get(slot.modelId) ?? 0;
+    if (available <= 0) continue;
+
+    const toServe = Math.min(remaining, available);
+    fleetState.remaining.set(slot.modelId, available - toServe);
+    fleetState.used.set(slot.modelId, (fleetState.used.get(slot.modelId) ?? 0) + toServe);
+
+    served += toServe;
+    if (isDegraded) degraded += toServe;
+    qualityWeightedSum += toServe * slot.qualityScore;
+    remaining -= toServe;
+  }
+
+  let queued = 0;
+  let rejected = 0;
+
+  if (remaining > 0) {
+    const behavior = policy.overflowBehavior[tier];
+    switch (behavior) {
+      case 'queue':
+        queued = remaining;
+        break;
+      case 'reject':
+        rejected = remaining;
+        break;
+      case 'degrade':
+        for (const slot of fleet) {
+          if (remaining <= 0) break;
+          const available = fleetState.remaining.get(slot.modelId) ?? 0;
+          if (available <= 0) continue;
+
+          const toServe = Math.min(remaining, available);
+          fleetState.remaining.set(slot.modelId, available - toServe);
+          fleetState.used.set(slot.modelId, (fleetState.used.get(slot.modelId) ?? 0) + toServe);
+          served += toServe;
+          degraded += toServe;
+          qualityWeightedSum += toServe * slot.qualityScore;
+          remaining -= toServe;
+        }
+        rejected = remaining;
+        break;
+    }
+  }
+
+  const avgQuality = served > 0 ? qualityWeightedSum / served : bestQuality;
+
+  return {
+    demandTokens: demand,
+    servedTokens: served,
+    queuedTokens: queued,
+    rejectedTokens: rejected,
+    degradedTokens: degraded,
+    avgQualityDelivered: avgQuality,
+  };
+}
+
+export function processServingPipeline(input: ServingPipelineInput): ServingPipelineResult {
+  const { modelsState, effectiveInferenceFlops, overloadPolicy, demandByTier, batchApi, modelQuality, researchUnlocks } = input;
+
+  const fleet = buildModelFleet(modelsState, effectiveInferenceFlops);
+  const totalFleetCapacity = fleet.reduce((sum, s) => sum + s.throughputCapacity, 0);
+
+  if (fleet.length === 0 || totalFleetCapacity <= 0) {
+    const metrics = makeInitialServingMetrics();
+    for (const tier of Object.keys(demandByTier) as TrafficPriority[]) {
+      const demand = demandByTier[tier] ?? 0;
+      if (demand > 0) {
+        metrics.tierMetrics[tier] = {
+          demandTokens: demand,
+          servedTokens: 0,
+          queuedTokens: 0,
+          rejectedTokens: demand,
+          degradedTokens: 0,
+          avgQualityDelivered: 0,
+        };
+        metrics.totalRejected += demand;
+      }
+    }
+    return {
+      servingMetrics: metrics,
+      batchApi: { ...batchApi, servedLastTick: 0, revenue: 0 },
+      batchRevenue: 0,
+    };
+  }
+
+  const totalDemand = Object.values(demandByTier).reduce((s, v) => s + v, 0);
+  const overallUtilization = totalFleetCapacity > 0 ? totalDemand / totalFleetCapacity : 0;
+
+  const effectiveStrategy = researchUnlocks.servingRoutingUnlocked
+    ? overloadPolicy.routingStrategy
+    : 'balanced';
+
+  const sortedFleet = sortFleetByStrategy(fleet, effectiveStrategy, overallUtilization);
+
+  const fleetState: FleetState = {
+    remaining: new Map(fleet.map(s => [s.modelId, s.throughputCapacity])),
+    used: new Map(fleet.map(s => [s.modelId, 0])),
+  };
+
+  const reservedCapacity = totalFleetCapacity * overloadPolicy.enterpriseReservation;
+  const enterpriseDemand = demandByTier['enterprise'] ?? 0;
+
+  if (reservedCapacity > 0 && enterpriseDemand > 0) {
+    const reservePerModel = reservedCapacity / fleet.length;
+    for (const slot of sortedFleet) {
+      const current = fleetState.remaining.get(slot.modelId) ?? 0;
+      const reserved = Math.min(reservePerModel, current);
+      fleetState.remaining.set(slot.modelId, current - reserved);
+    }
+  }
+
+  const effectivePriorityOrder = researchUnlocks.priorityQueuesUnlocked
+    ? overloadPolicy.priorityOrder
+    : ['enterprise', 'api-paid', 'consumer-paid', 'api-free', 'consumer-free'] as TrafficPriority[];
+
+  const tierResults: Record<TrafficPriority, TierServingMetrics> = {} as Record<TrafficPriority, TierServingMetrics>;
+
+  const nonEnterpriseTiers = effectivePriorityOrder.filter(t => t !== 'enterprise');
+
+  if (enterpriseDemand > 0) {
+    const enterpriseFleetState: FleetState = {
+      remaining: new Map(fleet.map(s => [s.modelId, s.throughputCapacity])),
+      used: new Map(fleet.map(s => [s.modelId, 0])),
+    };
+
+    const reserveLimit = reservedCapacity > 0 ? reservedCapacity : totalFleetCapacity;
+    let budgetLeft = reserveLimit;
+    for (const slot of sortedFleet) {
+      const cap = slot.throughputCapacity;
+      const alloc = Math.min(cap, budgetLeft);
+      enterpriseFleetState.remaining.set(slot.modelId, alloc);
+      budgetLeft -= alloc;
+      if (budgetLeft <= 0) break;
+    }
+
+    const effectiveEntDemand = researchUnlocks.servingRoutingUnlocked
+      ? Math.min(enterpriseDemand, overloadPolicy.rateLimitPerCustomer['enterprise'] * 100)
+      : enterpriseDemand;
+
+    tierResults['enterprise'] = serveFromFleet(
+      effectiveEntDemand, sortedFleet, enterpriseFleetState, overloadPolicy, 'enterprise', overallUtilization,
+    );
+
+    for (const slot of fleet) {
+      const entUsed = enterpriseFleetState.used.get(slot.modelId) ?? 0;
+      const mainRemaining = fleetState.remaining.get(slot.modelId) ?? 0;
+      fleetState.remaining.set(slot.modelId, Math.max(0, mainRemaining - entUsed + (reservedCapacity > 0 ? reservedCapacity / fleet.length : 0)));
+      fleetState.used.set(slot.modelId, entUsed);
+    }
+  } else {
+    tierResults['enterprise'] = { demandTokens: 0, servedTokens: 0, queuedTokens: 0, rejectedTokens: 0, degradedTokens: 0, avgQualityDelivered: 1 };
+
+    if (reservedCapacity > 0) {
+      const reservePerModel = reservedCapacity / fleet.length;
+      for (const slot of fleet) {
+        const current = fleetState.remaining.get(slot.modelId) ?? 0;
+        fleetState.remaining.set(slot.modelId, current + reservePerModel);
+      }
+    }
+  }
+
+  for (const tier of nonEnterpriseTiers) {
+    const rawDemand = demandByTier[tier] ?? 0;
+    const effectiveDemand = researchUnlocks.servingRoutingUnlocked
+      ? Math.min(rawDemand, overloadPolicy.rateLimitPerCustomer[tier] * 100)
+      : rawDemand;
+
+    tierResults[tier] = serveFromFleet(
+      effectiveDemand, sortedFleet, fleetState, overloadPolicy, tier, overallUtilization,
+    );
+  }
+
+  for (const tier of effectivePriorityOrder) {
+    if (!(tier in tierResults)) {
+      tierResults[tier] = { demandTokens: 0, servedTokens: 0, queuedTokens: 0, rejectedTokens: 0, degradedTokens: 0, avgQualityDelivered: 1 };
+    }
+  }
+
+  let batchTokensServed = 0;
+  let batchRevenue = 0;
+  const updatedBatchApi = { ...batchApi };
+
+  if (overloadPolicy.batchApiEnabled && researchUnlocks.batchApiUnlocked) {
+    let idleCapacity = 0;
+    for (const slot of fleet) {
+      const remaining = fleetState.remaining.get(slot.modelId) ?? 0;
+      idleCapacity += remaining;
+    }
+
+    const pendingBatch = Math.min(batchApi.pendingQueue + batchApi.totalBatchDemand, BATCH_API_MAX_PENDING);
+    batchTokensServed = Math.min(pendingBatch, idleCapacity);
+
+    const baseTokenPrice = 3.0;
+    batchRevenue = (batchTokensServed / 1_000_000) * baseTokenPrice * (1 - overloadPolicy.batchApiDiscount);
+
+    updatedBatchApi.pendingQueue = Math.max(0, pendingBatch - batchTokensServed);
+    updatedBatchApi.servedLastTick = batchTokensServed;
+    updatedBatchApi.revenue = batchRevenue;
+  }
+
+  const totalServed = Object.values(tierResults).reduce((s, t) => s + t.servedTokens, 0);
+  const totalQueued = Object.values(tierResults).reduce((s, t) => s + t.queuedTokens, 0);
+  const totalRejected = Object.values(tierResults).reduce((s, t) => s + t.rejectedTokens, 0);
+  const totalDegraded = Object.values(tierResults).reduce((s, t) => s + t.degradedTokens, 0);
+
+  let effectiveQuality = modelQuality;
+  if (totalServed > 0) {
+    let qualitySum = 0;
+    for (const t of Object.values(tierResults)) {
+      qualitySum += t.avgQualityDelivered * t.servedTokens;
+    }
+    effectiveQuality = qualitySum / totalServed;
+  }
+
+  const queuedFraction = totalDemand > 0 ? totalQueued / totalDemand : 0;
+  const avgLatencyMs = BASE_LATENCY_MS + queuedFraction * 100 * QUEUE_LATENCY_MS_PER_PERCENT;
+
+  const modelUtilization: ModelUtilizationEntry[] = fleet.map(slot => ({
+    modelId: slot.modelId,
+    modelName: slot.modelName,
+    quantization: slot.quantization,
+    qualityScore: slot.qualityScore,
+    throughputCapacity: slot.throughputCapacity,
+    throughputUsed: fleetState.used.get(slot.modelId) ?? 0,
+    utilization: slot.throughputCapacity > 0
+      ? Math.min(1, (fleetState.used.get(slot.modelId) ?? 0) / slot.throughputCapacity)
+      : 0,
+  }));
+
+  const autoScaleBoost = researchUnlocks.autoScalingBonus;
+  if (autoScaleBoost > 0) {
+    for (const tier of Object.keys(tierResults) as TrafficPriority[]) {
+      const metrics = tierResults[tier];
+      if (metrics.rejectedTokens > 0) {
+        const recovered = Math.min(metrics.rejectedTokens, metrics.rejectedTokens * autoScaleBoost);
+        tierResults[tier] = {
+          ...metrics,
+          servedTokens: metrics.servedTokens + recovered,
+          rejectedTokens: metrics.rejectedTokens - recovered,
+        };
+      }
+    }
+  }
+
+  return {
+    servingMetrics: {
+      tierMetrics: tierResults,
+      totalServed,
+      totalQueued,
+      totalRejected,
+      totalDegraded,
+      effectiveQuality,
+      avgLatencyMs,
+      modelUtilization,
+      batchApiTokensServed: batchTokensServed,
+      batchApiRevenue: batchRevenue,
+    },
+    batchApi: updatedBatchApi,
+    batchRevenue,
+  };
+}
+
+export function computeSatisfactionImpact(
+  metrics: TierServingMetrics,
+): number {
+  if (metrics.demandTokens <= 0) return 0;
+
+  const rejectedFraction = metrics.rejectedTokens / metrics.demandTokens;
+  const queuedFraction = metrics.queuedTokens / metrics.demandTokens;
+  const degradedFraction = metrics.servedTokens > 0 ? metrics.degradedTokens / metrics.servedTokens : 0;
+
+  const rejectionPenalty = rejectedFraction * REJECTION_SATISFACTION_PENALTY * 10;
+  const queuePenalty = queuedFraction * QUEUE_SATISFACTION_PENALTY * 10;
+  const degradationPenalty = degradedFraction * (1 - metrics.avgQualityDelivered) * DEGRADATION_SATISFACTION_PENALTY * 10;
+
+  return -(rejectionPenalty + queuePenalty + degradationPenalty);
+}

packages/game-engine/src/systems/marketSystem.ts

@@ -1,8 +1,9 @@
 import type { GameState } from '@ai-tycoon/shared';
 import { processMarketV2 } from './market/index';
+import type { ResearchBonuses } from './researchBonuses';
 
 export type { MarketTickResult } from './market/index';
 
-export function processMarket(state: GameState, currentTickCapacity: number) {
-  return processMarketV2(state, currentTickCapacity);
+export function processMarket(state: GameState, currentTickCapacity: number, effectiveInferenceFlops?: number, researchBonuses?: ResearchBonuses) {
+  return processMarketV2(state, currentTickCapacity, effectiveInferenceFlops, researchBonuses);
 }

packages/game-engine/src/systems/researchBonuses.ts

@@ -18,6 +18,7 @@ export interface ResearchBonuses {
   reputationBonus: number;
 
   safetyBonus: number;
+  autoScalingBonus: number;
 }
 
 export function getResearchBonuses(completedResearch: string[]): ResearchBonuses {
@@ -37,6 +38,7 @@ export function getResearchBonuses(completedResearch: string[]): ResearchBonuses
     agentsBonus: 0,
     reputationBonus: 0,
     safetyBonus: 0,
+    autoScalingBonus: 0,
   };
 
   for (const id of completedResearch) {
@@ -53,6 +55,7 @@ export function getResearchBonuses(completedResearch: string[]): ResearchBonuses
             case 'pipeline_speed': bonuses.pipelineSpeedBonus += effect.value; break;
             case 'data_quality': bonuses.dataQualityBonus += effect.value; break;
             case 'sdk_coverage': bonuses.sdkCoverageBonus += effect.value; break;
+            case 'auto_scaling': bonuses.autoScalingBonus += effect.value; break;
           }
           break;
         case 'capability_boost':

packages/game-engine/src/tick.ts

@@ -56,7 +56,7 @@ export function processTick(state: GameState): Partial<GameState> {
   const stateWithModels = { ...stateWithInfra, models: modelResult.modelsState };
 
   const capacity = computeCapacity(state, infrastructure, researchBonuses);
-  const market = processMarket(stateWithModels, capacity.tokensPerSecondCapacity);
+  const market = processMarket(stateWithModels, capacity.tokensPerSecondCapacity, capacity.effectiveInferenceFlops, researchBonuses);
   const compute = finalizeCompute(capacity, market.totalTokenDemand);
 
   const talent = processTalent(stateWithModels);

packages/shared/src/constants/gameBalance.ts

@@ -118,6 +118,34 @@ export const FLOPS_TO_TOKENS_MULTIPLIER = 26;
 
 export const OVERLOAD_PENALTY_EXPONENT = 1.5;
 
+// --- Serving Pipeline ---
+
+export const REJECTION_SATISFACTION_PENALTY = 0.15;
+export const QUEUE_SATISFACTION_PENALTY = 0.05;
+export const DEGRADATION_SATISFACTION_PENALTY = 0.08;
+
+export const REJECTION_CHURN_MULTIPLIER = 3.0;
+export const QUEUE_CHURN_MULTIPLIER = 1.5;
+
+export const ENTERPRISE_REJECTION_SLA_MULTIPLIER = 3.0;
+
+export const FREE_TIER_REJECTION_TOLERANCE = 0.3;
+export const PAID_TIER_REJECTION_TOLERANCE = 0.05;
+
+export const MODEL_SIZE_THROUGHPUT_SCALER: Record<SizeTier, number> = {
+  nano: 10.0, small: 5.0, medium: 2.0, large: 1.2, flagship: 1.0,
+};
+
+export const BATCH_API_DEMAND_PER_DEV: Record<ApiTierId, number> = {
+  free: 0, payg: 2, scale: 20, 'enterprise-api': 100,
+};
+export const BATCH_API_DEFAULT_DISCOUNT = 0.5;
+export const BATCH_API_MAX_PENDING = 100_000;
+
+export const BATCHING_THROUGHPUT_FACTOR = 0.15;
+export const BASE_LATENCY_MS = 50;
+export const QUEUE_LATENCY_MS_PER_PERCENT = 5;
+
 export const ERA_THRESHOLDS = {
   scaleup: { revenue: 10_000, capability: 15, reputation: 30 },
   bigtech: { revenue: 1_000_000, capability: 50, reputation: 60 },

packages/shared/src/types/gameState.ts

@@ -52,4 +52,4 @@ export const INITIAL_SETTINGS: GameSettings = {
   musicVolume: 0.5,
 };
 
-export const SAVE_VERSION = 8;
+export const SAVE_VERSION = 9;

packages/shared/src/types/market.ts

@@ -170,13 +170,93 @@ export interface ObsolescenceState {
   newModelBoostRemaining: number;
 }
 
-// --- Overload Policy (kept from original) ---
+// --- Serving Pipeline & Overload Policy ---
+
+export type TrafficPriority = 'enterprise' | 'api-paid' | 'consumer-paid' | 'api-free' | 'consumer-free';
+export type RoutingStrategy = 'quality-first' | 'speed-first' | 'balanced';
+export type OverflowBehavior = 'queue' | 'reject' | 'degrade';
+
+export const TRAFFIC_PRIORITIES: TrafficPriority[] = ['enterprise', 'api-paid', 'consumer-paid', 'api-free', 'consumer-free'];
 
 export interface OverloadPolicy {
+  priorityOrder: TrafficPriority[];
+  overflowBehavior: Record<TrafficPriority, OverflowBehavior>;
   maxQueueDepth: number;
-  rateLimitPerCustomer: number;
-  degradeQualityUnderLoad: boolean;
-  prioritizeEnterprise: boolean;
+  rateLimitPerCustomer: Record<TrafficPriority, number>;
+  enterpriseReservation: number;
+  routingStrategy: RoutingStrategy;
+  autoDegradation: {
+    enabled: boolean;
+    triggerThreshold: number;
+    minQualityFloor: number;
+  };
+  batchApiEnabled: boolean;
+  batchApiDiscount: number;
+  batchApiMaxDelay: number;
+}
+
+export interface TierServingMetrics {
+  demandTokens: number;
+  servedTokens: number;
+  queuedTokens: number;
+  rejectedTokens: number;
+  degradedTokens: number;
+  avgQualityDelivered: number;
+}
+
+export interface ModelUtilizationEntry {
+  modelId: string;
+  modelName: string;
+  quantization: string | null;
+  qualityScore: number;
+  throughputCapacity: number;
+  throughputUsed: number;
+  utilization: number;
+}
+
+export interface ServingMetrics {
+  tierMetrics: Record<TrafficPriority, TierServingMetrics>;
+  totalServed: number;
+  totalQueued: number;
+  totalRejected: number;
+  totalDegraded: number;
+  effectiveQuality: number;
+  avgLatencyMs: number;
+  modelUtilization: ModelUtilizationEntry[];
+  batchApiTokensServed: number;
+  batchApiRevenue: number;
+}
+
+export interface BatchApiState {
+  totalBatchDemand: number;
+  pendingQueue: number;
+  servedLastTick: number;
+  revenue: number;
+}
+
+function makeEmptyTierMetrics(): TierServingMetrics {
+  return { demandTokens: 0, servedTokens: 0, queuedTokens: 0, rejectedTokens: 0, degradedTokens: 0, avgQualityDelivered: 1 };
+}
+
+export function makeInitialServingMetrics(): ServingMetrics {
+  return {
+    tierMetrics: {
+      'enterprise': makeEmptyTierMetrics(),
+      'api-paid': makeEmptyTierMetrics(),
+      'consumer-paid': makeEmptyTierMetrics(),
+      'api-free': makeEmptyTierMetrics(),
+      'consumer-free': makeEmptyTierMetrics(),
+    },
+    totalServed: 0,
+    totalQueued: 0,
+    totalRejected: 0,
+    totalDegraded: 0,
+    effectiveQuality: 1,
+    avgLatencyMs: 0,
+    modelUtilization: [],
+    batchApiTokensServed: 0,
+    batchApiRevenue: 0,
+  };
 }
 
 // --- Root Market State ---
@@ -193,6 +273,8 @@ export interface MarketState {
   seasonalMultiplier: number;
   obsolescence: ObsolescenceState;
   overloadPolicy: OverloadPolicy;
+  servingMetrics: ServingMetrics;
+  batchApi: BatchApiState;
   openSourcedModels: string[];
   subscriberHistory: { tick: number; subscribers: number }[];
 }
@@ -315,10 +397,39 @@ export const INITIAL_MARKET: MarketState = {
     newModelBoostRemaining: 0,
   },
   overloadPolicy: {
+    priorityOrder: ['enterprise', 'api-paid', 'consumer-paid', 'api-free', 'consumer-free'],
+    overflowBehavior: {
+      'enterprise': 'queue' as OverflowBehavior,
+      'api-paid': 'queue' as OverflowBehavior,
+      'consumer-paid': 'degrade' as OverflowBehavior,
+      'api-free': 'reject' as OverflowBehavior,
+      'consumer-free': 'reject' as OverflowBehavior,
+    },
     maxQueueDepth: 100,
-    rateLimitPerCustomer: 1000,
-    degradeQualityUnderLoad: false,
-    prioritizeEnterprise: true,
+    rateLimitPerCustomer: {
+      'enterprise': 10000,
+      'api-paid': 1000,
+      'consumer-paid': 500,
+      'api-free': 100,
+      'consumer-free': 50,
+    },
+    enterpriseReservation: 0.2,
+    routingStrategy: 'balanced' as RoutingStrategy,
+    autoDegradation: {
+      enabled: true,
+      triggerThreshold: 0.85,
+      minQualityFloor: 0.75,
+    },
+    batchApiEnabled: false,
+    batchApiDiscount: 0.5,
+    batchApiMaxDelay: 60,
+  },
+  servingMetrics: makeInitialServingMetrics(),
+  batchApi: {
+    totalBatchDemand: 0,
+    pendingQueue: 0,
+    servedLastTick: 0,
+    revenue: 0,
   },
   openSourcedModels: [],
   subscriberHistory: [],