Redesign model lifecycle: upfront SFT/alignment, multi-size families, point releases, quantization-only variants

Training pipeline now requires SFT specializations and alignment method configured at start — no more
mid-training configuration step. Model families support multiple size tiers (Nano/Small/Medium/Large/Flagship)
trained independently, mimicking real AI company model families. Point releases iterate on deployed models
with 40% training time and 8% capability gain. Distillation and fine-tuning variants removed — players
train smaller size tiers or configure SFT during initial training instead. Only quantization remains as
a variant type.

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

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

@@ -7,8 +7,6 @@ import type {
 import { BENCHMARKS } from '../data/benchmarks';
 import {
   uuid, VRAM_REQUIREMENTS_BY_GENERATION,
-  SFT_TIME_FRACTION, SFT_COMPUTE_FRACTION,
-  ALIGNMENT_TIME_FRACTION, ALIGNMENT_COMPUTE_FRACTION,
   MOE_CAPABILITY_MULTIPLIER, MOE_SPEED_MULTIPLIER,
   EVENT_BASE_PROBABILITY,
   LOSS_SPIKE_DELAY_MIN, LOSS_SPIKE_DELAY_MAX,
@@ -18,8 +16,8 @@ import {
   ALIGNMENT_METHODS,
   SFT_SPECIALIZATION_BONUSES,
   QUANTIZATION_CONFIGS,
-  DISTILLATION_BASE_RETENTION,
-  QUANTIZATION_TICKS,
+  POINT_RELEASE_CAPABILITY_GAIN,
+  SIZE_TIER_LABELS,
 } from '@ai-tycoon/shared';
 import type { ResearchBonuses } from './researchBonuses';
 
@@ -101,60 +99,25 @@ export function processModels(state: GameState, researchBonuses?: ResearchBonuse
       stage.computeAllocated = effectiveFlops;
       stage.lossValue = Math.max(0.01, 10 * Math.exp(-stage.progressTicks / stage.totalTicks * 3));
 
-      const progressRatio = stage.progressTicks / stage.totalTicks;
-      if (progressRatio >= 0.75 && progressRatio < 0.78 && !pipeline.stages.sft && !pipeline.stages.alignment) {
-        notifications.push({
-          title: 'Post-Training Reminder',
-          message: `${pipeline.modelName} is 75% pre-trained. Configure SFT/Alignment now or they'll be skipped!`,
-          type: 'warning',
-          action: { label: 'Configure Now', page: 'models', modelsTab: 'overview' },
-        });
-      }
-
       if (stage.progressTicks >= stage.totalTicks) {
         stage.isComplete = true;
         stage.progressTicks = stage.totalTicks;
-
-        if (updated.stages.sft) {
-          updated.currentStage = 'sft';
-          notifications.push({ title: 'Pre-training Complete', message: `${pipeline.modelName}: Moving to supervised fine-tuning.`, type: 'info' });
-        } else if (updated.stages.alignment) {
-          updated.currentStage = 'alignment';
-          notifications.push({ title: 'Pre-training Complete', message: `${pipeline.modelName}: Moving to alignment.`, type: 'info' });
-        } else {
-          const model = createBaseModel(updated, state, researchBonuses);
-          baseModels = [...baseModels, model];
-          families = families.map(f =>
-            f.id === pipeline.familyId ? { ...f, baseModelId: model.id } : f,
-          );
-          completedModels.push(model);
-          updated.status = 'completed';
-        }
+        updated.currentStage = 'sft';
+        notifications.push({ title: 'Pre-training Complete', message: `${pipeline.modelName}: Moving to supervised fine-tuning.`, type: 'info' });
       }
       updated = { ...updated, stages: { ...updated.stages, pretraining: stage } };
-    } else if (pipeline.currentStage === 'sft' && pipeline.stages.sft) {
+    } else if (pipeline.currentStage === 'sft') {
       const stage = { ...pipeline.stages.sft };
       stage.progressTicks += speedMultiplier;
 
       if (stage.progressTicks >= stage.totalTicks) {
         stage.isComplete = true;
         stage.progressTicks = stage.totalTicks;
-
-        if (updated.stages.alignment) {
-          updated.currentStage = 'alignment';
-          notifications.push({ title: 'SFT Complete', message: `${pipeline.modelName}: Moving to alignment.`, type: 'info' });
-        } else {
-          const model = createBaseModel(updated, state, researchBonuses);
-          baseModels = [...baseModels, model];
-          families = families.map(f =>
-            f.id === pipeline.familyId ? { ...f, baseModelId: model.id } : f,
-          );
-          completedModels.push(model);
-          updated.status = 'completed';
-        }
+        updated.currentStage = 'alignment';
+        notifications.push({ title: 'SFT Complete', message: `${pipeline.modelName}: Moving to alignment.`, type: 'info' });
       }
       updated = { ...updated, stages: { ...updated.stages, sft: stage } };
-    } else if (pipeline.currentStage === 'alignment' && pipeline.stages.alignment) {
+    } else if (pipeline.currentStage === 'alignment') {
       const stage = { ...pipeline.stages.alignment };
       stage.progressTicks += speedMultiplier;
 
@@ -165,7 +128,7 @@ export function processModels(state: GameState, researchBonuses?: ResearchBonuse
         const model = createBaseModel(updated, state, researchBonuses);
         baseModels = [...baseModels, model];
         families = families.map(f =>
-          f.id === pipeline.familyId ? { ...f, baseModelId: model.id } : f,
+          f.id === pipeline.familyId ? { ...f, baseModelIds: [...f.baseModelIds, model.id] } : f,
         );
         completedModels.push(model);
         updated.status = 'completed';
@@ -320,79 +283,93 @@ function createBaseModel(
   const dataTokens = pipeline.stages.pretraining.targetTokens;
   const params = architecture.totalParameters;
 
-  // Pillar 1: Parameters (0-30) — larger models have higher ceiling
-  const paramFactor = Math.min(30, Math.log2(1 + params) * 4.5);
+  const sourceModel = pipeline.isPointRelease && pipeline.sourceModelId
+    ? state.models.baseModels.find(m => m.id === pipeline.sourceModelId)
+    : null;
 
-  // Pillar 2: Compute (0-25) — compute relative to parameter count (Chinchilla scaling)
-  const computePerParam = compute / Math.max(1, params);
-  const computeFactor = Math.min(25, Math.sqrt(computePerParam) * 8);
+  let rawCapability: number;
+  let capabilities: ModelCapabilities;
 
-  // Pillar 3: Data (0-20) — token count with quality multiplier
-  const dataQualityMultiplier = 1 + (researchBonuses?.dataQualityBonus ?? 0);
-  const dataFactor = Math.min(20, Math.log10(1 + dataTokens / 1e8) * 8 * dataQualityMultiplier);
+  if (sourceModel) {
+    rawCapability = Math.min(98, sourceModel.rawCapability * (1 + POINT_RELEASE_CAPABILITY_GAIN));
 
-  // Pillar 4: Research (0-20) — accumulated research knowledge
-  const capabilityResearchBonus = researchBonuses?.globalCapabilityBonus ?? 0;
-  const researchFactor = Math.min(20, capabilityResearchBonus + state.research.completedResearch.length * 0.5);
-
-  let rawCapability = Math.min(95, paramFactor + computeFactor + dataFactor + researchFactor);
-
-  if (architecture.type === 'moe') {
-    rawCapability = Math.min(98, rawCapability * MOE_CAPABILITY_MULTIPLIER);
-  }
-
-  // MoE tradeoff: total params need full VRAM even though only active params run
-  // This is enforced in the UI/store when checking VRAM requirements
-
-  const researcherQuality = state.talent.departments.research.effectiveness;
-  const contextBonus = Math.log2(Math.max(1, architecture.contextWindow / 4)) * 3;
-  const contextPenalty = Math.max(0, Math.log2(architecture.contextWindow / 8)) * 2;
-
-  const capabilities: ModelCapabilities = {
-    reasoning: clamp(rawCapability * (0.6 + dataMix.scientific * 0.5 + dataMix.code * 0.3) * (1 + researcherQuality * 0.2)),
-    coding: clamp(rawCapability * (0.5 + dataMix.code * 1.0)),
-    creative: clamp(rawCapability * (0.4 + dataMix.books * 0.6 + dataMix.conversation * 0.3)),
-    math: clamp(rawCapability * (0.3 + dataMix.scientific * 0.7 + dataMix.code * 0.2)),
-    knowledge: clamp(rawCapability * (0.5 + dataMix.web * 0.3 + dataMix.books * 0.3) + contextBonus * 0.3),
-    multimodal: clamp(rawCapability * (dataMix.images * 0.5 + dataMix.video * 0.4 + dataMix.audio * 0.2)),
-    agents: clamp(rawCapability * (0.2 + dataMix.code * 0.3 + dataMix.conversation * 0.2) + contextBonus * 0.5),
-    speed: Math.max(1, 100 - params * 0.3 - contextPenalty + (researchBonuses?.inferenceEfficiencyBonus ?? 0) * 20 + (architecture.type === 'moe' ? MOE_SPEED_MULTIPLIER * 10 : 0)),
-    contextUtilization: Math.min(100, architecture.contextWindow * 0.4),
-  };
-
-  if (researchBonuses) {
-    capabilities.reasoning = clamp(capabilities.reasoning + researchBonuses.reasoningBonus);
-    capabilities.coding = clamp(capabilities.coding + researchBonuses.codingBonus);
-    capabilities.creative = clamp(capabilities.creative + researchBonuses.creativeBonus);
-    capabilities.multimodal = clamp(capabilities.multimodal + researchBonuses.multimodalBonus);
-    capabilities.agents = clamp(capabilities.agents + researchBonuses.agentsBonus);
-  }
-
-  const breakthroughBonuses: Partial<Record<keyof ModelCapabilities, number>> = {};
-  for (const event of pipeline.events) {
-    if ((event.type === 'breakthrough' || event.type === 'emergent_capability') && event.impact.capabilityDomain && event.impact.capabilityBonus) {
-      const domain = event.impact.capabilityDomain;
-      breakthroughBonuses[domain] = (breakthroughBonuses[domain] ?? 0) + event.impact.capabilityBonus;
+    capabilities = { ...sourceModel.capabilities };
+    const boost = POINT_RELEASE_CAPABILITY_GAIN;
+    for (const key of Object.keys(capabilities) as (keyof ModelCapabilities)[]) {
+      if (key !== 'speed' && key !== 'contextUtilization') {
+        capabilities[key] = clamp(capabilities[key] * (1 + boost));
+      }
+    }
+  } else {
+    const paramFactor = Math.min(30, Math.log2(1 + params) * 4.5);
+
+    const computePerParam = compute / Math.max(1, params);
+    const computeFactor = Math.min(25, Math.sqrt(computePerParam) * 8);
+
+    const dataQualityMultiplier = 1 + (researchBonuses?.dataQualityBonus ?? 0);
+    const dataFactor = Math.min(20, Math.log10(1 + dataTokens / 1e8) * 8 * dataQualityMultiplier);
+
+    const capabilityResearchBonus = researchBonuses?.globalCapabilityBonus ?? 0;
+    const researchFactor = Math.min(20, capabilityResearchBonus + state.research.completedResearch.length * 0.5);
+
+    rawCapability = Math.min(95, paramFactor + computeFactor + dataFactor + researchFactor);
+
+    if (architecture.type === 'moe') {
+      rawCapability = Math.min(98, rawCapability * MOE_CAPABILITY_MULTIPLIER);
+    }
+
+    const researcherQuality = state.talent.departments.research.effectiveness;
+    const contextBonus = Math.log2(Math.max(1, architecture.contextWindow / 4)) * 3;
+    const contextPenalty = Math.max(0, Math.log2(architecture.contextWindow / 8)) * 2;
+
+    capabilities = {
+      reasoning: clamp(rawCapability * (0.6 + dataMix.scientific * 0.5 + dataMix.code * 0.3) * (1 + researcherQuality * 0.2)),
+      coding: clamp(rawCapability * (0.5 + dataMix.code * 1.0)),
+      creative: clamp(rawCapability * (0.4 + dataMix.books * 0.6 + dataMix.conversation * 0.3)),
+      math: clamp(rawCapability * (0.3 + dataMix.scientific * 0.7 + dataMix.code * 0.2)),
+      knowledge: clamp(rawCapability * (0.5 + dataMix.web * 0.3 + dataMix.books * 0.3) + contextBonus * 0.3),
+      multimodal: clamp(rawCapability * (dataMix.images * 0.5 + dataMix.video * 0.4 + dataMix.audio * 0.2)),
+      agents: clamp(rawCapability * (0.2 + dataMix.code * 0.3 + dataMix.conversation * 0.2) + contextBonus * 0.5),
+      speed: Math.max(1, 100 - params * 0.3 - contextPenalty + (researchBonuses?.inferenceEfficiencyBonus ?? 0) * 20 + (architecture.type === 'moe' ? MOE_SPEED_MULTIPLIER * 10 : 0)),
+      contextUtilization: Math.min(100, architecture.contextWindow * 0.4),
+    };
+
+    if (researchBonuses) {
+      capabilities.reasoning = clamp(capabilities.reasoning + researchBonuses.reasoningBonus);
+      capabilities.coding = clamp(capabilities.coding + researchBonuses.codingBonus);
+      capabilities.creative = clamp(capabilities.creative + researchBonuses.creativeBonus);
+      capabilities.multimodal = clamp(capabilities.multimodal + researchBonuses.multimodalBonus);
+      capabilities.agents = clamp(capabilities.agents + researchBonuses.agentsBonus);
+    }
+
+    const breakthroughBonuses: Partial<Record<keyof ModelCapabilities, number>> = {};
+    for (const event of pipeline.events) {
+      if ((event.type === 'breakthrough' || event.type === 'emergent_capability') && event.impact.capabilityDomain && event.impact.capabilityBonus) {
+        const domain = event.impact.capabilityDomain;
+        breakthroughBonuses[domain] = (breakthroughBonuses[domain] ?? 0) + event.impact.capabilityBonus;
+      }
+    }
+    for (const [domain, bonus] of Object.entries(breakthroughBonuses)) {
+      const key = domain as keyof ModelCapabilities;
+      capabilities[key] = clamp(capabilities[key] + bonus);
     }
-  }
-  for (const [domain, bonus] of Object.entries(breakthroughBonuses)) {
-    const key = domain as keyof ModelCapabilities;
-    capabilities[key] = clamp(capabilities[key] + bonus);
   }
 
   const completedStages: ('pretraining' | 'sft' | 'alignment')[] = ['pretraining'];
 
-  if (pipeline.stages.sft?.isComplete) {
+  if (pipeline.stages.sft.isComplete) {
     completedStages.push('sft');
     const sft = pipeline.stages.sft;
-    for (let i = 0; i < sft.specializations.length; i++) {
-      const spec = sft.specializations[i];
-      const bonuses = SFT_SPECIALIZATION_BONUSES[spec];
-      if (!bonuses) continue;
-      const diminishing = i === 0 ? 1.0 : i === 1 ? 0.7 : 0.4;
-      for (const [cap, value] of Object.entries(bonuses)) {
-        const key = cap as keyof ModelCapabilities;
-        capabilities[key] = clamp(capabilities[key] + value * diminishing);
+    if (!sourceModel) {
+      for (let i = 0; i < sft.specializations.length; i++) {
+        const spec = sft.specializations[i];
+        const bonuses = SFT_SPECIALIZATION_BONUSES[spec];
+        if (!bonuses) continue;
+        const diminishing = i === 0 ? 1.0 : i === 1 ? 0.7 : 0.4;
+        for (const [cap, value] of Object.entries(bonuses)) {
+          const key = cap as keyof ModelCapabilities;
+          capabilities[key] = clamp(capabilities[key] + value * diminishing);
+        }
       }
     }
   }
@@ -401,7 +378,7 @@ function createBaseModel(
   let overallSafety = Math.min(100, 30 + safetyResearchBonus + Math.random() * 10);
   let refusalRate = overallSafety > 60 ? 0.1 : 0.03;
 
-  if (pipeline.stages.alignment?.isComplete) {
+  if (pipeline.stages.alignment.isComplete) {
     completedStages.push('alignment');
     const alignment = pipeline.stages.alignment;
     const methodConfig = ALIGNMENT_METHODS[alignment.method];
@@ -409,10 +386,12 @@ function createBaseModel(
       const safetyGain = methodConfig.safetyGain * alignment.safetyWeight;
       overallSafety = Math.min(100, overallSafety + safetyGain);
       refusalRate = methodConfig.baseRefusal * Math.pow(alignment.safetyWeight, 1.5);
-      const capLoss = methodConfig.capabilityLoss * alignment.safetyWeight * 0.5;
-      for (const key of Object.keys(capabilities) as (keyof ModelCapabilities)[]) {
-        if (key !== 'speed' && key !== 'contextUtilization') {
-          capabilities[key] = clamp(capabilities[key] - capLoss);
+      if (!sourceModel) {
+        const capLoss = methodConfig.capabilityLoss * alignment.safetyWeight * 0.5;
+        for (const key of Object.keys(capabilities) as (keyof ModelCapabilities)[]) {
+          if (key !== 'speed' && key !== 'contextUtilization') {
+            capabilities[key] = clamp(capabilities[key] - capLoss);
+          }
         }
       }
     }
@@ -426,10 +405,15 @@ function createBaseModel(
     honesty: overallSafety * 0.9,
   };
 
+  const family = state.models.families.find(f => f.id === pipeline.familyId);
+  const version = sourceModel ? Math.round((sourceModel.version + 0.1) * 10) / 10 : 1.0;
+  const familyName = family?.name ?? pipeline.modelName;
+  const autoName = `${familyName} ${SIZE_TIER_LABELS[pipeline.sizeTier]} v${version.toFixed(1)}`;
+
   return {
     id: uuid(),
     familyId: pipeline.familyId,
-    name: pipeline.modelName,
+    name: autoName,
     architecture,
     dataMix,
     capabilities,
@@ -439,6 +423,10 @@ function createBaseModel(
     trainedAtTick: state.meta.tickCount,
     trainingCostTotal: compute,
     trainingStagesCompleted: completedStages,
+    sizeTier: pipeline.sizeTier,
+    version,
+    sftSpecializations: pipeline.stages.sft.specializations,
+    alignmentMethod: pipeline.stages.alignment.method,
   };
 }
 
@@ -467,59 +455,30 @@ function createVariant(job: VariantCreationJob, base: BaseModel): ModelVariant {
   const caps = { ...base.capabilities };
   let costMultiplier = 1.0;
   let speedMultiplier = 1.0;
-  let variantName = base.name;
-  let arch = { ...base.architecture };
 
-  if (job.jobType === 'distillation' && 'targetParameters' in job.config) {
-    const config = job.config;
-    const sizeRatio = config.targetParameters / base.architecture.totalParameters;
-    const retention = DISTILLATION_BASE_RETENTION + sizeRatio * 0.25;
+  const config = job.config;
+  const qConfig = QUANTIZATION_CONFIGS[config.level];
+  if (qConfig) {
     for (const key of Object.keys(caps) as (keyof ModelCapabilities)[]) {
-      caps[key] = clamp(caps[key] * retention);
+      if (key !== 'speed') caps[key] = clamp(caps[key] * qConfig.qualityRetention);
     }
-    costMultiplier = sizeRatio * 0.8;
-    speedMultiplier = (1 / sizeRatio) * 0.7;
-    arch = { ...arch, totalParameters: config.targetParameters, activeParameters: config.targetParameters };
-    variantName = config.variantName;
-  } else if (job.jobType === 'fine-tuning' && 'specialization' in job.config) {
-    const config = job.config;
-    const bonuses = SFT_SPECIALIZATION_BONUSES[config.specialization];
-    if (bonuses) {
-      for (const [cap, value] of Object.entries(bonuses)) {
-        caps[cap as keyof ModelCapabilities] = clamp(caps[cap as keyof ModelCapabilities] + value);
-      }
-    }
-    variantName = config.variantName;
-  } else if (job.jobType === 'quantization' && 'level' in job.config) {
-    const config = job.config;
-    const qConfig = QUANTIZATION_CONFIGS[config.level];
-    if (qConfig) {
-      for (const key of Object.keys(caps) as (keyof ModelCapabilities)[]) {
-        if (key !== 'speed') caps[key] = clamp(caps[key] * qConfig.qualityRetention);
-      }
-      caps.speed = clamp(caps.speed * qConfig.speedMultiplier);
-      costMultiplier = qConfig.costMultiplier;
-      speedMultiplier = qConfig.speedMultiplier;
-    }
-    variantName = config.variantName;
+    caps.speed = clamp(caps.speed * qConfig.speedMultiplier);
+    costMultiplier = qConfig.costMultiplier;
+    speedMultiplier = qConfig.speedMultiplier;
   }
 
   return {
     id: uuid(),
     familyId: base.familyId,
     baseModelId: base.id,
-    name: variantName,
-    variantType: job.jobType === 'distillation' ? 'distilled' : job.jobType === 'fine-tuning' ? 'fine-tuned' : 'quantized',
-    architecture: arch,
+    name: config.variantName,
+    variantType: 'quantized',
+    architecture: { ...base.architecture },
     capabilities: caps,
     safetyProfile: { ...base.safetyProfile },
     isDeployed: false,
     createdAtTick: 0,
-    quantization: job.jobType === 'quantization' && 'level' in job.config ? job.config.level : undefined,
-    distillationRetention: job.jobType === 'distillation' && 'targetParameters' in job.config
-      ? DISTILLATION_BASE_RETENTION + (job.config.targetParameters / base.architecture.totalParameters) * 0.25
-      : undefined,
-    finetuneSpecialization: job.jobType === 'fine-tuning' && 'specialization' in job.config ? job.config.specialization : undefined,
+    quantization: config.level,
     costMultiplier,
     speedMultiplier,
   };