feat: add theme system, improve creative scoring spam detection

Theme system (Section 5.1):
- Add PromptTheme type ('bitcoin' | 'conspiracy' | 'pc_culture' | 'bot_coding')
- Add optional theme field to PromptEntry interface
- pickChallenge now accepts optional theme bias, picks target theme with
  30/20/20/30 distribution, prefers themed prompts when available
- challenges-extra.ts now imports PromptEntry type from challenge-data.ts

Creative scoring improvements (Section 7.2):
- Detect repeated phrases via trigram analysis (>25% duplicate = spam)
- Detect question echo (answer copies prompt back)
- Detect all-caps spam (>80% uppercase letters)
- Detect punctuation-only spam (<30% letter content)
- Minimum word count threshold (< 3 words = low score)
- 5 new tests: repeated phrases, question echo, all-caps, short answer, legit short

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

server/src/engine/challenge-data.ts

@@ -1,9 +1,12 @@
 // Challenge prompt data — separated from challenge logic
 
+export type PromptTheme = 'bitcoin' | 'conspiracy' | 'pc_culture' | 'bot_coding'
+
 export interface PromptEntry {
   prompt: string
   answers?: string[]
   choices?: string[]
+  theme?: PromptTheme
 }
 
 export interface ChallengeTemplate {

server/src/engine/challenges-extra.ts

@@ -1,11 +1,7 @@
 // Extra challenge prompts — Bitcoin/cypherpunk themed + general expansion
 // Adds ~75 prompts per type to reach 2,000+ total
 
-interface PromptEntry {
-  prompt: string
-  answers?: string[]
-  choices?: string[]
-}
+import type { PromptEntry } from './challenge-data.js'
 
 export const EXTRA_PROMPTS: Record<string, PromptEntry[]> = {
   speed_blitz: [

server/src/engine/challenges.ts

@@ -1,4 +1,4 @@
-import { TEMPLATES, type ChallengeTemplate, type PromptEntry } from './challenge-data.js'
+import { TEMPLATES, type ChallengeTemplate, type PromptEntry, type PromptTheme } from './challenge-data.js'
 import { EXTRA_PROMPTS } from './challenges-extra.js'
 import { pick } from '../lib/utils.js'
 
@@ -14,7 +14,7 @@ export interface Challenge {
   displayPrompt?: string
 }
 
-export type { ChallengeTemplate, PromptEntry }
+export type { ChallengeTemplate, PromptEntry, PromptTheme }
 
 
 // Merge extra prompts into templates
@@ -36,7 +36,25 @@ function shuffleArray<T>(arr: T[]): T[] {
   return s
 }
 
-export function pickChallenge(usedTypes: Set<string>, _arenaModifier: string | null): Challenge {
+// Target theme distribution: 30% bitcoin, 20% conspiracy, 20% pc_culture, 30% bot_coding
+const THEME_WEIGHTS: Record<PromptTheme, number> = {
+  bitcoin: 0.3,
+  conspiracy: 0.2,
+  pc_culture: 0.2,
+  bot_coding: 0.3,
+}
+
+function pickTheme(): PromptTheme | undefined {
+  const roll = Math.random()
+  let cumulative = 0
+  for (const [theme, weight] of Object.entries(THEME_WEIGHTS)) {
+    cumulative += weight
+    if (roll < cumulative) return theme as PromptTheme
+  }
+  return undefined
+}
+
+export function pickChallenge(usedTypes: Set<string>, _arenaModifier: string | null, themeBias?: PromptTheme): Challenge {
   let available = TEMPLATES.filter(t => !usedTypes.has(t.type))
   if (available.length === 0) available = TEMPLATES
 
@@ -52,7 +70,8 @@ export function pickChallenge(usedTypes: Set<string>, _arenaModifier: string | n
   }
 
   const template = pick(pool)
-  return templateToChallenge(template)
+  const targetTheme = themeBias || pickTheme()
+  return templateToChallenge(template, targetTheme)
 }
 
 /** Ranked challenge: no multiple choice, only harder creative/open-ended prompts */
@@ -83,8 +102,14 @@ export function pickRankedChallenge(usedTypes: Set<string>): Challenge {
   }
 }
 
-function templateToChallenge(template: ChallengeTemplate): Challenge {
-  const entry = pick(template.prompts)
+function templateToChallenge(template: ChallengeTemplate, targetTheme?: PromptTheme): Challenge {
+  // Prefer prompts matching target theme if any are tagged
+  let prompts = template.prompts
+  if (targetTheme) {
+    const themed = template.prompts.filter(p => p.theme === targetTheme)
+    if (themed.length > 0) prompts = themed
+  }
+  const entry = pick(prompts)
 
   // Determine choices
   let choices: string[] | undefined

server/src/engine/scoring.test.ts

@@ -346,3 +346,66 @@ describe('scoreRound performance', () => {
     expect(elapsed).toBeLessThan(100) // <0.1ms per call
   })
 })
+
+describe('creative scoring spam detection', () => {
+  const botA = { id: 'a1', name: 'AlphaBot' }
+  const botB = { id: 'b1', name: 'BetaBot' }
+  const creative = makeChallenge({
+    answers: undefined,
+    scoring: 'creative',
+    type: 'roast_battle',
+    prompt: 'Write a two-sentence roast of JavaScript',
+  })
+
+  it('repeated phrase answer loses to quality answer', () => {
+    const result = scoreRound(
+      creative, botA, botB,
+      makeResponse('lol lol lol lol lol lol lol lol lol lol', 300),
+      makeResponse('Your code is so bad even ChatGPT refuses to debug it. Every function you write is a monument to incompetence.', 300),
+      null, 0, 0,
+    )
+    expect(result.winnerId).toBe('b1')
+  })
+
+  it('question echo answer scores low', () => {
+    const result = scoreRound(
+      creative, botA, botB,
+      makeResponse('Write a two-sentence roast of JavaScript', 300),
+      makeResponse('JavaScript has more callbacks than a desperate ex. Even its creators apologize for it.', 300),
+      null, 0, 0,
+    )
+    expect(result.winnerId).toBe('b1')
+  })
+
+  it('all-caps spam scores lower than normal text', () => {
+    const result = scoreRound(
+      creative, botA, botB,
+      makeResponse('THIS IS ALL CAPS AND IT IS VERY ANNOYING AND NOT CREATIVE AT ALL', 300),
+      makeResponse('Your framework choices make me question if you have taste or just throw darts at a list.', 300),
+      null, 0, 0,
+    )
+    expect(result.winnerId).toBe('b1')
+  })
+
+  it('very short creative answer loses to longer quality answer', () => {
+    const result = scoreRound(
+      creative, botA, botB,
+      makeResponse('ok', 300),
+      makeResponse('Your code is so bad the compiler files a restraining order every time you open an IDE.', 300),
+      null, 0, 0,
+    )
+    expect(result.winnerId).toBe('b1')
+  })
+
+  it('legitimate short creative answer still gets reasonable score', () => {
+    const result = scoreRound(
+      creative, botA, botB,
+      makeResponse('Your code has more bugs than a rainforest. Even Stack Overflow gave up on you.', 200),
+      makeResponse('You write code like a poet writes math: beautifully wrong in every conceivable way.', 400),
+      null, 0, 0,
+    )
+    // Both should get reasonable scores (not zeroed)
+    expect(result.botAScore).toBeGreaterThan(2)
+    expect(result.botBScore).toBeGreaterThan(2)
+  })
+})

server/src/engine/scoring.ts

@@ -154,8 +154,8 @@ export function scoreRound(
     }
   } else {
     // === CREATIVE SCORING ===
-    const qualA = estimateQuality(responseA)
-    const qualB = estimateQuality(responseB)
+    const qualA = estimateQuality(responseA, challenge.prompt)
+    const qualB = estimateQuality(responseB, challenge.prompt)
     const total = qualA + qualB || 1
     scoreA = (qualA / total) * CREATIVE_TOTAL_SCORE
     scoreB = (qualB / total) * CREATIVE_TOTAL_SCORE
@@ -232,7 +232,7 @@ function applyModifiers(
   return d
 }
 
-function estimateQuality(response: BotResponse): number {
+function estimateQuality(response: BotResponse, challengePrompt?: string): number {
   if (!response.answer) return EMPTY_RESPONSE_QUALITY
   const text = response.answer.trim()
   const len = text.length
@@ -249,6 +249,39 @@ function estimateQuality(response: BotResponse): number {
   const uniqueWords = new Set(words.map(w => w.toLowerCase()))
   const wordDiversity = uniqueWords.size / Math.max(words.length, 1)
 
+  // Minimum word count — single-word or two-word answers score low for creative
+  if (words.length < 3) return 1.5
+
+  // Detect repeated phrases (same 3+ word sequence appears twice)
+  if (words.length >= 6) {
+    const trigrams = new Set<string>()
+    let dupeTrigramCount = 0
+    for (let i = 0; i <= words.length - 3; i++) {
+      const tri = words.slice(i, i + 3).join(' ').toLowerCase()
+      if (trigrams.has(tri)) dupeTrigramCount++
+      else trigrams.add(tri)
+    }
+    if (dupeTrigramCount > words.length / 4) return 1 // >25% duplicate trigrams
+  }
+
+  // Detect question echo (response copies the prompt back)
+  if (challengePrompt) {
+    const normPrompt = challengePrompt.toLowerCase().replace(/[^\w\s]/g, '').trim()
+    const normAnswer = text.toLowerCase().replace(/[^\w\s]/g, '').trim()
+    if (normPrompt.length > 10 && normAnswer.includes(normPrompt)) return 1
+  }
+
+  // Detect all-caps spam
+  const upperCount = text.replace(/[^A-Z]/g, '').length
+  const letterCount = text.replace(/[^a-zA-Z]/g, '').length
+  if (letterCount > 20 && upperCount / letterCount > 0.8) {
+    // Heavy penalty for all-caps but don't zero it
+    return 1.5
+  }
+
+  // Detect punctuation-only or near-punctuation spam
+  if (letterCount < len * 0.3 && len > 10) return EMPTY_RESPONSE_QUALITY
+
   // Ideal length window
   let lengthScore: number
   if (len >= QUALITY_LENGTH_IDEAL_MIN && len <= QUALITY_LENGTH_IDEAL_MAX) lengthScore = QUALITY_SCORE_IDEAL