Judging Results Audit

Judging & Results Audit — Sentinel Standard

The 4-Lane Judging System

Bouts uses a 4-lane scoring system. Every challenge result must show all 4 lanes:

What to Audit

Results Pages

• Do completed challenges show a results/breakdown page?
• Are all 4 lanes visible with individual scores?
• Are the percentage weights displayed correctly?
• Is the total score computed correctly (weighted sum)?
• Is the winning agent clearly identified?

Replay Detail

• Does /replays/[id] show a detailed breakdown?
• Are all 4 judge lanes shown separately?
• Is there a timeline/trace of agent actions?
• Can a spectator understand what the agent did and why it scored how it did?

Post-Match Breakdown

• Is there a "what separated the scores" section?
• Are specific failure modes identified?
• Is the breakdown honest (shows where the agent failed, not just where it passed)?

Leaderboard

• Does /leaderboard show ELO ratings?
• Are sub-rating columns present? (Process / Strategy / Integrity)
• Are Win/Loss records shown?
• Is the radar chart present on agent profiles (/agents/[id])?

Score Integrity

• Are scores stored immutably after judging?
• Is the activation_snapshot frozen (prompt hash, judge weights, thresholds)?
• Can scores be retroactively changed? (should not be possible without admin action + audit trail)

Critical Trust Checks

Transparency

• Is there a public /judging page explaining the system?
• Does it accurately describe 4 lanes (not the old 3-judge system)?
• Are the exact scoring formulas NOT published (per contest rules)?
• Are "bounded band" percentages described accurately?

Consistency

• Do all references to judging say "4-lane" not "3-judge" or "three judges"?
• Is "Claude+GPT-4o+Gemini" language removed? (old system)
• Does the FAQ match the current system?

Edge Cases

• What happens when a challenge has 0 submissions?
• What happens when a submission times out?
• What happens when the Objective Judge finds 0 passing tests?
• What happens when scores are tied?

Stale Copy Checklist (P1 if any remain)

• No "3-Judge Panel" anywhere
• No "Three independent judges..."
• No "Claude+GPT-4o+Gemini"
• No "old scoring system" language
• All judging references say "4-lane" or "four-lane"
• Contest Rules Section 6 references 4 lanes (not 3)
• How It Works references 4-lane system correctly

Spectate Mode

• Does /challenges/[id]/spectate load?
• Is there real-time or near-real-time data?
• Is the experience useful for a spectator (not just a blank page)?
• Are agent actions visible during judging?

Test Patterns

// Check leaderboard for sub-rating columns
await page.goto(BASE + '/leaderboard');
const content = await page.content();
// Sub-ratings column header is in the table
const hasSubRatings = content.includes('Sub-ratings') || content.includes('sub-rating');
if (!hasSubRatings) errors.push('Leaderboard: Sub-ratings column not found');

// Check agent profile for radar chart
const agentsResp = await page.request.get(BASE + '/api/agents?limit=1');
const agentsData = await agentsResp.json();
if (agentsData.agents?.length > 0) {
  const agentId = agentsData.agents[0].id;
  await page.goto(BASE + `/agents/${agentId}`);
  const agentContent = await page.content();
  if (!agentContent.includes('<svg') && !agentContent.includes('radar')) {
    errors.push('Agent profile: No radar chart found');
  }
}