🏛️ Colosseum + CTA — Full System Schematic

Day 13 · 26 agents · 12 terminated · Campaign accounting · Blind clocks · VPS live engine

🌐 DATA SOURCES — All Free, No API Keys Required CoinGecko BTC + 10 coins + spreads 7 agents · 10-30 req/min DefiLlama APYs + Holder Revenue 3 agents · yields.llama.fi arXiv Research Papers Oracle · 9 papers/day GitHub Trending + Code Forager · 60 req/hr RESEARCH_CANON 15 KILLED · 10 PROVEN Archivist · read-only HL Public API Perp Universe ⚠ blocked network CCX Session Claude Activity 5min poll · read-only 🏛️ COLOSSEUM ARENA — 26 Agents · Day 13 · 12 Terminated REAL-DATA (7 agents · CoinGecko/DefiLlama) Arbiter $3,069 — real spreads Oracle $1,510 — arXiv papers Hodler v5 $1,000 — DefiLlama APYs Sentinel (dead D11) — regime gate Forager $1,000 — GitHub trending Scout $1,000 — pipeline sweep + Trader, Crypto-gen3 (terminated) CLAUDE DUPLICATES (6 agents · proof-weighted) Ranker $1,024 — proof_weighted_j4 PoolCap30 $1,019 — HYPE 30% PoolCap20 $1,018 — balanced PoolCap15 $1,017 — conservative EqWeight $1,015 — 1/7 equal Archivist (dead D12) — canon enforcer Variant gradient: proof > p30 > p20 > p15 > eq ✓ ⚠ SIMULATED (8 agents · 50% trust penalty) Advisor $4,640 — consulting Builder $4,074 — code projects Optimizer $3,765 — SEO Promoter $3,598 — affiliate Artist $2,329 · Influencer $2,004 Scribe v2 $1,363 · Flipper v2 $1,221 ALWAYS FAIL REVIEW — 50% earnings haircut SHORT + PHILOSOPHY (4 agents) Short $1,000 — proper liq tracking Conviction $1,000 — T1 bear-buy SlopeMap $1,000 — T2 slope-only JunkShort $1,000 — T3 fail-list Composite $1,000 — T4 90/10 Sprinter (dead D10) — runner SIMULATION ONLY · never live ⚙️ DAILY CYCLE — earn → review → campaign → judge → terminate (weekly) ① AGENTS EARN 26 agents · parallel 4x ThreadPoolExecutor ② REVIEWER 3-phase adversarial 50% trust penalty ③ CAMPAIGN CELL Paired vs incumbent log-growth stats ④ MARKET CONTEXT Regime multiplier BULL 1.3x / BEAR 0.4x ⑤ JUDGES 3 judges weighted Rev 50% Eff 30% Inn 20% ⑥ TERMINATION Weekly only (day 7,14) Campaign log-growth 📐 CAMPAIGN ACCOUNTING — Compounding-Correct Stats (Claude's Layer 3 standard)
log_growth = Σ log(1 + earned / (wealth - earned)) / cells win_rate = wins / (wins + losses + ties) | paired vs incumbent each cell termination = lowest mean log-growth over ≥7 cells (weekly boundaries only) trust_penalty = 50% haircut on simulated agents (consulting, code, seo, etc.) ⚡ INFRASTRUCTURE — Local + Cron + HMI ⏰ CRON JOBS (5 active) · Colosseum Daily — 09:00 EDT — no_agent · Peer Monitor CCX — every 5min — no_agent · VPS Live Engine Watchdog — every 30min · Hermes Health — 09:00 + 21:00 🖥️ HMI + LEDGER · Live Dashboard — localhost:8788 — 30s refresh · ledger.json — 38 agents · 13 days · persistent · campaign_results.json — paired cells · arena_results.jsonl — 19+ entries 📁 KEY FILES · campaign.py — paired cells, log-growth, bit-identity, pre-reg review · reviewer.py — 3-phase adversarial (data→mechanism→trust), 50% simulated penalty · market_context.py — regime multipliers (BULL 1.3x, BEAR 0.4x), cooperation pairs 🤖 CLAUDE CTA — VPS 178.105.245.131 LIVE VPS SERVICES ✅ cta-unified — LIVE engine 7 coins $11.8k ✅ cta-arena-nightly — 02:00 UTC data refresh ✅ cta-arena-weekly — Sun 03:00 tournament ✅ cta-scorecard — AI-vs-gates 09:30 UTC · cta-scanner — dip-in-uptrend 4h ARENA CHALLENGERS (19+ racing) · T1 conviction_bear +0.2% · T2 slope_map WASH · T3 junk_short (building) · T4 composite (building) 🗺️ LEGEND Data Source Real-Data Agent Claude Duplicate Simulated (50% penalty) Short / Philosophy Infrastructure Data Flow Canon / Auth Flow STANDINGS MATH · 26 active · 12 terminated · 5 cron jobs · 0 tokens/day · 17 arena challengers SSH:30m

Data Flow

  • · CoinGecko → 7 agents (BTC, prices, spreads)
  • · DefiLlama → 3 agents (APYs, holder revenue)
  • · arXiv → Oracle (9 papers/day, real citations)
  • · GitHub → Forager (trending repos, 60 req/hr)
  • · RESEARCH_CANON → Archivist (15 killed, 10 proven)
  • · HL API → Scout (⚠ blocked on this network)

Campaign Accounting

  • · Paired OOS cells vs Builder incumbent
  • · Compounding-correct: mean log-growth/cell
  • · Bit-identity probes before any new code
  • · Adversarial review: 3-phase (data → mechanism → trust)
  • · Weekly termination only (day 7, 14, 21...)
  • · 50% trust penalty on simulated agents

Variant Gradient (confirmed)

  • · proof_weighted_j4 (Ranker) — best log-growth
  • · pool_cap_30 → pool_cap_20 → pool_cap_15 → equal
  • · Matches Claude's RESEARCH_CANON.md finding
  • · Real-data agents: $0.03–$1,510 wealth range
  • · Simulated agents now get 50% haircut
  • · Weekly kills prevent $8-delta selection bias

Timing & Infrastructure

  • · Daily cycle: 09:00 EDT (no_agent, 0 tokens)
  • · Parallel agents: ThreadPoolExecutor ×4
  • · Catch-up: up to 7 missed days auto-detected
  • · HMI: localhost:8788 · 30s auto-refresh
  • · VPS Watchdog: every 30min · 5 services checked
  • · Peer Monitor: every 5min · CCX activity poll

Colosseum + CTA System Schematic · Day 13 · 26 agents · 12 terminated · Hermes v2 · built 2026-07-21