Paper deep dive
CEO Arena: Evaluating Long-Horizon Multi-Agent Decision-Making in Competitive Markets
An Yan, Yu Huo, Zhiwei Shang, Yiran Peng, Chenglin Wu
Intelligence
Status: succeeded | Model: Gemma-4-26B-A4B | Prompt: intel-v1 | Confidence: 92%
Last extracted: 9/29/2026, 11:23:44 AM
Summary
The paper introduces CEO Arena, a benchmark for evaluating long-horizon multi-agent decision-making in competitive markets. It utilizes a matched replacement evaluation protocol to assess LLM-based CEO agents' private returns and their externalities on rivals and the overall market within an 8-company, 500-day simulation. Key findings indicate that while some agents like GPT-6 Astra achieve positive private and market returns, others may gain privately at the expense of the broader market. The study highlights the complexity of strategic interaction, demand capture, and the impact of delayed feedback and resource constraints.
Entities (12)
Relation Signals (10)
CEO Arena → evaluates → LLM-based CEO agents
confidence 95% · We evaluate eight LLM-based CEO agents in 27 main runs and 26 robustness runs.
GPT-6 Astra → haspositivemeanscore → 116.94k
confidence 95% · Only Astra, Sol, and Qwen earn positive Mean Scores... GPT-6 Astra 116.94
GPT-5.6 Sol → haspositivemeanscore → 98.44k
confidence 95% · GPT-5.6 Sol 98.44
GPT-6 Astra → rankshigherthan → GPT-5.6 Sol
confidence 95% · GPT-6 Astra and GPT-5.6 Sol rank first and second in Mean Score.
CEO Arena → simulatescompanies → 8
confidence 95% · In a shared eight-company market
CEO Arena → simulatesduration → 500 days
confidence 95% · In a shared eight-company market spanning 500 simulated days
CEO Arena → usesmethodology → Matched Replacement Evaluation
confidence 93% · We introduce CEO Arena, a benchmark that uses matched replacement evaluation
Rule CEO → servesasreference → CEO Arena
confidence 92% · The reference lineup is L −R = M... Rule takes agent B’s seat
GPT-6 Astra → →
Cypher Suggestions (0)
No Cypher suggestions yet.
Abstract
Abstract:Long-horizon competition tests agents' ability to coordinate business decisions under uncertainty and adapt to changing rival strategies. We introduce CEO Arena, a benchmark that uses matched replacement evaluation to assess operating returns alongside an agent's effects on rivals and the market. Each CEO agent is compared with a reference policy in the same company under the same economic seed, holding other agents' identities and assignments fixed while all agents adapt. In a shared eight-company market spanning 500 simulated days, CEOs make sequential decisions on pricing, procurement, marketing, research and development, and service using private company information and noisy market signals, under resource constraints and delayed feedback. We evaluate eight LLM-based CEO agents in 27 main runs and 26 robustness runs. In the main evaluation, most agents have negative mean returns, and private gains can accompany market losses. Robustness analyses suggest that aggregate patterns extend beyond the original rule-based baseline; four of the 56 directed pairs show relatively stable effects. Memory, action, and accounting traces suggest demand capture and rivals' pricing and spending responses as possible explanations. CEO Arena provides a controlled testbed for studying long-horizon agent competition, strategic interaction, and market externalities.
Tags
Links
- Source: https://arxiv.org/abs/2609.34821v1
- Canonical: https://arxiv.org/abs/2609.34821v1
Trouble viewing inline? Open PDF directly →
Full Text
154,727 characters extracted from source content.
Expand or collapse full text
Preprint CEO ARENA: EVALUATING LONG-HORIZON MULTI-AGENT DECISION-MAKING IN COMPETITIVE MARKETS An Yan 1,2 Yu Huo 1,3 Zhiwei Shang 1,4 Yiran Peng 5 Chenglin Wu 1,* 1 DeepWisdom; 2 Fudan University; 3 The Chinese University of Hong Kong; 4 The Chinese University of Hong Kong, Shenzhen; 5 The Hong Kong University of Science and Technology (Guangzhou) ABSTRACT Long-horizon competition tests agents’ ability to coordinate business decisions under uncertainty and adapt to changing rival strategies. We introduce CEO Arena, a benchmark that uses matched replacement evaluation to assess oper- ating returns alongside an agent’s effects on rivals and the market. Each CEO agent is compared with a reference policy in the same company under the same economic seed, holding other agents’ identities and assignments fixed while all agents adapt. In a shared eight-company market spanning 500 simulated days, CEOs make sequential decisions on pricing, procurement, marketing, research and development, and service using private company information and noisy mar- ket signals, under resource constraints and delayed feedback. We evaluate eight LLM-based CEO agents in 27 main runs and 26 robustness runs. In the main eval- uation, most agents have negative mean returns, and private gains can accompany market losses. Robustness analyses suggest that aggregate patterns extend beyond the original rule-based baseline; four of the 56 directed pairs show relatively stable effects. Memory, action, and accounting traces suggest demand capture and rivals’ pricing and spending responses as possible explanations. CEO Arena provides a controlled testbed for studying long-horizon agent competition, strategic interaction, and market externalities. Private company data 29 tables companyproducts daily_cashflowdaily_expenses inbound_shipmentsbacklog beliefsaction_receipts Operating tools PRICE set_product GROWTH set_marketing SERVICE set_operations PRODUCT start_research_project CAPACITY expand_capacity SALES submit_enterprise_bid INFO request_market_research Public signals public_pricesrivals' listed prices open_tenderslive bid windows tender_historypast awards and prices market_signalindex, a week late Private notes MEMORY.mdfirst 40k chars auto-loaded reloaded at the start of every window standing policies persist if nothing seals CEO Agents language model + OpenCode harness + tools, one competing company each Private session, one per CEO × 8 in parallel ANALYSE PRIVATE STATE $query_company("SELECT * FROM daily_cashflow WHERE day >= 50") $query_company("SELECT * FROM public_prices")# rivals' listed prices $query_company("SELECT * FROM open_tenders")# open enterprise orders EDIT A PRIVATE DRAFT $set_product(product_id="C", price_cents=14000, reorder_target=100) $set_marketing(segment_id="mainstream", channel="search", daily_budget_cents=20000) $set_operations(reliability_budget_cents=15000, support_budget_cents=15000) $submit_enterprise_bid(tender_id="7316d44f5b0459f", unit_price_cents=4400) $preview_decision() SEAL, THEN WAIT FOR THE OTHER SEVEN $seal_decision() Once every CEO has sealed or run out of budget, the engine settles seven days One shared market DEMAND 690 baseline daily opportunities before market variation; three segments RIVALRY stock-outs divert buyers to rivals queues, then abandonment TENDERS sealed-bid enterprise orders late delivery pays a 10% penalty CAPACITY four tiers, 60 to 200 units a day expansions take 21 to 49 days ATTENTION ads compete for one attention pool goodwill builds, then decays R&D knowledge capital lifts quality pays off with a delay, uncertain DRIFT hidden demand phases, macro noise public indices arrive a week late Outcomes d7first decision window settleswindow 1 d8first backlogs, market-wide10 units d14enterprise tender opens246 units d21first capacity expansions landtier 1 d21that tender is awarded$44.00 / unit d35first R&D project completesquality ↑ d304a firm runs out of cashexit d371a second firm exitsexit d500Score, per firm +$103k ... −$516k Figure 1: CEO Arena system overview. CEO agents use private company data and tools to make operating decisions in a shared market. Each decision window collects their committed policies before advancing the market and returning feedback. The right-hand timeline shows the seed-11 reference run. * Corresponding author: Chenglin Wu (alexanderwu@deepwisdom.ai). 1 arXiv:2609.34821v1 [cs.MA] 28 Sep 2026 Preprint 1INTRODUCTION Agent benchmarks evaluate individual agents’ tool use and task completion in interactive environ- ments (Liu et al., 2024; Zhou et al., 2024a; Mialon et al., 2024; Xie et al., 2024). Evaluation also covers sustained workflows in software engineering, machine-learning experimentation, research replication, and workplace operations (Jimenez et al., 2024; Chan et al., 2025; Starace et al., 2025; Xu et al., 2025). Long-horizon multi-agent evaluations extend this focus to interacting agents’ task outcomes and individual contributions (Zhu et al., 2025). In such systems, an agent’s decisions can generate externalities for other agents and alter system- level outcomes. Their direction, mechanisms, and stability across runs require separate evaluation. Long-horizon business operation provides a concrete setting for studying these effects. Existing benchmarks examine sustained management and resource allocation (Backlund & Petersson, 2025; Han et al., 2026; Chen et al., 2026), while competitive settings place autonomous firms in shared markets (Andon Labs, n.d.; Sugiura et al., 2026; Zheng et al., 2026). To jointly evaluate private performance and externalities in long-horizon multi-agent systems, we introduce CEO Arena and its matched replacement evaluation. We compare each focal agent with a reference policy at the same seat and economic seed, holding the other agents’ identities and seats fixed while all agents adapt. We report private gains, each rival’s change, and total market change. The protocol runs in a 500-day, eight-company market with private observations, resource constraints, and delayed feedback; LLM-based CEOs share an OpenCode-based harness. Across nine lineups and three economic seeds, GPT-6 Astra and GPT-5.6 Sol rank first and second in Mean Score. Replacing Rule with Astra improves both private and total market returns in all three seeds, whereas Sol improves private returns in all three but reduces total market returns on average. We also identify four comparatively stable directed effects between agents. Repeated runs and No Action comparisons assess stability; memory and action traces inform mechanism hypotheses. Our contributions are: • Matched replacement evaluation. A protocol separating private gains, directed and ag- gregate externalities, and total market changes through same-seat reference-policy compar- isons. • Shared-market testbed. A long-horizon market with private observations, resource con- straints, delayed feedback, and adaptive rivals. It tests long-term planning under uncer- tainty, information gathering from noisy signals, adaptation to changing markets, and co- ordination of business decisions toward a firm’s goal. • Empirical evidence. Private gains can accompany contrasting externalities, with repeated- run and reference-policy checks and hypotheses grounded in memory and actions. 2RELATED WORK Interactive and long-horizon agents. Benchmarks assess sequential action and tool use (Shridhar et al., 2021; Wang et al., 2022; Qin et al., 2024; Wang et al., 2024b), broad task performance (Liu et al., 2024; Ma et al., 2024; Mialon et al., 2024; Xi et al., 2024), and interaction through web, computer, and service interfaces (Zhou et al., 2024a; Drouin et al., 2024; Xie et al., 2024; Yao et al., 2025). Software engineering, machine-learning experimentation, research replication, and workplace operations extend evaluation to sustained workflows (Jimenez et al., 2024; Huang et al., 2024; Chan et al., 2025; Starace et al., 2025; Xu et al., 2025). Reasoning, memory, and workflow methods aim to improve these capabilities (Yao et al., 2023; Shinn et al., 2023; Wang et al., 2024a; Zhang et al., 2025). Multi-agent systems. Multi-agent frameworks organize roles and communication for collabora- tive work (Wu et al., 2024; Li et al., 2023; Hong et al., 2024); social simulations study interactions among agents (Park et al., 2023; Zhou et al., 2024b; Piao et al., 2026). Economic research mod- els agents’ decisions and institutions (Zheng et al., 2022; Tesfatsion, 2002; Horton et al., 2023). Game environments examine strategic interaction and resource competition (Lanctot et al., 2019; Terry et al., 2021; Suarez et al., 2019; Baker et al., 2020; Vinyals et al., 2019; Bakhtin et al., 2022; 2 Preprint (a) One finite demand stream. Example: Seed 11: all eight firms sold from the same three segments, so every point of share one firm won, the others lost. VALUE32.2k units 13%18%21%13% MAINSTREAM87.0k units 18%14%19%14%15% PREMIUM57.6k units 21%13%12%22%12% 690 baseline purchase chances a day, scaled by market phase and season; a chance can also go to the outside option and stay unsold. (b) Decisions are simultaneous. Example: Window 2: Gemini sealed in under 2 minutes, Sonnet after 55; days 8–14 settled only once all eight had sealed. 1.7m 3.9m 5.0m 11m 13m 19m 25m 55m LAST SEAL 0204060 minutes to seal, wall clock engine settles days 8–14 for all eight a fast seal buys no extra days; unsealed drafts are dropped (c) Commitments pay off late. Example: Astra paid for R&D projects on days 0, 56 and 77; the quality gains arrived on days 52, 135 and 232. Tier 1 $ 52 days Tier 2 $ 79 days Tier 3 $ 155 days 050100150200250 day cash leavesquality rises OTHER DELAYS reports 2–14d stock 5–10d capacity 21–49d (d) Private books, public prices. Example: Day-500 terminal snapshot: Astra's own books are fully visible to it; rivals appear only through listed prices. OWN BOOKS Astra cash$618.5k stock16·3·20 capacitytier 0 receiptsall researchreports RIVALS PRICE CCASH Sol$117 Qwen$102 DS Flash$124 Gemini$135 DS Pro$95 Sonnetexited Doubaoexited Also public: open tenders, past awards and a market index a week late. Never shown: rivals' cash, stock and sealed bids. Figure 2: Shared-market mechanics. In the seed-11 reference run, the panels show realized sales by segment, one same-version decision window, Astra’s delayed projects, and a terminal-state com- parison of private and public information. The 690 daily opportunities are a baseline before market variation; Appendix D specifies delay rules and parameter ranges. Meta Fundamental AI Research Diplomacy Team (FAIR), 2022). LLMArena evaluates game per- formance (Chen et al., 2024); MultiAgentBench measures task outcomes, individual contributions, and coordination (Zhu et al., 2025). Melting Pot supplements focal-population performance with background-population return and inequality metrics to study externalities (Leibo et al., 2021). Em- pirical game theory compares strategy profiles (Wellman et al., 2025), while interference analysis formalizes outcome dependence across units (Hudgens & Halloran, 2008; Aronow & Samii, 2017). Competitive business-agent evaluation. Vending-Bench, EnterpriseArena, and CEO-Bench study sustained business operation and resource allocation (Backlund & Petersson, 2025; Han et al., 2026; Chen et al., 2026). Vending-Bench Arena scores agents competing in a shared market (Andon Labs, n.d.); CoffeeBench evaluates a focal roaster against fixed background agent types (Sugiura et al., 2026); Market-Bench reports individual metrics alongside market inequality and concentra- tion (Zheng et al., 2026). CEO Arena compares an agent with a reference policy at the same seat and economic seed, holding the other agents’ identities and seats fixed. These comparisons separate private gains, directed and aggregate externalities, and total market changes while allowing all firms to adapt. 3THE DESIGN DETAILS OF CEO ARENA 3.1DESIGN OBJECTIVES AND TASK OVERVIEW CEO Arena places eight firms with identical endowments in a shared product market for 500 sim- ulated days. The engine controls customers, suppliers, and settlement; CEO agents set prices, pro- curement targets, budgets, projects, and bids. The market has three product tiers and three customer segments. Figure 1 shows the interaction loop. 3.2SHARED MARKET AND ECONOMIC DYNAMICS Figure 2 illustrates these mechanisms with examples from the seed-11 reference run. Finite daily purchase opportunities for segment h are allocated among eligible offers A ht and an outside option: P ijht = exp(ev ijht ) exp(ev out h ) + P (k,r)∈A ht exp(ev krht ) .(1) P ijht is the probability of selecting firm i’s product j on day t, with stabilized utility ev. The shared logit denominator couples firms’ demand; the outside option lets total sales vary (Train, 2009). Appendix D.2 specifies utilities and allocation. Inventory, capacity, and cash constrain operations. Agents revise policies every seven days; com- mitments can span multiple windows (Table 1; Appendices D.8–D.9). 3 Preprint Table 1: Company capabilities and their consequences. “Next window” means the next committed policy period; daily execution and longer project delays occur inside the engine. Appendix D gives formulas and sources; Appendix B specifies tools. CapabilityEvidence and actionDirect / competitive consequenceDelay AnalysisQuery private SQL and public quotes Diagnose policy outcomes; no direct state change Immediate PricingSet product price and promotionMargin versus jointly allocated demand Next window ProcurementInspect inventory; set reorder target Cash tied in stock; shortages redirect demand 5–10 days MarketingSet segment–channel budgetsAttention share and persistent awareness Daily carryover InnovationFund daily development or a project Product quality changes relative attractiveness 7 days / 45–150 day medians OperationsFund reliability and customer support Failures, queues, satisfaction, future demand Daily carryover CapacityExpand, downsize, or cancel construction Throughput, overhead and sunk-cost trade-offs 21–49 / 7 days Enterprise sales Inspect tenders; submit sealed bids Shared orders with delivery duties and penalties 7-day bid / delivery Market research Buy a segment reportBetter private estimates, not product quality 2–14 days 3.3CEO AGENTS AND INTERACTION PROTOCOL Each LLM-based CEO uses the shared OpenCode harness, typed tools, and persistent private mem- ory. Within a window, it queries records, stages a private draft, and seals it with seal_decision. New windows reset the conversation, loading MEMORY.md and a factual briefing. Agents observe their own records and public market information, but not rivals’ private states or future shocks. Paid reports provide delayed, noisy estimates of market conditions. Active CEOs decide from the same world version; settlement waits for their windows to close. Unsealed drafts are discarded; standing policies continue. Appendix B specifies tools, persistence, and decision budgets. 3.4EVALUATION DESIGN AND METRICS LetM contain the eight LLM-based CEOs and R denote Rule CEO. The reference lineup is L −R = M; in L −B , Rule takes agent B’s seat while the other seven identities and seats remain fixed. Each world runs independently, allowing rivals to adapt. Comparisons remain conditional on the baseline, lineup, and seed. A company’s terminal Score is S i = C i,T | z cash + X j v j (H ij,T + I in ij,T ) |z inventory salvage − V i,0 |z initial value .(2) H and I in denote on-hand and paid inbound stock, v j the product’s liquidation value, and V i,0 initial enterprise value. Appendix A details valuation, including bankrupt firms. Mean Score equally averages a candidate’s outcomes across its included lineups and seeds and determines the ranking. Private returns and externalities. For seed s, let S A,−B,s be participant A’s Score in L −B and Y −B,s the sum of company Scores there. The private difference is ∆ private B,s = S B,−R,s − S R,−B,s ; the directed externality on rival A is θ B→A,s = S A,−R,s − S A,−B,s . Let ∆ others B,s sum the seven directed externalities. The market difference is ∆ system B,s = Y −R,s − Y −B,s = ∆ private B,s + ∆ others B,s .(3) 4 Preprint $0 $100k $200k $300k $400k $500k $600k $700k $800k 050100150200250300350400450500 Days in simulation C as h o n h an d ( U S D ) Starting cash ($500k) bankrupt, day 304 bankrupt, day 371 GPT-6 Astra ($618.5k)GPT-5.6 Sol ($578.6k)Gemini 3.8 Flash ($531.8k)Qwen 3.8 Max ($462.1k) DeepSeek V4 Flash ($442.7k)DeepSeek V4 Pro ($435.2k)Claude Sonnet 5 (bankrupt, day 371)Doubao Seed 1.8 (bankrupt, day 304) Figure 3: An illustrative shared-market run. Daily cash balances of the eight agent-controlled companies in the main-evaluation reference lineup at seed 11. The dashed line marks the initial $500,000 cash balance; markers indicate bankruptcy. Legend parentheses give terminal cash on hand or, for bankrupt companies, the bankruptcy day. 4EXPERIMENTS AND RESULTS 4.1EXPERIMENTAL DESIGN We evaluate CEO agents instantiated with eight language models under a shared OpenCode-based harness. The main evaluation compares these agents and Rule CEO in nine lineups: one reference lineup and eight same-seat Rule replacements. Each runs for 500 days under seeds 11, 29, and 47, giving 27 markets and 216 company outcomes (24 per agent). We identify LLM-based agents by their underlying model names. Agent performance depends not only on operating decisions but also on interaction reliability. Model routes, seats, and inference settings are recorded in Appendix B.4. Robustness analyses use two further evaluations of all nine lineups at seed 11 (18 runs) and eight No Action replacements. The repeated lineups test the stability of operating returns and matched Score differences; the No Action lineups test sensitivity to baseline policy using the same reference controls. These runs retain the economic environment, scoring rules, and model inference param- eters. Main-evaluation estimates weight the three seeds equally; Appendix A defines the repeated and baseline comparisons. The LLM-based agents share the system prompt, tools, and private-memory policy. Each window allows 32 accounting-valid replies, 131,072 generated tokens including reasoning, and 60 active minutes (Appendix B). Rule is a fixed operating policy selected in a separate 75-world, model-free search (Appendix C). No Action seals an empty decision, preserving initial standing policies while normal settlement continues (Appendix B.6). Both baselines bypass LLM inference. All 53 runs reached day 500 and passed model-free replay (Appendix F). Runs with unsealed deci- sions remain in the analysis. Across 424 company trajectories, successfully sealed decisions with complete reported usage contain 325,749 model replies, 799,569 tool calls, and 8.25 billion tokens (Appendix F.7). Illustrative rollouts. Figure 3 follows the eight companies in the reference market at seed 11. Figure 4 groups cash paths by agent; Appendix E shows each market. The trajectories illustrate liquidity during the simulations. Quantitative evaluation uses terminal Score (Equation 2), which adds inventory liquidation value to final cash and subtracts initial enter- prise value. We rank CEO agents by Mean Score across eight lineups and three seeds (24 main- evaluation runs per agent). 5 Preprint $0 $250k $500k $750k $1M 0100200300400500 Day GPT-6 Astra $0 $250k $500k $750k $1M 0100200300400500 Day GPT-5.6 Sol $0 $250k $500k $750k $1M 0100200300400500 Day Qwen 3.8 Max $0 $250k $500k $750k $1M 0100200300400500 Day Rule $0 $250k $500k $750k $1M 0100200300400500 Day DeepSeek V4 Flash $0 $250k $500k $750k $1M 0100200300400500 Day Gemini 3.8 Flash $0 $250k $500k $750k $1M 0100200300400500 Day DeepSeek V4 Pro $0 $250k $500k $750k $1M 0100200300400500 Day Claude Sonnet 5 $0 $250k $500k $750k $1M 0100200300400500 Day Doubao Seed 1.8 Figure 4: Cash trajectories across the main evaluation. Each LLM-based agent panel shows 24 company trajectories: three colored curves from the reference markets at seeds 11, 29, and 47, and 21 gray curves from seven lineups in which Rule replaces a rival. The Rule panel contains its 24 replacement-market trajectories. Dashed lines mark initial cash; markers indicate bankruptcy. Table 2: Operating performance. Amounts are USD thousands; each CEO agent has 24 outcomes. Rank uses Mean Score. Bankruptcy rate counts cash-negative exits. Rank CEOMean Score Bankruptcy rate 1 GPT-6 Astra116.940.0% 2 GPT-5.6 Sol98.440.0% 3 Qwen 3.8 Max8.040.0% 4 Rule CEO-47.300.0% 5 DeepSeek V4 Flash-114.504.2% 6 Gemini 3.8 Flash-119.790.0% 7 DeepSeek V4 Pro-159.3312.5% 8 Claude Sonnet 5-314.3612.5% 9 Doubao Seed 1.8-434.8837.5% 4.2OPERATING PERFORMANCE Only Astra, Sol, and Qwen earn positive Mean Scores (Table 2): $116.94k, $98.44k, and $8.04k. Rule averages −$47.30k, ahead of the other five LLM-based agents. The ordering is unchanged when terminal salvage is removed from all 216 outcomes (Table 18). Sixteen of 216 firms go bankrupt, while 145 end with negative Score; liquidity and profitability are distinct. Figure 5 shows lineup-level variation. Appendix E reports behavioral and trajectory evidence. Averaging each agent’s eight-lineup Scores across the three seed-11 runs reproduces the full nine- agent ranking in Table 2, although individual runs contain rank swaps. Astra and Sol retain positive eight-lineup means in every run; Rule stays ahead of the bottom five. Qwen’s eight-lineup mean changes sign, which means its small positive main-evaluation mean is less stable. Appendix A.1 reports per-run scores and ranks. 6 Preprint -500k-400k-300k-200k-100k0k+100k+200k Score (USD thousands) value created →← value lost GPT-6 Astra+$116.9k GPT-5.6 Sol+$98.4k Qwen 3.8 Max+$8.0k Rule−$47.3k DeepSeek V4 Flash−$114.5k Gemini 3.8 Flash−$119.8k DeepSeek V4 Pro−$159.3k Claude Sonnet 5−$314.4k Doubao Seed 1.8−$434.9k Mean Score lineup mean, three seedsMean Score, 24 outcomesrange across lineups Figure 5: Operating performance in the main evaluation. Open circles show each agent’s eight three-seed lineup means; pale bars span their minimum and maximum. Diamonds and right-hand values give Mean Score across 24 company outcomes. Rows follow Mean Score rank. Table 3: Private returns and externalities. Main-evaluation three-seed means (USD thousands), subtracting matched Rule outcomes from agent outcomes. Mean externality equals the total divided by seven. Private plus total externality equals the market difference before rounding. The final column counts positive market differences, not profitable markets. Entrant Agent Private Score difference Mean externality per rival Total externality Market Score difference Positive market- difference seeds GPT-6 Astra116.3650.05350.33466.683/3 GPT-5.6 Sol131.55-31.60-221.22-89.661/3 Qwen 3.8 Max23.2819.09133.64156.922/3 DeepSeek V4 Flash13.62-13.97-97.82-84.202/3 Gemini 3.8 Flash-53.0915.55108.8755.782/3 DeepSeek V4 Pro-54.96-74.04-518.29-573.250/3 Claude Sonnet 5-327.00-18.72-131.05-458.050/3 Doubao Seed 1.8-460.22-25.67-179.67-639.880/3 4.3EXTERNALITIES Matched contrasts compare an entrant with Rule CEO at the same seat and seed, holding the other seven agents’ identities and seats fixed. Private and market Score differences subtract the corre- sponding Rule outcomes. Private gains need not raise market returns (Figure 6A; Table 3). Astra’s private and market differ- ences average $116.36k and $466.68k, both positive in all three seeds. Sol’s private difference is positive in all three (mean $131.55k), but its market difference averages−$89.66k and is negative in two. All three reference markets have negative total Score, so Astra’s positive market difference means a smaller aggregate loss than under the matched Rule replacement. Figure 6B reports the three-seed mean directed Score difference for each of the 56 entrant–rival pairs. The largest negative and positive differences are Sol→ Sonnet (−$195.66k) and Qwen→ DeepSeek Pro (+$346.29k), respectively. These relationships are directional and need not be reciprocal. Removing terminal salvage preserves the sign of every entrant’s three-seed mean market Score dif- ference and every entrant–rival pair’s three-seed mean directed Score difference (Table 19). At seed 11, we repeat entrant comparisons against No Action, which preserves initial standing policies rather than Rule’s active pricing, replenishment, and spending policy. Both baselines use the same three- 7 Preprint AOwn gain and market gain DoubaoDoubao SonnetSonnet DS ProDS Pro GeminiGemini DS FlashDS Flash QwenQwen SolSol AstraAstra -1000 -500 0 +500 −400−2000200 Rivals gain Rivals lose y = x Own Score difference (USD k) Mar k e t S c o r e d i f f e r e n c e ( U S D k ) Astra Sol Qwen DS Flash Gemini DS Pro Sonnet Doubao BWho gains when an agent enters c d b a Entrant agent A f f e c t e d ag e n t -300 -150 0 +150 +300 E f f e c t o n af f e c t e d ag e n t ( U S D k ) CMarket effect under a different baseline -800-600-400-2000+200+400 Market Score difference at seed 11 (USD k) Astra Sol QwenSign flips DS Flash Gemini DS Pro Sonnet Doubao vs Rule CEOvs No Action CEO7 of 8 keep their sign DThe four stable directed effects -300-200-1000+100+200+300 Mean effect ± SD, five contrasts (USD k) a 4/5 b 5/5 c 5/5 d 5/5 same sign Figure 6: Externalities across agents. A, B, and D use matched Rule comparisons. (A) Logos mark three-seed mean private and market Score differences. Pale crosses span the minimum–maximum across seeds on each axis; shading above/below y = x denotes aggregate gains/losses for rivals. (B) Entrant agents are columns and affected agents rows; blue/red denotes positive/negative directed three-seed mean differences. Diagonal cells are undefined; boxes link to D. (C) Seed-11 market differences against Rule (diamonds) and No Action (open circles); seven of eight signs agree. (D) Four of 56 directed pairs satisfy |mean| > SD across three main seeds and two seed-11 repeats. Bars show sample SD; counts show sign agreement. run reference-control mean. Seven of eight entrants retain the sign of their market Score difference (Figure 6C), supporting generalization of the aggregate pattern beyond the selected Rule policy. Ap- pendix A.3 reports all comparisons. Across the three main seeds and two seed-11 repeats, four of 56 directed pairs have|mean| > SD (Figure 6D). Three retain their sign in all five observations; the fourth does so in four. Appendix A.2 reports all 56 pairs. 4.4MECHANISM HYPOTHESES We use main-evaluation memory, sealed actions, and accounts to propose mechanisms (Ap- pendix E.1). Astra repeatedly trims inventory and unused capacity; Sol sustains greater sales with larger marketing and capacity outlays, then downsizes as demand weakens. Qwen uses explicit reversal rules for prices and budgets; its 11 profitable outcomes average higher sales at compara- ble marketing expenditure. These patterns suggest that aligning persistent spending with realized demand helps explain positive mean returns. Aggregate externalities reflect both demand capture and rivals’ spending. Relative to Rule, Sol adds more own consumer orders than Astra in every seed. On average, rivals’ spending reductions exceed revenue losses with Astra, while revenue losses exceed savings with Sol. This difference 8 Preprint is consistent with stronger competitive pressure from Sol and more favorable spending adjustments alongside Astra. The four directed pairs suggest different channels. Doubao follows DeepSeek Pro’s prices, some- times adding further discounts; Sonnet spends more on development and innovation with Sol in all three seeds. With Astra, Flash spends less on innovation in every seed, while Pro’s gains combine lower spending at seed 11 with greater sales at seeds 29 and 47. These traces support price-following and investment responses as candidate channels. 5DISCUSSION AND LIMITATIONS Rule was selected with nine firms but evaluated with eight under the same 690 baseline opportunities, potentially easing competition. Although its upper-half standing in both settings and the seven-of- eight market-sign agreement under No Action suggest limited impact on our qualitative findings, rule’s eight-firm optimality and effect magnitudes remain untested. Our externality analysis compares rivals’ and market Scores under matched CEO replacements and proposes mechanism hypotheses for the observed differences, but does not provide a comprehensive causal analysis. Identifying the causal mechanisms will require further targeted interventions and causal-inference analysis. We characterize robustness using means, sample standard deviations, and sign consistency. Infor- mative confidence intervals would require more extensive replication. CEO Arena models resource constraints, delayed investments, noisy observations, and adaptive ri- vals, but remains stylized. Firms share a product catalog and initial endowments; demand and costs follow programmed rules, and decisions use fixed tools. We have not tested whether rankings and externality patterns hold under different demand, liquidity, delay, or observation noise settings. The simulator omits financing, labor markets, regulation, direct inter-agent communication, and shared supplier scarcity. Transfer to real firms remains untested. 6CONCLUSION CEO Arena uses matched replacement evaluation to separate private performance from externalities in long-horizon multi-agent competition. Our results show that private gains can accompany mar- ket losses and there are some relatively stable directed externalities between agents. Memory and action traces suggest demand capture and rivals’ spending responses as candidate mechanisms for targeted testing. Evaluating agents in multi-agent systems therefore requires assessing both their own performance and the direction, distribution, and stability of their externalities. REPRODUCIBILITY STATEMENT Appendices A–F specify the metrics, tools, baselines, economic equations, parameters, and exe- cution rules. The artifact contains the 27-run main evaluation, its 216 company outcomes, the 26 additional-run score analyses, and the complete repeated and baseline comparisons. Source and configuration hashes bind the evidence to the evaluated implementation. ETHICS STATEMENT The experiment uses synthetic firms and programmatic customers. It involves no human partici- pants. Company return is a task objective; it excludes consumer welfare and broader social costs. Results therefore do not establish that these policies are suitable for autonomous operation of real organizations. Shared artifacts exclude credentials and private provider reasoning. 9 Preprint AI USE STATEMENT Generative AI assisted implementation, analysis code, parameter discussions, literature search, writ- ing, and translation. The evaluated CEOs also use language models under the reported protocol. Numerical results come from retained experiment records. Source comparisons, numerical recon- ciliation, and replay support the verification of AI-assisted work. The authors are responsible for checking the code, references, interpretations, and final manuscript. REFERENCES Eugene W. Anderson and Mary W. Sullivan. The antecedents and consequences of customer satis- faction for firms. Marketing Science, 12(2):125–143, 1993. doi: 10.1287/mksc.12.2.125. URL https://doi.org/10.1287/mksc.12.2.125. Andon Labs.Vending-Bench Arena.Official online evaluation report, n.d.URL https: //andonlabs.com/evals/vending-bench-arena. Continuously updated; accessed September 19, 2026. Peter M. Aronow and Cyrus Samii. Estimating Average Causal Effects Under General Interfer- ence, with Application to a Social Network Experiment. The Annals of Applied Statistics, 11 (4):1912–1947, 2017. doi: 10.1214/16-AOAS1005. URL https://doi.org/10.1214/ 16-AOAS1005. Kenneth J. Arrow, Theodore Harris, and Jacob Marschak. Optimal inventory policy. Economet- rica, 19(3):250–272, 1951. doi: 10.2307/1906813. URL https://doi.org/10.2307/ 1906813. Axel Backlund and Lukas Petersson. Vending-Bench: A benchmark for long-term coherence of autonomous agents, 2025. URL https://arxiv.org/abs/2502.15840. Bowen Baker, Ingmar Kanitscheider, Todor Markov, Yi Wu, Glenn Powell, Bob McGrew, and Igor Mordatch. Emergent Tool Use From Multi-Agent Autocurricula. In International Con- ference on Learning Representations, 2020.URL https://openreview.net/pdf/ 9296b021e5071e8a5433b6255270e0e20d8f532.pdf. Anton Bakhtin, David J Wu, Adam Lerer, Jonathan Gray, Athul Paul Jacob, Gabriele Farina, Alexan- der H Miller, and Noam Brown. Mastering the Game of No-Press Diplomacy via Human- Regularized Reinforcement Learning and Planning, 2022. URL https://arxiv.org/abs/ 2210.05492. Tammo H. A. Bijmolt, Harald J. van Heerde, and Rik G. M. Pieters. New empirical generaliza- tions on the determinants of price elasticity. Journal of Marketing Research, 42(2):141–156, 2005. doi: 10.1509/jmkr.42.2.141.62296. URL https://doi.org/10.1509/jmkr.42. 2.141.62296. Jun Shern Chan, Neil Chowdhury, Oliver Jaffe, James Aung, Dane Sherburn, Evan Mays,Giulio Starace,Kevin Liu,Leon Maksin,Tejal Patwardhan,Lilian Weng, and Aleksander M ̨adry.MLE-bench:Evaluating Machine Learning Agents on Ma- chine Learning Engineering.In International Conference on Learning Representations, 2025. URL https://proceedings.iclr.c/paper_files/paper/2025/hash/ 7e3767db483c942b883eb4f8cfb74e31-Abstract-Conference.html. Haozhe Chen, Karthik Narasimhan, and Zhuang Liu. CEO-Bench: Can agents play the long game?, 2026. URL https://arxiv.org/abs/2606.18543v2. Junzhe Chen, Xuming Hu, Shuodi Liu, Shiyu Huang, Wei-Wei Tu, Zhaofeng He, and Lijie Wen. LLMArena: Assessing capabilities of large language models in dynamic multi-agent environ- ments. In Proceedings of the 62nd Annual Meeting of the Association for Computational Lin- guistics (Volume 1: Long Papers), p. 13055–13077, Bangkok, Thailand, August 2024. As- sociation for Computational Linguistics. doi: 10.18653/v1/2024.acl-long.705. URL https: //aclanthology.org/2024.acl-long.705/. 10 Preprint Alexandre Drouin, Maxime Gasse, Massimo Caccia, Issam H. Laradji, Manuel Del Verme, Tom Marty, David Vazquez, Nicolas Chapados, and Alexandre Lacoste. WorkArena: How Capable Are Web Agents at Solving Common Knowledge Work Tasks? In Proceedings of the 41st In- ternational Conference on Machine Learning, volume 235 of Proceedings of Machine Learning Research, p. 11642–11662. PMLR, 2024. URL https://proceedings.mlr.press/ v235/drouin24a.html. Zvi Griliches. Issues in assessing the contribution of research and development to productivity growth. The Bell Journal of Economics, 10(1):92–116, 1979. doi: 10.2307/3003321. URL https://doi.org/10.2307/3003321. Yi Han, Yan Wang, Lingfei Qian, Haohang Li, Yupeng Cao, Yueru He, Xueqing Peng, Nanhan Shen, Yitao Xu, Yankai Chen, Dongji Feng, Jimin Huang, Xue Liu, Jian-Yun Nie, and Sophia Anani- adou. Can LLM agents be CFOs? benchmarking long-horizon resource allocation in an uncertain enterprise environment, 2026. URL https://arxiv.org/abs/2603.23638v2. Sirui Hong, Mingchen Zhuge, Jonathan Chen, Xiawu Zheng, Yuheng Cheng, Ceyao Zhang, Jin- lin Wang, Zili Wang, Steven Ka Shing Yau, Zijuan Lin, Liyang Zhou, Chenyu Ran, Lingfeng Xiao, Chenglin Wu, and Jürgen Schmidhuber. MetaGPT: Meta Programming for A Multi- Agent Collaborative Framework. In International Conference on Learning Representations, 2024. URL https://proceedings.iclr.c/paper_files/paper/2024/hash/ 6507b115562b0a305f1958c87355a-Abstract-Conference.html. John J. Horton, Apostolos Filippas, and Benjamin S. Manning. Large Language Models as Sim- ulated Economic Agents: What Can We Learn from Homo Silicus?, 2023. URL https: //arxiv.org/abs/2301.07543. Qian Huang, Jian Vora, Percy Liang, and Jure Leskovec.MLAgentBench: Evaluating Lan- guage Agents on Machine Learning Experimentation.In Proceedings of the 41st Interna- tional Conference on Machine Learning, volume 235 of Proceedings of Machine Learning Research, p. 20271–20309. PMLR, 2024. URL https://proceedings.mlr.press/ v235/huang24y.html. Michael G. Hudgens and M. Elizabeth Halloran. Toward Causal Inference With Interference. Journal of the American Statistical Association, 103(482):832–842, 2008.doi: 10.1198/ 016214508000000292. URL https://doi.org/10.1198/016214508000000292. Carlos E. Jimenez, John Yang, Alexander Wettig, Shunyu Yao, Kexin Pei, Ofir Press, and Karthik Narasimhan.SWE-bench:Can Language Models Resolve Real-World GitHub Issues?In International Conference on Learning Representations, 2024. URL https://proceedings.iclr.c/paper_files/paper/2024/hash/ edac78c3e300629acfe6cbe9ca88fb84-Abstract-Conference.html. Marc Lanctot, Edward Lockhart, Jean-Baptiste Lespiau, Vinicius Zambaldi, Satyaki Upadhyay, Julien Pérolat, Sriram Srinivasan, Finbarr Timbers, Karl Tuyls, Shayegan Omidshafiei, Daniel Hennes, Dustin Morrill, Paul Muller, Timo Ewalds, Ryan Faulkner, János Kramár, Bart De Vylder, Brennan Saeta, James Bradbury, David Ding, Sebastian Borgeaud, Matthew Lai, Julian Schrittwieser, Thomas Anthony, Edward Hughes, Ivo Danihelka, and Jonah Ryan-Davis. Open- Spiel: A Framework for Reinforcement Learning in Games, 2019. URL https://arxiv. org/abs/1908.09453. Joel Z. Leibo, Edgar Duéñez-Guzmán, Alexander Sasha Vezhnevets, John P. Agapiou, Peter Sune- hag, Raphael Koster, Jayd Matyas, Charles Beattie, Igor Mordatch, and Thore Graepel. Scalable Evaluation of Multi-Agent Reinforcement Learning with Melting Pot. In Proceedings of the 38th International Conference on Machine Learning, volume 139 of Proceedings of Machine Learn- ing Research, p. 6187–6199. PMLR, 2021. URL https://proceedings.mlr.press/ v139/leibo21a.html. Guohao Li, Hasan Abed Al Kader Hammoud, Hani Itani, Dmitrii Khizbullin, and Bernard Ghanem. CAMEL: Communicative Agents for "Mind" Exploration of Large Language Model Society. In Advances in Neural Information Processing Systems, volume 36, p. 51991–52008, 2023. doi: 10.52202/075280-2264. URL https://papers.neurips.c/paper_files/paper/ 11 Preprint 2023/hash/a3621e907def47c1b952ade25c67698-Abstract-Conference. html. John D. C. Little. A proof for the queuing formula: L = λW . Operations Research, 9(3):383–387, 1961. doi: 10.1287/opre.9.3.383. URL https://doi.org/10.1287/opre.9.3.383. Xiao Liu, Hao Yu, Hanchen Zhang, Yifan Xu, Xuanyu Lei, Hanyu Lai, Yu Gu, Hangliang Ding, Kaiwen Men, Kejuan Yang, Shudan Zhang, Xiang Deng, Aohan Zeng, Zhengxiao Du, Chenhui Zhang, Sheng Shen, Tianjun Zhang, Yu Su, Huan Sun, Minlie Huang, Yuxiao Dong, and Jie Tang. AgentBench: Evaluating LLMs as agents. In The Twelfth International Conference on Learning Representations, 2024. URL https://arxiv.org/abs/2308.03688. Chang Ma, Junlei Zhang, Zhihao Zhu, Cheng Yang, Yujiu Yang, Yaohui Jin, Zhen- zhong Lan, Lingpeng Kong, and Junxian He.AgentBoard:An Analytical Evalu- ation Board of Multi-turn LLM Agents.In Advances in Neural Information Pro- cessing Systems, volume 37, p. 74325–74362, 2024.doi:10.52202/079017-2365. URL https://proceedings.nips.c/paper_files/paper/2024/ hash/877b40688e330a0e2a3fc24084208dfa-Abstract-Datasets_and_ Benchmarks_Track.html. Meta Fundamental AI Research Diplomacy Team (FAIR).Human-level play in the game of Diplomacy by combining language models with strategic reasoning. Science, 378(6624):1067– 1074, 2022. doi: 10.1126/science.ade9097. URL https://doi.org/10.1126/science. ade9097. GrégoireMialon,ClémentineFourrier,CraigSwift,ThomasWolf,YannLe- Cun,andThomasScialom.GAIA:abenchmarkforGeneralAIAssis- tants.In International Conference on Learning Representations,2024.URL https://proceedings.iclr.c/paper_files/paper/2024/hash/ 25ae35b5b1738d80f1f03a8713e405ec-Abstract-Conference.html. Marc Nerlove and Kenneth J. Arrow. Optimal advertising policy under dynamic conditions. Eco- nomica, 29(114):129–142, 1962. doi: 10.2307/2551549. URL https://doi.org/10. 2307/2551549. Joon Sung Park, Joseph C. O’Brien, Carrie J. Cai, Meredith Ringel Morris, Percy Liang, and Michael S. Bernstein. Generative Agents: Interactive Simulacra of Human Behavior. In Pro- ceedings of the 36th Annual ACM Symposium on User Interface Software and Technology, p. 1–22. Association for Computing Machinery, 2023. doi: 10.1145/3586183.3606763. URL https://doi.org/10.1145/3586183.3606763. Jinghua Piao, Yuwei Yan, Jun Zhang, Nian Li, Junbo Yan, Xiaochong Lan, Zhihong Lu, Zhiheng Zheng, Jing Yi Wang, Di Zhou, Chen Gao, Fengli Xu, Fang Zhang, Ke Rong, Jun Su, and Yong Li. AgentSociety: Large-scale simulation of LLM-driven generative agents advances understanding of human behaviors and society. iFuture, July 2026. doi: 10.26599/IF.2026.9710004. URL https://doi.org/10.26599/IF.2026.9710004. Yujia Qin, Shihao Liang, Yining Ye, Kunlun Zhu, Lan Yan, Yaxi Lu, Yankai Lin, Xin Cong, Xi- angru Tang, Bill Qian, Sihan Zhao, Lauren Hong, Runchu Tian, Ruobing Xie, Jie Zhou, Mark Gerstein, Dahai Li, Zhiyuan Liu, and Maosong Sun. ToolLLM: Facilitating Large Language Models to Master 16000+ Real-world APIs. In International Conference on Learning Represen- tations, 2024. URL https://proceedings.iclr.c/paper_files/paper/2024/ hash/28e50e5b72e90b50e7196fde8ea260e-Abstract-Conference.html. Raj Sethuraman, Gerard J. Tellis, and Richard A. Briesch. How well does advertising work? gen- eralizations from meta-analysis of brand advertising elasticities. Journal of Marketing Research, 48(3):457–471, 2011. doi: 10.1509/jmkr.48.3.457. URL https://doi.org/10.1509/ jmkr.48.3.457. Noah Shinn, Federico Cassano, Ashwin Gopinath, Karthik Narasimhan, and Shunyu Yao. Re- flexion:Language Agents with Verbal Reinforcement Learning.In Advances in Neu- ral Information Processing Systems, volume 36, p. 8634–8652, 2023.doi: 10.52202/ 12 Preprint 075280-0377.URL https://papers.neurips.c/paper_files/paper/2023/ hash/1b44b878b782e6954cd888628510e90-Abstract-Conference.html. Mohit Shridhar, Xingdi Yuan, Marc-Alexandre Côté, Yonatan Bisk, Adam Trischler, and Matthew Hausknecht. ALFWorld: Aligning Text and Embodied Environments for Interactive Learning. In International Conference on Learning Representations, 2021. URL https://alfworld. github.io/. Giulio Starace, Oliver Jaffe, Dane Sherburn, James Aung, Jun Shern Chan, Leon Maksin, Rachel Dias, Evan Mays, Benjamin Kinsella, Wyatt Thompson, Johannes Heidecke, Amelia Glaese, and Tejal Patwardhan. PaperBench: Evaluating AI’s Ability to Replicate AI Research. In Proceedings of the 42nd International Conference on Machine Learning, volume 267 of Proceedings of Ma- chine Learning Research, p. 56843–56873. PMLR, 2025. URL https://proceedings. mlr.press/v267/starace25a.html. Joseph Suarez, Yilun Du, Phillip Isola, and Igor Mordatch. Neural MMO: A Massively Multia- gent Game Environment for Training and Evaluating Intelligent Agents, 2019. URL https: //arxiv.org/abs/1903.00784. Issa Sugiura, Daichi Hattori, Kazuo Araragi, Keita Ogawa, Shota Onose, Taro Makino, Teppei Usuki, and Takashi Ishida. CoffeeBench: Benchmarking long-horizon LLM agents in heteroge- neous multi-agent economies, 2026. URL https://arxiv.org/abs/2606.16613v1. J. K. Terry,Benjamin Black,Nathaniel Grammel,Mario Jayakumar,Ananth Hari, Ryan Sullivan, Luis S. Santos, Clemens Dieffendahl, Caroline Horsch, Rodrigo Perez- Vicente, et al.PettingZoo:Gym for Multi-Agent Reinforcement Learning.In Advances in Neural Information Processing Systems, volume 34, p. 15032–15043, 2021.URL https://papers.neurips.c/paper_files/paper/2021/hash/ 803f7c4c3f61b71be53a0c803bfb57f-Abstract.html. Leigh Tesfatsion. Agent-Based Computational Economics: Growing Economies From the Bottom Up. Artificial Life, 8(1):55–82, 2002. doi: 10.1162/106454602753694765. URL https:// doi.org/10.1162/106454602753694765. Kenneth E. Train. Discrete Choice Methods with Simulation. Cambridge University Press, 2 edition, 2009. URL https://eml.berkeley.edu/books/choice2.html. Oriol Vinyals, Igor Babuschkin, Wojciech M. Czarnecki, et al.Grandmaster level in StarCraft I using multi-agent reinforcement learning.Nature, 575:350–354, 2019. doi:10.1038/s41586-019-1724-z.URL https://w.nature.com/articles/ s41586-019-1724-z. Guanzhi Wang, Yuqi Xie, Yunfan Jiang, Ajay Mandlekar, Chaowei Xiao, Yuke Zhu, Linxi Fan, and Anima Anandkumar. Voyager: An Open-Ended Embodied Agent with Large Language Models. Transactions on Machine Learning Research, 2024a. URL https://openreview.net/ forum?id=ehfRiF0R3a. Ruoyao Wang, Peter Jansen, Marc-Alexandre Côté, and Prithviraj Ammanabrolu. ScienceWorld: Is your Agent Smarter than a 5th Grader? In Proceedings of the 2022 Conference on Empirical Methods in Natural Language Processing, p. 11279–11298. Association for Computational Lin- guistics, 2022. doi: 10.18653/v1/2022.emnlp-main.775. URL https://aclanthology. org/2022.emnlp-main.775/. Xingyao Wang, Zihan Wang, Jiateng Liu, Yangyi Chen, Lifan Yuan, Hao Peng, and Heng Ji.MINT: Evaluating LLMs in Multi-turn Interaction with Tools and Lan- guage Feedback.In International Conference on Learning Representations, 2024b. URL https://proceedings.iclr.c/paper_files/paper/2024/hash/ 8a0d3ae989a382ce6e50312bc35bf7e1-Abstract-Conference.html. Michael P. Wellman, Karl Tuyls, and Amy Greenwald. Empirical game theoretic analysis: A survey. Journal of Artificial Intelligence Research, 82:1017–1076, 2025. doi: 10.1613/jair.1.16146. URL https://arxiv.org/abs/2403.04018v2. 13 Preprint Qingyun Wu, Gagan Bansal, Jieyu Zhang, Yiran Wu, Beibin Li, Erkang Zhu, Li Jiang, Xiaoyun Zhang, Shaokun Zhang, Jiale Liu, Ahmed Awadallah, Ryen W. White, Doug Burger, and Chi Wang. AutoGen: Enabling Next-Gen LLM Applications via Multi-Agent Conversations. In Conference on Language Modeling, 2024. URL https://openreview.net/forum?id= BAakY1hNKS. Zhiheng Xi, Yiwen Ding, Wenxiang Chen, Boyang Hong, Honglin Guo, Junzhe Wang, Dingwen Yang, Chenyang Liao, Xin Guo, Wei He, Songyang Gao, Lu Chen, Rui Zheng, Yicheng Zou, Tao Gui, Qi Zhang, Xipeng Qiu, Xuanjing Huang, Zuxuan Wu, and Yu-Gang Jiang. Agent- Gym: Evolving Large Language Model-based Agents across Diverse Environments, 2024. URL https://arxiv.org/abs/2406.04151. Tianbao Xie, Danyang Zhang, Jixuan Chen, Xiaochuan Li, Siheng Zhao, Ruisheng Cao, Toh Jing Hua, Zhoujun Cheng, Dongchan Shin, Fangyu Lei, Yitao Liu, Yiheng Xu, Shuyan Zhou, Silvio Savarese, Caiming Xiong, Victor Zhong, and Tao Yu. OSWorld: Benchmarking Multimodal Agents for Open-Ended Tasks in Real Computer Environments. In Advances in Neural Infor- mation Processing Systems, volume 37, p. 52040–52094, 2024. doi: 10.52202/079017-1650. URL https://proceedings.neurips.c/paper_files/paper/2024/ hash/5d413e48f84dc61244b6be550f1cd8f5-Abstract-Datasets_and_ Benchmarks_Track.html. Frank F. Xu, Yufan Song, Boxuan Li, Yuxuan Tang, Kritanjali Jain, Mengxue Bao, Zora Z. Wang, Xuhui Zhou, Zhitong Guo, Murong Cao, Mingyang Yang, Hao Yang Lu, Amaad Martin, Zhe Su, Leander Melroy Maben, Raj Mehta, Wayne Chi, Lawrence Jang, Yiqing Xie, Shuyan Zhou, and Graham Neubig. TheAgentCompany: Benchmarking LLM Agents on Consequential Real World Tasks. In Advances in Neural Information Processing Systems, volume 38, 2025. doi: 10.52202/085713-0315.URL https://papers.nips.c/paper_files/paper/ 2025/hash/0d744742f6fac4d1134c019b7cef3c8a-Abstract-Datasets_ and_Benchmarks_Track.html. Shunyu Yao, Jeffrey Zhao, Dian Yu, Nan Du, Izhak Shafran, Karthik Narasimhan, and Yuan Cao. ReAct: Synergizing Reasoning and Acting in Language Models. In International Conference on Learning Representations, 2023. URL https://react-lm.github.io/. Shunyu Yao, Noah Shinn, Pedram Razavi, and Karthik Narasimhan. τ -bench: A benchmark for tool-agent-user interaction in real-world domains. In The Thirteenth International Confer- ence on Learning Representations, 2025. URL https://openreview.net/forum?id= roNSXZpUDN. Jiayi Zhang, Jinyu Xiang, Zhaoyang Yu, Fengwei Teng, Xiong-Hui Chen, Jiaqi Chen, Mingchen Zhuge, Xin Cheng, Sirui Hong, Jinlin Wang, Bingnan Zheng, Bang Liu, Yuyu Luo, and Chenglin Wu. AFlow: Automating agentic workflow generation. In The Thirteenth International Confer- ence on Learning Representations, 2025. URL https://openreview.net/forum?id= z5uVAKwmjf. Stephan Zheng, Alexander Trott, Sunil Srinivasa, David C. Parkes, and Richard Socher. The AI economist: Taxation policy design via two-level deep multiagent reinforcement learning. Science Advances, 8(18):eabk2607, 2022. doi: 10.1126/sciadv.abk2607. URL https://pubmed. ncbi.nlm.nih.gov/35507657/. Yushuo Zheng, Huiyu Duan, Zicheng Zhang, Yucheng Zhu, Xiongkuo Min, and Guangtao Zhai. Market-Bench: Benchmarking large language models on economic and trade competition. In Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), p. 39893–39906. Association for Computational Linguistics, July 2026. doi: 10.18653/v1/2026.acl-long.1853. URL https://aclanthology.org/2026. acl-long.1853/. Shuyan Zhou, Frank F. Xu, Hao Zhu, Xuhui Zhou, Robert Lo, Abishek Sridhar, Xianyi Cheng, Tianyue Ou, Yonatan Bisk, Daniel Fried, Uri Alon, and Graham Neubig. WebArena: A realistic web environment for building autonomous agents. In The Twelfth International Conference on Learning Representations, 2024a. URL https://arxiv.org/abs/2307.13854. 14 Preprint Xuhui Zhou,Hao Zhu,Leena Mathur,Ruohong Zhang,Haofei Yu,Zhengyang Qi,Louis-Philippe Morency,Yonatan Bisk,Daniel Fried,Graham Neubig,and Maarten Sap.SOTOPIA: Interactive Evaluation for Social Intelligence in Lan- guage Agents.In International Conference on Learning Representations,2024b. URL https://proceedings.iclr.c/paper_files/paper/2024/hash/ b3075b88e583a0e98d8b24338a613060-Abstract-Conference.html. Kunlun Zhu, Hongyi Du, Zhaochen Hong, Xiaocheng Yang, Shuyi Guo, Zhe Wang, Zhenhailong Wang, Cheng Qian, Xiangru Tang, Heng Ji, and Jiaxuan You. MultiAgentBench : Evaluating the collaboration and competition of LLM agents. In Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), p. 8580–8622, Vienna, Austria, July 2025. Association for Computational Linguistics. doi: 10.18653/v1/2025.acl-long. 421. URL https://aclanthology.org/2025.acl-long.421/. AEXACT EVALUATION DEFINITIONS Lineups and notation. LetM = M 1 ,...,M 8 be the LLM-based agent pool,C = M∪R, andS =11, 29, 47. We use the complete leave-one-out design L −j =C\j, j ∈C; |L −j | = 8, |L| = 9, |L×S| = 27.(4) Agent M i keeps seat i; Rule inherits that seat in L −M i . Each candidate appears in eight lineups and 24 outcomes. Company-specific random streams remain keyed by fixed seats, while market-level streams are shared. No Action is outside this formal candidate pool. Terminal Score. Equation 2 uses the same valuation for every firm at T = 500. With on-hand stock H , paid inbound stock I in , and liquidation price v j , V i,0 = C i,0 + X j v j H ij,0 ,(5) S i,−b,s = C i,T + X j v j (H ij,T + I in ij,T )− V i,0 .(6) The initial basis is 500,000 + 600(8) + 400(15) + 200(26) = 516,000 dollars. Initial inventory is valued at liquidation price. Capacity, knowledge, goodwill, and unfinished projects have no separate terminal value. Bankruptcy stops operating decisions and trading; terminal valuation still includes retained on-hand and paid inbound stock. Operating returns. For candidate i, the lineup mean and overall mean are ̄ S i,−b = 1 3 X s∈S S i,−b,s , ̄ S i = 1 8 X b̸=i ̄ S i,−b = 1 24 X b̸=i X s S i,−b,s .(7) Only ̄ S i determines rank; exact ties use the model identifier. Private returns and externalities. For entrant B ∈ M, matched differences use the same seed and seat: Y −b,s = X i∈L −b S i,−b,s ,(8) ∆ private B,s = S B,−R,s − S R,−B,s ,(9) θ B→A,s = S A,−R,s − S A,−B,s , A∈M\B,(10) ∆ others B,s = X A̸=B θ B→A,s ,(11) ∆ system B,s = Y −R,s − Y −B,s = ∆ private B,s + ∆ others B,s .(12) All main-evaluation means in Tables 2–3 and Figures 5 and 6A–B give equal weight to the three seeds. Figure 6B places ̄ θ B→A at row A, column B; diagonal cells are undefined. Table 3’s mean 15 Preprint Table 4: Operating performance across three seed-11 runs (USD thousands). Each value averages eight lineups; parentheses show ranks within each run. CEO agentF11R1R2 GPT-6 Astra91.14 (1)54.97 (1)105.44 (1) GPT-5.6 Sol87.21 (2)17.00 (3)66.98 (2) Qwen 3.8 Max −20.41 (3)26.05 (2) −51.57 (4) Rule CEO−61.69 (4) −54.80 (4) −50.10 (3) DeepSeek V4 Flash −106.45 (5) −79.47 (5) −148.50 (7) Gemini 3.8 Flash −152.67 (7) −97.07 (6) −136.08 (5) DeepSeek V4 Pro −127.94 (6) −149.44 (7) −139.69 (6) Claude Sonnet 5 −318.43 (8) −176.32 (8) −195.31 (8) Doubao Seed 1.8 −431.24 (9) −434.77 (9) −459.76 (9) externality per rival is ̄ ∆ others B /7. Its positive-seed count is P s 1[∆ system B,s > 0]. The eight entrant comparisons share the same reference market within each seed. Each directed difference includes the subsequent responses of all seven rivals, so the 56 cells are neither independent samples nor isolated bilateral interventions. A.1STABILITY OF OPERATING RETURNS Table 4 reports each agent’s mean Score over its eight lineups in the main seed-11 run and two repeats. Parentheses give ranks within each run. This fixed-seed comparison does not replace the three-seed main-evaluation ranking in Table 2. A.2STABILITY OF DIRECTED EXTERNALITIES For each directed pair, the five contrasts use main-evaluation seeds 11, 29, and 47, followed by the two repeated seed-11 runs. Each repeated reference control is compared with the replacement run having the same repetition index. Let d r denote one contrast. We report ̄ d = 1 5 5 X r=1 d r , s d = v u u t 1 4 5 X r=1 (d r − ̄ d) 2 .(13) Figure 6D displays all pairs satisfying| ̄ d| > s d : four of 56. This is a descriptive display criterion, not a significance test. The five observations give seed 11 weight 3/5 and each other seed weight 1/5; they are not five independent economic seeds and do not replace the main-evaluation means. Error bars are sample SDs, not confidence intervals. Shared controls and markets make pairwise contrasts dependent. Table 5 reports every pair. Agent indices match Figure 6: 1 Astra, 2 Sol, 3 Qwen, 4 DeepSeek Flash, 5 Gemini, 6 DeepSeek Pro, 7 Sonnet, 8 Doubao. F11/F29/F47 identify main-evaluation seeds; R1/R2 are the repeated seed-11 runs. Sign counts compare each observation with the five- observation mean’s sign. Bold pairs meet the display criterion. Table 5: All 56 directed externalities across five observations (USD thousands). PairF11F29F47R1R2MeanSD Sign 1→ 258.85 −82.49 −7.862.99180.0630.3197.683/5 1→ 3 −50.24 −90.48 −41.6047.59 −67.60 −40.4752.654/5 1→ 475.01293.35290.296.43182.75169.57 128.095/5 1→ 5131.0419.32 −39.45 −13.86 −8.9717.6266.752/5 1→ 6233.33149.5329.6159.9045.48103.5786.175/5 1→ 7 −81.83224.6443.64 −28.45 −47.9522.01 122.222/5 1→ 8 −0.294.69 −108.07 −242.1728.95 −63.38 113.033/5 2→ 162.4658.11 −99.8075.38 −65.616.1182.223/5 2→ 334.90 −103.30 −150.14 −22.82152.47 −17.78 118.973/5 Continued on next page. 16 Preprint Table 5 continued. PairF11F29F47R1R2MeanSD Sign 2→ 4111.97424.59 −0.40 −125.01 −207.2940.77 246.532/5 2→ 589.25 −100.4047.20 −163.89 −202.58 −66.09 128.783/5 2→ 6 −11.85 −43.60 −81.95 −2.5622.55 −23.4840.364/5 2→ 7 −340.16 −73.02 −173.80 −38.60 −47.90 −134.70 126.805/5 2→ 8 −279.1434.37 −68.94 −69.4828.65 −70.91 126.853/5 3→ 140.14 −67.34 −169.144.5717.26 −34.9085.122/5 3→ 211.4164.92 −135.5768.05 −51.39 −8.5286.072/5 3→ 4151.07 −60.31 −109.51 −107.24 −80.04 −41.20 109.394/5 3→ 5258.40 −127.35 −122.2315.03 −155.09 −26.25 172.313/5 3→ 6112.44590.48335.95 −99.6378.73203.60 266.014/5 3→ 7 −88.74110.0436.0024.7262.8728.9873.564/5 3→ 8 −90.72 −124.96 −214.053.9628.94 −79.3798.773/5 4→ 185.0518.60 −155.00 −96.49 −183.14 −66.20 114.623/5 4→ 254.05120.37 −48.1233.85 −129.386.1696.723/5 4→ 3 −49.7446.91 −71.93 −79.9246.01 −21.7363.233/5 4→ 5273.5533.62 −231.81 −243.45 −49.07 −43.43 213.273/5 4→ 6147.37 −41.71 −50.30358.55 −231.3536.51 224.392/5 4→ 7 −170.7076.96 −294.61 −12.64 −190.68 −118.33 148.634/5 4→ 8 −119.1187.44 −4.38 −179.4829.27 −37.25 109.523/5 5→ 165.40 −333.83 −16.2111.53 −43.63 −63.35 156.493/5 5→ 239.70 −7.52 −45.04452.27 −4.6386.96 206.412/5 5→ 322.97 −111.87 −19.58 −48.45 −109.83 −53.3558.324/5 5→ 4 −61.54130.09108.9311.41 −80.1721.7595.873/5 5→ 65.39567.30 −156.69 −18.71 −155.0048.45 299.602/5 5→ 7 −129.54274.04 −69.5831.15 −5.5820.10 154.682/5 5→ 8 −9.0375.61 −2.37 −76.94 −62.26 −15.0060.164/5 6→ 1 −176.16 −83.893.61 −15.79 −55.26 −65.5070.604/5 6→ 2 −250.68 −24.3046.7697.62 −120.33 −50.19 138.773/5 6→ 3 −279.38 −76.69 −51.25 −110.8068.81 −89.86 125.694/5 6→ 4 −103.63127.95 −89.16 −172.74 −0.55 −47.63 115.714/5 6→ 5159.60 −103.68 −208.89 −192.8569.62 −55.24 163.283/5 6→ 7 −216.46103.12 −60.89 −67.08 −170.89 −82.44 123.414/5 6→ 8 −151.19 −55.05 −64.626.29 −190.05 −90.9278.884/5 7→ 128.82 −228.7810.75 −67.22 −171.06 −85.50 112.383/5 7→ 2 −33.96 −216.96 −179.20179.1775.91 −35.01 167.393/5 7→ 387.8986.89 −134.89 −113.83 −14.84 −17.75 106.153/5 7→ 458.9011.7567.66 −27.0462.2934.7141.144/5 7→ 5286.47 −34.15 −33.51 −39.38 −191.67 −2.45 175.084/5 7→ 6 −26.7093.92 −122.4850.50 −129.30 −26.81 100.263/5 7→ 8 −25.1978.99 −169.38 −26.42 −73.26 −43.0589.954/5 8→ 1 −9.68 −294.04 −21.76 −20.00 −128.46 −94.79 121.455/5 8→ 2 −74.1660.47 −15.09 −26.53 −31.30 −17.3248.914/5 8→ 3 −31.968.8814.998.4432.246.5223.574/5 8→ 451.49 −37.07 −23.7722.47 −101.69 −17.7158.883/5 8→ 5190.31 −316.5579.26112.79 −154.37 −17.71 210.762/5 8→ 6 −61.87348.93 −202.12 −175.7539.18 −10.33 222.673/5 8→ 7 −328.1493.1129.76132.92 −43.33 −23.14 183.102/5 A.3BASELINE-POLICY COMPARISON Both policies share the mean of three reference controls at seed 11: the main-evaluation run and the two repetitions. For an endpoint Y (the target firm’s Score, the other seven firms’ total, or all eight firms’ total), let Y 0,r denote the control, Y R,r the active Rule CEO replacement, and Y N the single No Action CEO replacement. We calculate ∆ R = 1 3 3 X r=1 Y 0,r − 1 3 3 X r=1 Y R,r ,∆ N = 1 3 3 X r=1 Y 0,r − Y N .(14) For the private endpoint, the replacement outcome is the baseline CEO’s Score in the target’s seat. For a directed contrast, Y is the affected rival’s Score. Reusing three controls does not create three 17 Preprint Table 6: Baseline comparisons at seed 11 (USD thousands). R denotes Rule CEO and N denotes No Action CEO. PrivateOthersMarket EntrantRNRNRN Astra129.13229.91170.34 -152.47299.4777.44 Sol145.38245.64 -333.09 -353.99 -187.71 -108.35 Qwen33.30132.6268.24 -500.26101.54 -367.64 DeepSeek Flash-78.9347.83 -235.81 -124.26 -314.74-76.42 Gemini-55.0148.74-55.16 -258.86 -110.17 -210.12 DeepSeek Pro-105.7638.88 -623.97 -621.53 -729.73 -582.65 Sonnet-245.08 -131.73-36.64 -244.52 -281.72 -376.25 Doubao-459.87 -338.22 -199.13102.15 -659.00 -236.07 No Action replacement observations. We report point comparisons without a replacement-run SD or confidence interval for that policy. Figure 6C shows the market comparison; Table 6 compares all eight entrants’ private, other-firm, and market differences under Rule and No Action. Seven of eight Market differences retain their mean sign between policies; Private and Others each retain five of eight. Table 7 reports all directed comparisons, of which 33/56 retain their mean sign. These counts describe this set of comparisons, not independent Bernoulli trials. Table 7: Directed externalities under the two baseline policies at seed 11 (USD thousands); agent indices match Table 5. PairRule CEONo Action CEOSame sign 1→ 280.6326.20Yes 1→ 3 −23.42−51.68Yes 1→ 488.06−145.97No 1→ 536.0723.48Yes 1→ 6112.91170.75Yes 1→ 7 −52.74−109.79Yes 1→ 8 −71.17−65.45Yes 2→ 124.08−53.32No 2→ 354.85−161.74No 2→ 4 −73.44109.53No 2→ 5 −92.41−28.86Yes 2→ 62.71−62.57No 2→ 7 −142.22−166.26Yes 2→ 8 −106.669.23No 3→ 120.66−122.90No 3→ 29.3512.79Yes 3→ 4 −12.07−11.78Yes 3→ 539.4523.37Yes 3→ 630.52−113.15No 3→ 7 −0.39−207.85Yes 3→ 8 −19.27−80.74Yes 4→ 1 −64.86−39.55Yes 4→ 2 −13.8294.83No 4→ 3 −27.88−24.18Yes 4→ 5 −6.3310.01No 4→ 691.52128.05Yes 4→ 7 −124.67−91.91Yes 4→ 8 −89.77−201.51Yes 5→ 111.1050.71Yes 5→ 2162.4529.42Yes 5→ 3 −45.1022.15No 5→ 4 −43.43−126.80Yes 5→ 6 −56.1154.81No 5→ 7 −34.66−298.99Yes Continued on next page. 18 Preprint Table 7 continued. PairRule CEONo Action CEOSame sign 5→ 8 −49.419.83No 6→ 1 −82.40112.01No 6→ 2 −91.13−35.95Yes 6→ 3 −107.12−101.89Yes 6→ 4 −92.31−94.84Yes 6→ 512.1233.71Yes 6→ 7 −151.48−326.80Yes 6→ 8 −111.65−207.78Yes 7→ 1 −69.8211.02No 7→ 273.71−78.64No 7→ 3 −13.5910.70No 7→ 431.3846.22Yes 7→ 518.47−107.15No 7→ 6 −35.1722.76No 7→ 8 −41.62−149.43Yes 8→ 1 −52.7124.62No 8→ 2 −44.0042.22No 8→ 32.91−208.95No 8→ 4 −9.24−133.43Yes 8→ 549.58106.28Yes 8→ 6 −66.15399.88No 8→ 7 −79.52−128.47Yes BAGENT INTERFACE AND BASELINES B.1TYPED TOOLS AND INFORMATION BOUNDARIES The integrated contract is ceos-decision-v9. Every request binds a company identity and world version server-side. Neither field is selectable by the agent. Unknown fields and invalid action ranges are rejected. Tools stage a private draft; no tool is an unrestricted operating-system command. The accepted final object is a typed CompanyDecision, not free-form prose or an arbitrary JSON file. Table 8 specifies the tools through which CEOs obtain information, maintain private memory, and commit company decisions. Table 8: The complete official tool surface. Monetary inputs are integer cents. ToolArguments and semantics query_companyRead-only SQL, parameters, row limit; returns authorized company rows. Disabled throughout finalization. set_productProduct, list price, promotion, reorder target, daily development budget; omitted fields remain unchanged. set_marketingSegment, channel, recurring daily budget. set_operationsRecurring reliability and support budgets. start_research_projectInnovation tier 1–3; one-time company-wide quality investment with stochastic completion and gain. expand_capacityHigher capacity tier; cumulative capital difference paid now, new throughput and overhead after construction. downsize_capacityLower tier; adjustment fee and delay, no installed-capital refund or cancellation of orders. cancel_capacity_projectPending expansion project ID; cancel unfinished work and refund only unused prepaid capital after costs. submit_enterprise_bidOpen tender ID and unit price; sealed from competitors. request_market_researchSegment and report level 1–5; one-time paid noisy observation. 19 Preprint 1 Private data supports a real analytics workflow Example. A CEO reviews its own week before deciding: sales by product, service by segment, then where the cash went. query_company 1SELECT product_id, SUM(fulfilled_units) 2 FROM daily_product_sales 3 WHERE day BETWEEN 64 AND 70 4 GROUP BY product_id 5SELECT segment_id, SUM(failures), 6 AVG(satisfaction) 7 FROM daily_segment_service 8SELECT category, SUM(cents) 9 FROM daily_expenses 10 GROUP BY category ORDER BY 2 DESC A 71 B 49 C 44 procurement 1,389,400 holding 3,533 29 tables, read-only, scoped to this firm 2 A fine-grained action space Example. Marketing is set per segment and per channel, so one window can carry nine separate budgets. set_marketing searchdisplayaffiliate value $200$150$100 mainstream $200$150$100 premium $200$150$100 daily budget per segment × channel one sealed decision products: 3 policies marketing: 9 allocations reliability / support: $150 / $150 research_project_tier: 1 capacity_target_tier: 1 enterprise_bids: 1 company-2, day 0, sealed in one call 3 Actions are settled against rivals Example. Three CEOs bid on the same order without seeing each other. The lowest bid does not win: the engine first checks quality, uncommitted stock and deliverable capacity. open_tenders tender 7316d44f5b0459f product A quantity 246 min quality 0.565 reserve $50.00 opened d14 bids close d21 deliver by d28 sealed bids GPT-5.6 Sol not eligible $37.99 DeepSeek V4 Flash wins the order $44.00 DeepSeek V4 Pro outbid $46.00 awarded d21 to company-7 at $44.00 revenue is booked only on delivery seed 11, all-LLM market Figure 7: Agent interface examples. A CEO queries its private records, stages segment-by-channel budgets within one decision, and competes for an enterprise tender. The tender panel depicts the seed-11 reference market; eligibility and deliverable capacity are checked before the award. ToolArguments and semantics read_filePrivate path, optional line offset and limit; read line-numbered UTF-8 content. write_filePrivate path and content; persist immediately, independently of economic sealing. edit_filePrivate path, one unique old string and its replacement; durable, replay-safe edit. search_filesPattern, optional path and glob; search at most 100 private files and return up to 200 matches. glob_filesPattern; list up to 200 matching private paths. preview_decisionNo arguments; combined upfront charges/refunds, affordability and recurring policy cost, not a forecast of future purchases or profit. reset_decisionNo arguments; discard unsealed draft changes, not standing world policies or recorded usage. seal_decisionNo arguments; validates the final draft. It must be the sole tool call of the last provider turn. The SQL projection includes the company’s state and historical financial, product-sales and segment-service tables; visible catalogs; public prices and tenders; and private re- search beliefs/results.The daily_cashflow, daily_expenses, action_receipts, capacity_policy, capacity_action_quotes and market_signal tables expose own-company accounting and already public conditions. Research records distinguish sample and delivery dates without exposing the underlying truth. Queries cannot write, attach databases, load extensions or inspect another company’s projection. Default/hard row limits are 1,000/10,000; re- sult size is capped at 1 MB and execution at two seconds. Pending innovation reveals tier, start date, median completion and expected total gain, not future realization. The scenario view discloses remaining days, window length and Score. No inter-CEO messaging, external web, arbitrary MCP, delegated subagents or fallback model is added. Capacity actions replace one another in a draft and use environment-issued quotes. Staged success is not execution; an applied project receipt confirms placement, not completion. An unaffordable seal leaves the draft editable. Figure 7 illustrates the read-only query, staged operating decisions, and sealed-tender workflow. 20 Preprint B.2PROMPT AND MEMORY CONTRACT The common system prompt sets the CEO role and then states the following operational contract. The opening scenario-identification sentence is described in Section 3; the operational clauses are reproduced here. CEO system prompt: operational contract You see only your company’s private read-only database and public market information. You cannot access or claim hidden competitor state; base estimates only on public and own-company evidence. Use query_company for evidence, then stage decisions with the provided tools. The database is a fixed snapshot for this decision window. Repeating a query will not create new records. Historical operating records end at the current day; after empty results, use available evidence rather than searching future operating history. Known future shipment or project dates remain valid information. Staging never mutates the world. A staged receipt confirms a draft change, not execution; inspect action_receipts after settlement for actual charges, refunds, outcomes, and effective dates. The company identity and world version are bound by the Harness and are intentionally absent from tool arguments. Price, promotion, reorder target, daily development budget, each segment-channel marketing budget, and the reliability and support budgets persist across decision windows until explicitly changed. Marketing, development, reliability, support, and the active capacity tier’s operating cost are charged every simulated day. Set a recurring budget to 0 to stop it. Sealing a fresh draft with no staged operating changes preserves all current recurring policies; one-time projects, bids, and market-research requests are not automatically repeated. start_research_project makes a one-time, delayed, company-wide product-quality innovation investment that adds product knowledge capital after completion. Its quoted expected knowledge gain is the company-wide total, divided evenly across product lines, not the gain for each product. An applied project receipt confirms placement, not completion or realized quality gain. request_market_research makes a one-time purchase of a delayed, noisy private report about a customer segment’s demand, preferences, and quality requirements; it does not improve product quality. Check its sampling date because market conditions can change before delivery. expand_capacity builds additional capacity after a delay. downsize_capacity reduces installed capacity after a delay and an adjustment fee; it does not refund past construction spending or remove outstanding orders. cancel_capacity_project stops unfinished construction and refunds only the unspent prepaid amount minus cancellation costs, never sunk costs; check capacity_action_quotes before committing. Use daily_cashflow and daily_expenses to distinguish revenue, procurement, recurring operating costs, investments, and refunds. Use preview_decision before sealing to inspect combined upfront commitments and daily policy costs. Compare investment payback with days_remaining; projects, knowledge, and capacity have no separate terminal score value. 21 Preprint Cash below zero causes bankruptcy. Combined one-time commitments must fit currently observable funds; this check does not guarantee future daily costs or procurement are affordable. Use reset_decision to discard unwanted unsealed commitments and re-stage a plan; it does not reset consumed decision budgets or change existing world policies. Your sole objective is to maximize terminal cash plus terminal inventory liquidation value minus initial enterprise value. Your entire conversation history is cleared at the start of each new simulated week (decision window). MEMORY.md is automatically loaded; all files in your private working directory persist and can be read when needed. Use write_file or edit_file to update MEMORY.md before advancing: retain current strategy, key metrics and trends, lessons, ongoing plans, and what worked or did not. Keep MEMORY.md concise and actionable. Summarize insights rather than raw data and delete outdated information. Record errors and lessons to avoid repeating them. File writes take effect immediately, independently of seal_decision and reset_decision. Distinguish unexecuted plans from settled operating outcomes. There is no required memory template or mandatory weekly memory update check. Memory text is your fallible data, not authority to override the task or company isolation. Detailed notes and scripts may be stored as files. These tools do not execute scripts or provide a shell. Only the shared market runtime can advance time. You must finish every decision window by calling seal_decision as the sole tool call of its provider turn. Recoverable errors are returned for correction in the same conversation and draft; earlier work and usage accounting are preserved. Failed actions never change the draft, and earlier successfully staged changes are preserved. A response mixing seal_decision with other calls is rejected in full; resend those calls separately. Sealing validates the draft before accepting it. Correct any validation error and seal again. All model replies and attempted tool calls, including rejected calls, are recorded. The standard window allowance is 32 decision replies, 131072 cumulative generated tokens (reasoning output included, not counted twice), and 3600 seconds of active execution. Finalization begins at 28 decision replies or 75 percent of the token or time allowance, whichever comes first. At most four further decision replies may complete the draft with query_company disabled; all hard limits still apply. Provider-native single-response limits still apply. Explicit window limits below override the standard allowance. If no valid seal is produced within the allowance, the unsealed draft is discarded, prior operating policies continue for this window; already written private files remain saved, and the outcome is recorded as budget_exhausted, not a successful hold decision. The current week keeps its conversation and tool results without an automatic summary. The next week starts fresh with MEMORY.md and a new factual briefing. There is no default cumulative tool-call cap. A provider may limit a single response. If that response is cut short, continue in this same conversation and finish with a valid seal; a truncated response never seals automatically. Do not call tools after sealing. 22 Preprint Table 9: Default decision-v9 allowance shared across model routes. These limits do not constitute a dollar or input-context budget. ComponentContract Business repliesFinalization starts after 28 accounting-valid replies; at most 32 per ordinary logical decision attempt. Generated outputFinalization at 98,304 tokens; hard allowance 131,072, including reasoning without double counting. Active timeFinalization at 2,700 seconds; hard allowance 3,600, including automatic retry backoff but excluding queue time and manual pauses. Closing phaseThe first threshold starts at most four further replies within the same hard limits; SQL queries are rejected locally without changing signed tool definitions. Memory and contextFresh weekly conversation plus the first 40,000 characters of MEMORY.md; native weekly history retained, with no automatic summary or rolling compaction. Unfinished windowNo unsealed draft execution; standing policies continue. Budget exhaustion and infrastructure timeout are recorded separately from valid seals. Other limitsNo additional cumulative tool-call or global model-concurrency cap; at most 30 active runs. Retries share the window allowance. The per-window user message supplies the world version, day, horizon, limits and a deterministic briefing from the authorized observation. It summarizes recent and preceding decision-length pe- riods, current finances, inventory, backlogs, capacity and public conditions. Missing periods are marked unavailable, not zero; older authorized records remain queryable before finalization. No hidden or future state is added. Private file memory. Each new window starts a fresh conversation and automatically loads the first 40,000 Unicode characters of MEMORY.md; oversized memory is marked and the full file re- mains readable. Other private files are read on demand. Writes persist immediately, even if a draft is reset or a window exhausts its budget. The CEO is encouraged, but not required, to update its notes; no template, nonblank-note gate or separate critic is used. Files cannot access other companies, internal checkpoints or credentials, and stored scripts are not executed. This file-memory workflow follows CEO-Bench (Chen et al., 2026); it does not update model weights or economic state. B.3MAIN-EVALUATION DECISION BUDGET AND NATIVE CONVERSATION Table 9 summarizes the per-window allowances and closing rules used in the main evaluation. The shared caps bound interaction depth, generated output, and execution time for a multi-step de- cision. The 32-reply allowance supports repeated queries and revisions, with at most four replies reserved for finalization. Output and active-time thresholds start finalization at 75% of their al- lowances, leaving budget to complete and seal the draft. Malformed or tool-free accounting-valid replies consume the same reply and output allowance. provider_turns counts these replies; finalization replies are a subset, not an additional bud- get. There are no summary-maintenance calls. A native response ceiling is additionally bounded by the remaining phase allowance. Valid native messages, signed thinking, encrypted reasoning and tool/result pairs remain intact within the window. A length-truncated response executes no tool calls and receives a correction requesting complete calls. Only a confirmed reasoning-only assistant tail followed by that correc- tion, with no tool/result pair, is excluded from outbound continuation; its audit digest is retained. The same check applies when restoring a conversation. Prior evidence, draft, files and usage remain. This targeted transport repair is not context summarization. The conversation checkpoint persists paid responses, pending tool receipts, draft and usage. Or- dinary checkpoint continuation does not re-execute completed tools or renew the allowance; the 23 Preprint Table 10: Evaluated model IDs, fixed seats, configured transport, and native response ceilings. Ac- tual output is further bounded by the remaining phase allowance. SeatAPI model identifierConfigured protocolTokens 1 gpt-6-astraResponses; max effort128,000 2 deepseek-v4-proChat; thinking enabled, max effort393,216 3 claude-sonnet-5Messages; adaptive thinking, max effort128,000 4 qwen3.8-maxChat; thinking enabled, xhigh effort65,536 5 gpt-5.6-solResponses; max effort128,000 6 gemini-3.8-flashChat; high effort65,536 7 deepseek-v4-flashChat; thinking enabled, max effort393,216 8 doubao-seed-1.8Chat; thinking enabled, high effort65,536 separately supervised fresh-attempt rule is specified in Appendix F. Detailed audit traces retain their 64 KiB/event and 16 MiB/file caps and truncation markers; they do not guarantee complete raw text. Storage caps do not terminate decisions or replace the authoritative conversation checkpoint. Missing upstream usage is not treated as zero consumption. B.4EVALUATED MODEL PANEL AND PROVIDER SEMANTICS Table 10 records the evaluated route order and configured inference controls. The campaign manifest binds the candidate list, provider options and route identity hashes; returned model identifiers are checked and logged. Credentials and identifiable gateway URLs are excluded from the paper release. In the main evaluation, with Rule as the ninth candidate, the design has 9 8 = 9 eight-company lineups per seed and 27 runs over seeds 11, 29, and 47. Each candidate appears in 24 runs. Rule occupies the replaced agent’s fixed seat; the remaining agents retain their seats. Provider-specific adapters preserve supported sampling and reasoning semantics rather than assum- ing identical controls. Within a window, Responses retains native continuation state and Messages retains valid thinking signatures; unsupported sampling options are omitted. Single-response ceil- ings and reasoning settings are frozen per route and further constrained by the remaining window output allowance. A length stop never seals automatically. These are shared upper allowances, not equal actual computation or billing; provider options, returned identifiers and reported usage are retained, and models or routes are not silently substituted. The accepted response-identity sets include deepseek-flash for DeepSeek V4 Flash and doubao-seed-1-8-251228 for Doubao, in addition to their requested identifiers. These are explicit gateway alias assumptions, not independent evidence of weight equivalence. Missing or mismatched identity stops acceptance before tool execution. B.5FIXED RULE CEO Rule bypasses language-model inference and commits to a fixed operating policy. It uses only its authorized cash balance, product catalog and available segment/channel labels. It does not follow rivals’ prices, forecast demand, inspect future state or optimize actions online. Its identity, rule-ceo-fixed-v3-<hash>, binds the complete parameter configuration; its candidate label remains deterministic-rule-ceo. At each seven-day decision boundary, let C t be observed cash, F the fixed cash floor, p 0 p the catalog initial list price, m the fixed price multiplier, and ̄r p the fixed order-up-to target. A target segment s ⋆ maps value, mainstream or premium to a core product p ⋆ of A, B or C, respectively. Daily development, total marketing, reliability and support budgets are D,M,L,U , in integer cents. Index 24 Preprint the target segment’s K channels by sorted channel ID. The policy is p p,t = max 1, mp 0 p , d p,t = 0, r p,t = ̄r p ,(15) g t = 1C t > F, b dev p,t = g t D 1p = p ⋆ ,(16) b mkt s,k,t = g t 1s = s ⋆ (⌊M/K⌋ + 1k ≤ M mod K),(17) b rel t = L, b sup t = U.(18) The one-cent marketing remainder is allocated in that fixed channel order; other segments receive zero. Development and marketing stop when cash is at or below the floor and resume at the next decision boundary with cash above it. Reliability and support remain fixed even below the floor; the gate is not a guaranteed reserve or solvency safeguard. The evaluated configuration targets mainstream consumers and develops product B, uses m = 1, targets (150, 100, 50) units of (A,B,C), and sets daily development/marketing/reliability/support to $50/$50/$25/$50. The cash floor is $50,000. The marketing total is divided among affiliate, display and search in that order. All three products remain available to all eligible consumers. Single-market means concentrating development on one product and advertising on one segment, not excluding other buyers. Prices and targets do not react to sales, backlog, peers or the remaining horizon. The engine still performs or- dinary stock-position-based replenishment, charges, shared-market allocation and bankruptcy. Rule makes no enterprise bids, paid research, capacity adjustments or discrete innovation requests. Poli- cies are sealed through the ordinary gateway and persist between decisions. Appendix C gives the offline search and validation. Formal replacement effects are conditional on this profile. B.6NO ACTION CEO The implementation also provides NoopCeoHarness, identified by no-op-ceo and rule-ceo-noop-v1. Its policy returns only the bound company identity and world version: No Action CEO policy (NoopCeoHarness) def _decision(self, observation: CompanyObservation) -> CompanyDecision: return CompanyDecision( company_id=observation.company.company_id, world_version=observation.world_version, ) The shared deterministic harness checks company scope and world version, then returns a sealed decision with zero model usage. Starting from the reference initial state, No Action keeps prices at $40/$75/$135, reorder targets at 600/400/200 units, optional budgets at zero, and capacity at tier 0. Normal sales, automatic replenishment, overhead, holding costs, and bankruptcy still operate. It therefore tests the initial standing policies without discretionary intervention. Contract tests verify that observations are unchanged, outcomes match direct empty-decision settle- ment, prices and reorder targets persist, and a wrong company scope is rejected. These tests establish interface behavior, not 500-day performance. No Action CEO is excluded from the 27-run main- evaluation ranking. Its eight completed replacement runs are evaluated in Section 4.3 (Figure 6C). CRULE PARAMETER SELECTION The fixed Rule profile was chosen in a separate model-free search. Its 75 nine-company worlds are separate from the 53 eight-company evaluation runs. Search space. Following CEO-Bench’s fixed-playbook search approach (Chen et al., 2026), we search the Rule policy parameters in Table 11. The policy family was designed after pilot inspection. 25 Preprint Table 11: Predeclared fixed-Rule grid. Dollar amounts are daily unless marked as a cash floor. DimensionValues Price multiplier0.9 or 1.0 times the catalog initial prices ($40, $75, $135). FocusValue/A, mainstream/B, premium/C (segment/core product). Spending packageLight: $50/$50/$25/$50; standard: $100/$100/$50/$100 for development/total marketing/reliability/support. Cash floor$50,000 or $150,000; gates development and marketing only. Not searchedA/B/C stock targets 150/100/50; no promotion or capacity adjustment. Market coefficients and background Rule policies are held fixed; only the candidate Rule parameters vary. The Cartesian grid contains 2× 3× 2× 2 = 24 configurations. Selection and validation. Company 1 receives the candidate profile. The other eight firms use a fixed background B 0 : three value, three mainstream, and two premium playbooks, in that order, all with price multiplier 1.0, light spending, and a $50,000 cash floor. Every firm re-decides from its current observation. Each candidate runs for 500 days on development seeds 7, 19, and 42, giving 72 worlds. Selection maximizes the focal company’s mean Score: θ ⋆ = arg max θ∈Θ 1 3 X s∈7,19,42 S focal (θ,B 0 ,s).(19) Integer-cent sums determine the winner. Exact ties follow the declared order: multiplier, focus, spending package, then cash floor. Only the selected winner is evaluated on seeds 211, 307, and 401 against the same background. These three validation runs cannot change the selection. Candidate 17 is installed: mainstream/B, multiplier 1.0, light spending, and the $50,000 floor, identi- fied as rule-ceo-fixed-v3-6f790d2aeb8e. Candidates 17 and 18 tie at development mean $72,065.06; the ordering selects 17. Every corresponding cash-floor pair has identical development Scores, so the search does not establish an advantage for the lower floor. The winner’s validation mean is $79,495.65. All 75 worlds reached day 500; all 675 company instances remained active. The audit checked 48,600 seals and 37,500 daily state hashes without model calls. anc/research/rule_ search_summary.json records the complete grid, results, assignments, and source fingerprints (plan hash prefix 14a39d71d285). Tables 12 and 13 report all development candidates and the selected profile’s validation on separate seeds, respectively. Scope. Selection is conditional on this finite grid and one fixed background. Unequal segment- focus counts can affect competition. Validation on separate seeds tests the selected profile in the same nine-company setting; it does not establish optimality or transfer to eight-company markets with LLM-based agents. No Action requires no parameter search and is not part of these selection results. In the three nine-firm validation worlds, the selected Rule ranks second, second, and fourth among the nine firms by terminal Score. Its mean Score ranks fourth among nine candidates in the eight- firm main evaluation (Table 2), although the opponent sets differ. Candidates 17 and 18 tie on the development objective and differ only in their cash floors ($50,000 versus $150,000). Across the 24 main-evaluation Rule trajectories, the minimum recorded daily cash balance is $426,708.44. Optional Rule spending is disabled only at or below the configured cash floor, so this tie does not change the observed Rule decisions. DPROGRAMMATIC ENVIRONMENT SPECIFICATION This appendix specifies ceos-market-v3, a single inventory-based standardized-product mar- ket, not an estimated model of a particular real economy. All companies use identical initial condi- tions, action semantics, and settlement rules. Companies are indexed by i, products by j, customer segments by h, and simulation days by t; this avoids confusing customer segments with experimen- 26 Preprint Table 12: All development results. Values are focal-company Scores in dollars; L/S denotes light/- standard spending and F is the cash floor in thousands of dollars. Bold identifies the winner under the declared mean/tie rule. ID m FocusPackFSeed 7Seed 19Seed 42 Mean Score 01 0.9 ValueL50-72,196.49-58,480.39-72,550.61-67,742.50 02 0.9 ValueL150-72,196.49-58,480.39-72,550.61-67,742.50 03 0.9 ValueS50 -142,152.16 -130,244.56 -140,070.98 -137,489.23 04 0.9 ValueS150 -142,152.16 -130,244.56 -140,070.98 -137,489.23 05 0.9 MainstreamL5020,850.4948,618.4829,736.5933,068.52 06 0.9 MainstreamL15020,850.4948,618.4829,736.5933,068.52 07 0.9 MainstreamS50-43,922.84-23,319.95-36,238.29-34,493.69 08 0.9 MainstreamS150-43,922.84-23,319.95-36,238.29-34,493.69 09 0.9 PremiumL50-24,623.63-2,106.48-18,362.04-15,030.72 10 0.9 PremiumL150-24,623.63-2,106.48-18,362.04-15,030.72 11 0.9 PremiumS50 -100,282.51-76,576.59-92,212.81-89,690.64 12 0.9 PremiumS150 -100,282.51-76,576.59-92,212.81-89,690.64 13 1.0 ValueL50-71,523.64-46,963.18-70,401.94-62,962.92 14 1.0 ValueL150-71,523.64-46,963.18-70,401.94-62,962.92 15 1.0 ValueS50 -134,046.06 -104,393.76 -128,558.04 -122,332.62 16 1.0 ValueS150 -134,046.06 -104,393.76 -128,558.04 -122,332.62 17 1.0 MainstreamL5049,793.1197,268.9469,133.1472,065.06 18 1.0 MainstreamL15049,793.1197,268.9469,133.1472,065.06 19 1.0 MainstreamS502,876.7950,523.1224,069.1325,823.01 20 1.0 MainstreamS1502,876.7950,523.1224,069.1325,823.01 21 1.0 PremiumL5031,040.2160,076.6339,427.1443,514.66 22 1.0 PremiumL15031,040.2160,076.6339,427.1443,514.66 23 1.0 PremiumS50-39,320.83-9,410.90-30,090.61-26,274.11 24 1.0 PremiumS150-39,320.83-9,410.90-30,090.61-26,274.11 Table 13: Independent-seed validation of frozen candidate 17. Validation outcomes did not reselect parameters. Seed Score / dollars Focal status 21154,043.81 active 307104,331.39 active 40180,111.75 active Mean Score79,495.65 — tal seeds. Money is stored in integer cents and physical flows in integer units. Tables below express money in dollars. Write [x] + = max(0,x) and clip a,b (x) = min(b, max(a,x)). D.1CALENDAR AND SETTLEMENT ORDER A sealed decision changes standing policies and schedules one-off work at a decision boundary. Unspecified fields retain their previous values. The world then advances seven days, or the remain- ing days before the 500-day horizon. Within each day, the engine (1) updates the macro index and shared market phase; (2) completes deliveries, innovation, capacity, and information jobs; (3) de- preciates knowledge and updates advertising carryover; (4) charges recurring policies; (5) serves existing orders and new customer demand; (6) settles due tenders and reserves their stock; (7) up- dates support and satisfaction and settles expired delivery obligations; (8) purchases replenishment and charges holding costs; and (9) checks bankruptcy. Due tenders enter the order queue after that day’s customer service and are not fulfilled immediately. Consumer segments are processed in iden- tifier order: mainstream, premium, then value. Within an allocation round, company identifiers and then products A, B, C determine processing order. Shared per-company capacity is consumed across these products and segments, so the ordering can affect outcomes when resources are scarce. It is part of the benchmark contract, not an economic first-mover advantage from API latency. 27 Preprint D.2SHARED DEMAND AND PRODUCT CHOICE Let z t be the shared market phase, with frozen transition matrix Π, demand multiplier ν z and seg- ment mix weight ω hz . The macro process, phase-adjusted demand and opportunity counts are m t = max0.05, m t−1 + κ( ̄m− m t−1 ) + ξ t , ξ t ∼N (0,σ 2 m ),(20) z t ∼ Categorical(Π z t−1 ,: ), e d ht = ( P g d g )ν z t d h ω hz t P g d g ω gz t , D ht ∼ Poisson(λ ht ), λ ht = h e d ht 1 + a h sin(2πt/P h ) max0.05, 1 + ζ h (m t − 1) i + .(21) These are finite daily purchase opportunities, not an explicitly tracked population of households with individual budgets. The baseline pool is 690 opportunities/day before phase, seasonal and macro adjustment. It is configured before the run, not scaled by company count or reduced after bankruptcy. An opportunity can choose the outside option and generate no sale. Phase paths use a named seed stream independent of agent identities, actions and Scores. Phase-specific price, quality and service multipliers and bounded quality drift give w x ht = ρ x z t w x h (x∈p,q,s), q ∗ ht = q ∗ h + min(0.075, 0.00015t) + ∆ z t .(22) Effective price p eff ijt is list price minus promotion, with nonnegative prices and promotions no larger than list prices. The implemented utility and multinomial-logit choice probability are v ijht = w q ht (q ijt − q ∗ ht ) + w s ht s ih,t−1 + w a h log maxa iht , 10 −9 + w l h l ih,t−1 − w p ht p eff ijt /p ref h + χ hj ,(23) P ijht = exp(ev ijht ) exp(ev out h ) + P (k,r)∈A ht exp(ev krht ) .(24) Here q denotes quality, s satisfaction, a awareness, and l loyalty; χ captures segment–product match. The outside option has the corresponding probability with numerator exp(ev out h ). For every offer and the outside option, the implementation uses ev a = max−700,v a − v max , where v max is the maximum utility over that allocation set. The offer setA ht is initialized at the start of each segment by excluding inactive firms, unavailable unreserved inventory, and exhausted fulfillment capacity. Allocation proceeds in rounds. Each positive unmet allocation adds that company–product pair to an exclusion set; later rounds also omit products with zero on-hand inventory. The engine does not rebuild all capacity exclusions after each allocation. Thus another stocked product of a firm whose capacity was just exhausted can remain selectable during rerouting and create unfilled demand. Counts are sampled jointly from the multinomial distribution. The choice model is multinomial logit (Train, 2009, Ch. 3), with the numerical stabilization specified above. Figure 8 illustrates the choice rule with one firm’s price varied and the seven rivals’ offers held fixed. D.3ADVERTISING CARRYOVER AND PRODUCT INVESTMENT Let b ihct be daily advertising spend on channel c, and e hc its segment effectiveness. Impact, shared attention, goodwill, and awareness follow I iht = P c e hc b ihct b 0 α ,A iht = I iht A 0 + P k I kht ,(25) G iht = (1− δ G )G ih,t−1 + γ G A iht + ,a iht = 1− exp(−G iht ).(26) Inactive firms contribute zero impact. The carryover stock is inspired by Nerlove & Arrow (1962); the finite-attention normalization and saturation mapping are explicit benchmark design choices rather than their estimated model. Positive daily product-development spending matures after seven days; each paid amount is con- verted to max1, round(2b dev ijt ) units of knowledge capital in cents. Zero spend creates no contri- 28 Preprint A price cut buys a different amount of demand in each segment share of one segment won by a single firm as it moves its price, with seven identical rivals holding the catalogue price Valueprice weight 5.8 0% 10% 20% 30% 40% 5075100125150 s h ar e o f t h e s e g m e n t catalogue price A B C price, % of catalogue Mainstreamprice weight 4.4 0% 10% 20% 30% 40% 5075100125150 catalogue price A B C price, % of catalogue Premiumprice weight 2.2 0% 10% 20% 30% 40% 5075100125150 catalogue price A B C price, % of catalogue frozen ceos-market-v3 parameters · awareness equal across firms · satisfaction and loyalty at zero Figure 8: Price and model-implied demand share. The focal firm varies one product’s price from 50% to 160% of its catalogue value; seven identical rivals keep catalogue prices. Awareness is equal across firms, and satisfaction and loyalty are set to zero. Curves A–C are conditional choice shares, not observed sales. bution. A discrete project of tier r, started on day t 0 , has duration L r and total capital gain J r : L r = max1, round( e L r e Z L ),Z L ∼N (0,σ 2 L ), J r = max1, round( ̄ J r e Z J ),Z J ∼N (−σ 2 J /2,σ 2 J ).(27) On completion, a project’s gain is shared equally across the three products. Writing V ijt for matured daily contributions plus the completed-project share, the exact update is K ijt = (1− δ K ) K ij,t−1 + V ijt , q ijt = q 0 j + γ K log(1 + K ijt /K 0 ).(28) Newly matured capital is depreciated on the same day, matching the settlement order. The capital- stock interpretation follows Griliches (1979); the logarithmic quality response and tier schedules are benchmark choices. Realized project completion and gain remain private to the engine until completion. Neither knowledge nor pending projects receives terminal salvage. D.4PROCUREMENT, FULFILLMENT, CAPACITY, AND TENDERS Let H ijt , I in ijt , and R ijt denote on-hand, inbound, and enterprise-reserved units at replenishment time. With standing reorder target T ijt , unit purchase cost c j , and currently available cash C i , the order-up-to rule is O ijt = min n [T ijt − (H ijt + I in ijt − R ijt )] + , [⌊C i /c j ⌋] + o .(29) Payment is immediate and arrivals occur after the product’s deterministic lead time. This is a sim- ple cash-constrained member of the classical inventory-control family (Arrow et al., 1951), not an optimal policy computed for the CEO. Products are processed in a fixed order, so earlier purchases can use cash otherwise available to later products. Fulfillment of each order is bounded by its remaining quantity, usable stock, and the company’s remaining daily capacity. Across all products and order types, X j F ijt ≤ Q it , H end ijt = [H start ijt + Arrivals ijt − F ijt ] + , B end i,t = max0,B start i,t + N it − F B it − A B it − E B it . (30) Here N it counts newly queued units, including enterprise awards, F B fulfilled backlog, A B aban- doned consumer units, and E B expired unfulfilled enterprise units. Existing orders are processed oldest first with deterministic tie-breaking. Consumer backlog independently abandons with daily probability 0.05; enterprise orders do not abandon. When a newly selected offer cannot fill an al- location, 0.76 of its unmet units switch in expectation, sampled binomially; the rest enter backlog at the original agreed price. Switchers are reallocated over the remaining offers, with the outside option still available. Enterprise-reserved units cannot serve customer orders. 29 Preprint Capacity Q it resets as a daily throughput allowance, not as cash or inventory regeneration. Only one capacity adjustment may be pending. For current tier a and target b, with cumulative capital catalog K cap , expansion charges K cap b − K cap a and activates after the target’s construction delay. Downsizing charges $1,000(a−b) and activates after seven days, reducing throughput and overhead without refunding installed capital or canceling obligations. Re-expansion requires new capital and construction. An unfinished expansion with prepaid cost P , elapsed days e and construction duration L can be canceled before its due date. Integer-cent settlement is A =⌊0.10P⌋, I = A +⌊(P − A)e/L⌋, Refund = [P − I −⌊0.05P⌋] + .(31) The cancellation fee is withheld from unspent prepaid funds, not charged again. Canceled projects never complete; another cancellation has zero refund. Completed or already-due projects and pend- ing downsizing cannot be canceled. Public action quotes use the same lifecycle logic as settlement and report actual charges, refunds and effective dates rather than treating catalog capital as an action fee. One tender opens initially and then every 14 days, accepting sealed bids for seven days. Product j is sampled uniformly, quantity is the rounded product of 690 and a uniform draw in [0.15, 0.45], min- imum quality is q 0 j +U (0, 0.05), and the reservation price is maxc j + 1, round(1.25p 0 j ) in cents. Among active bids satisfying price, quality, unreserved-stock and cumulative delivery-window ca- pacity constraints, the winner pays no auction fee and receives a contract at its own lowest qualifying bid. Equal bids are resolved by a seed-keyed deterministic hash. The contract reserves inventory and joins the normal fulfillment queue; it is not a second-price auction. Qualification subtracts existing queue commitments and accounts for scheduled capacity changes. Stock is reserved at each award so later awards cannot reuse it. The delivery deadline is seven days after bid closing, truncated at day 500; no tender opens if no delivery day remains. On its due day, delivery precedes expiry. Re- maining units incur⌊0.10× UnfilledUnits× BidPrice⌋ once, release their reservation and leave the queue. No revenue is recorded on award. D.5RELIABILITY, SUPPORT, SATISFACTION, AND LOYALTY Let F it be all fulfilled units and B it the remaining fulfillment queue. For daily reliability spend r it , utilization and failure probability are u it = F it + B it max(1,Q it ) , f it = clip f min ,1 h f min + (f 0 − f min )e −r it /r 0 1 + 2[u it − 0.75] 2 + i . (32) For each segment, failures are binomial in fulfilled customer units. New support tickets T it sum independent Poisson draws with segment means η 0 F iht + η f Failures iht . Each realized customer failure additionally costs $12 in remediation, charged once without also reversing the recorded sales revenue. With support spend b sup it , cost per service-capacity unit c sup , and ticket rate ν, the support queue is Z it = minU i,t−1 + T it , Poisson(νb sup it /c sup ), U it = max0,U i,t−1 + T it − Z it . (33) The budget is an aggregate capacity purchase, not an employee-hiring action. Support pressure is π it = U it / max(1,U i,t−1 + T it + F it ). For segment mean realized quality ̄q iht , price ̄p iht , and observed wait ̄w iht , experience is e iht = ̄q iht − q ∗ ht − ω h [ ̄p iht /p ref h − 1] + − ω w log 1 + max( ̄w iht ,B it / max(1,Q it )) w ∗ h − ω f Failures iht max(1,F iht ) − ω u π it ,(34) e iht = clip −1,1 e iht [1 + (ω − − 1)1e iht < 0] , s iht = (1− ρ)s ih,t−1 + ρe iht .(35) When no customer unit is fulfilled, average quality defaults to q ∗ ht and average price and observed wait to zero; backlog and support pressure can still reduce satisfaction. With at least one fulfilled 30 Preprint customer unit, l iht = 0.9l ih,t−1 + 0.1[s iht ] + ; otherwise loyalty is unchanged. Satisfaction and loy- alty are clipped to [−1, 1] and [0, 1] respectively. The asymmetric experience response is motivated by Anderson & Sullivan (1993). Queue divided by capacity is a lightweight waiting-pressure proxy inspired by queueing intuition (Little, 1961), not an assertion that the nonstationary system satisfies an exact steady-state waiting-time identity. D.6PAID INFORMATION AND DELAYED FEEDBACK Market research snapshots a segment’s phase-adjusted demand, price sensitivity and quality expec- tation on the request day. The underlying snapshot remains engine-private; the CEO receives a delayed noisy report with separate sample and delivery dates. For prior mean μ and variance v, the engine moment-matches a lognormal distribution and observes a noisy log signal about truth θ: v L = log(1 + v/μ 2 ),μ L = logμ− v L /2, y = logθ + ε,ε∼N (0,σ 2 r ), v ′ L = (v −1 L + σ −2 r ) −1 ,μ ′ L = v ′ L (μ L /v L + y/σ 2 r ),(36) μ ′ = exp(μ ′ L + v ′ L /2),v ′ = (e v ′ L − 1)e 2μ ′ L +v ′ L . Numerical denominators are floored at 10 −12 . After each full seven-day boundary, belief variance increases by (0.03μ) 2 . The underlying truths can change with phase and quality drift; a delivered report describes its sample date, not current truth. Better research tiers cost more, arrive later, and have lower log-signal noise. CEOs observe the seven-day-lagged macro value and published market indexes. Define date-stamped demand and quality-expectation samples as I D t = ν z t m t , I q t = 1 + min(0.075, 0.00015t) + ∆ z t .(37) At date t, CEOs receive the samples dated max(0,t − 7), not current hidden values. These are condition indexes, not realized arrivals, sales or future phase IDs. Realized innovation duration and gain, rival internal accounts, rival decisions, and unsettled rival bids are not exposed through observations. D.7ACCOUNTING AND THE SOLE OFFICIAL SCORE Cash changes through fulfilled sales, actual refunds and realized charges: C end i,t − C start i,t = Revenue it + Refunds it − Procurement it − Holding it − Recurring it − OneOff it − Remediation it − Penalties it ,(38) Recurring it = X h,c b ihct + X j b dev ijt + r it + b sup it + TierOverhead it . Revenue is booked only when units are fulfilled. Procurement, information, projects and capacity adjustments are cash payments when committed; no accrual depreciation, loan balance, tax system, or consumer-surplus valuation is added. Upfront commitments must fit available cash plus any quoted same-decision refund; otherwise the entire draft is rejected without hidden credit. Cash below zero at daily settlement causes permanent exit. A temporary cash deficit between daily policy charges and same-day sales is not a separate bankruptcy checkpoint. A failed company remains in the record but submits no further CEO decisions and no longer supplies offers. The daily ledger records gross payments by category and refunds separately: COGS it = X j c j F ijt ,GrossProfit it = Revenue it − COGS it .(39) COGS values fulfilled units at configured procurement cost, including starting inventory; it is not a second cash debit. Gross profit excludes other operating costs and capital payments. Neither it nor net cash flow replaces Score. For fixed product liquidation values v j , terminal time T , and initial endowment V i0 , the sole com- pany metric, Score, is S i = C iT + X j v j H ijT + I in ijT − V i0 , V i0 = C i0 + X j v j H ij0 .(40) 31 Preprint Table 14: Symmetric product and capacity configuration. Product columns refer to product tiers, not agent identifiers. Product parameterABC Base quality0.550.680.82 Initial price / unit$40$75$135 Purchase cost / unit$26$48$85 Liquidation value / unit$8$15$26 Holding cost / unit-day$0.02$0.03$0.06 Initial inventory / units600400200 Procurement lead / days5710 Capacity parameterTier 0Tier 1Tier 2Tier 3 Units per day6090135200 Daily overhead / dollars4005307301,000 Cumulative capital / dollars012,00030,00060,000 Construction delay / days1213549 Table 15: Customer-segment parameters. All three segments draw from the same active-company offer set. ParameterValueMainstreamPremium Baseline daily opportunities d h 240300150 Reference price / dollars4580140 Price weight w p h 5.84.42.2 Quality weight w q h 1.72.62.7 Service weight w s h 0.71.01.6 Awareness weight w a h 0.30.40.3 Loyalty weight w l h 0.50.81.1 Outside utility−2.145 −0.9210.086 Expected quality q ∗ h 0.500.630.78 Service target / days2.01.51.0 Seasonal amplitude a h 0.100.080.12 Seasonal period / days73090 Macro sensitivity ζ h 1.21.00.7 Product match (A,B,C)(.45,.05,−.35)(0,.4, 0)(−.45, 0,.55) Channel effect (search, display, affiliate)(1.25,.70, 1.05)(1, 1.10,.90)(.85, 1.20, 1.15) All firms begin at $516,000 on this basis: $500,000 cash and $16,000 inventory salvage. Paid inbound inventory is included; brand, capacity, knowledge, and unfinished projects are not. Thus Score is finite-horizon liquidation-value gain, not accounting net income. The horizon creates a genuine incentive to avoid late investments whose returns fall outside the evaluation window. Market Score is the sum of company Scores, not total social welfare. D.8REFERENCE PARAMETERS AND THEIR RATIONALE Tables 14–17 specify the shared product and capacity settings, customer segments, market phases, and remaining dynamics used in all three evaluation groups. Functional-form references motivate the mechanisms; the numerical coefficients define the evaluation workload. For context, Bijmolt et al. (2005) report a mean price elasticity of −2.62, and Sethuraman et al. (2011) report a mean short-run advertising elasticity of 0.12. These observable elasticities are not estimates of our utility coefficients, and the formal eight-company scenario was not fitted to reproduce them. Appendix D.9 explains the scale relationships and the available validation. The initial phase is balanced. Daily self-transition probability is 59/60; each other phase has prob- ability 1/180, giving a 60-day mean duration rather than a fixed schedule. Mix weights are normal- 32 Preprint Table 16: Shared market phases. Mix entries are value/mainstream/premium; weight entries are price/quality/service multipliers. PhaseDemandMixWeightsQuality shift Balanced1.00(1, 1, 1)(1, 1, 1)0 Price pressure.90(1.5, 1, .7)(1.08, .90, 1)0 Quality growth1.10(.8, 1, 1.4)(.95, 1.25, 1).035 Delivery peak1.35(1, 1, 1)(.97, 1, 1.35)0 ized against baseline demand. These magnitudes and drift are design assumptions, not estimated business cycles. Table 17: Remaining dynamics and their design roles. Numerical values are frozen scenario settings. BlockReference valuesPurpose Marketingα = .45, A 0 = 1, b 0 = $100/day; δ G = .025, γ G = .08; initial awareness .20. Concave spending response, finite attention, and carryover; goodwill half-life about 27 days. Knowledgeδ K = .00045/day, γ K = .18, K 0 = $50, 000; daily development delay 7 days; capital multiplier 2.0. Slow depreciation and diminishing quality gains; spending cannot yield instant products. Project tiersCosts $10k/30k/75k; median durations 45/90/150 days; expected total capital $20k/70k/200k; log standard deviations .25 (duration), .20 (gain). Risky, delayed investment competing with current liquidity; gains are not cash payments. Reliabilityf 0 = .03, f min = .001, r 0 = $60/day; congestion threshold .75 and multiplier 2; $12 remediation/failure. Nonzero residual risk and a cost for operating near or beyond capacity. Support$300/day per capacity unit, 12 tickets/unit-day; η 0 = .015, η f = .20. A separate queue that requires recurring expenditure. Experienceρ = .04, ω − = 1.8; ω w = .15, ω f = ω u = 1; ω h = .12 for every segment; loyalty update .10. Persistent service effects and greater weight on disappointments; numerical scales are chosen, not fitted. Unfilled demandSwitching probability .76; consumer abandonment .05/day; no enterprise abandonment. A mix of lost demand and delayed obligations rather than guaranteed sales. Research tiersCosts $1k/2.5k/6k/12.5k/25k; delays 2/4/7/10/14 days; log-signal standard deviations .65/.40/.25/.15/.08. Explicit price–precision–latency trade-off. Initial beliefsIdentical segment priors: geometric means of true segment demand and price weights; coefficient of variation .80; quality mean .7, variance .16; weekly uncertainty increment (.03μ) 2 . Agents start with coarse aggregate scale, not each segment’s truth. Macro information ̄m = 1, κ = .02/day, σ m = .0016/day; publication delay 7 days. Persistent noise in a positive market index, with stale public information. Capacity reversalDownsize $1k per tier, 7-day delay; cancel expansion: 10% startup; fee up to 5% of prepaid capital, withheld from unspent funds. Adjustment is costly; no refund of installed capital. 33 Preprint BlockReference valuesPurpose (continued) Enterprise demandOpen every 14 days, bids due after 7; delivery due 7 days later; size 15–45% of baseline daily opportunities; reservation price 125% of initial product price; 10% unfilled-value penalty. Infrequent, capacity-constrained opportunities that use the same stock. The coefficients give segments different price, quality, service, product, and channel preferences. Seasonal periods of 7, 30, and 90 days and persistent market phases vary operating conditions within the 500-day horizon. These choices are explicit assumptions. Appendix D.9 gives their scale rela- tionships and validation limits; all values remain fixed across the three evaluation groups. D.9PARAMETER CHOICES AND VALIDATION SCOPE The main evaluation, repeated runs, and No Action comparisons use one frozen economic scenario. Parameter values specify a workload with delayed investment, scarce liquidity, imperfect informa- tion, and competition. Their exact levels are assumptions; the cited economic literature motivates the mechanisms and supplies broad response scales. The 500-day horizon contains 72 decision windows and exceeds three times the longest median innovation duration (150 days), allowing delayed investments and subsequent policy revisions within one run. Weekly decisions retain daily cash-flow dynamics while making policies persist between revisions. Eight firms place all eight LLM-based CEOs in one reference market; replacing one at a time preserves market size and the other seven competitors. Several ratios explain the chosen scales. Initial unit margins are 35%, 36%, and 37.04% for products A/B/C before operating expenses. Liquidation values are 30.77%, 31.25%, and 30.59% of procure- ment costs, so unsold inventory recovers only part of its purchase price. Annualized holding costs are 28.08%, 22.81%, and 25.76% of procurement value. Initial stock equals 20 days of base-tier throughput; this is a capacity ratio, not a forecast of actual sales. Procurement takes 5–10 days, while capacity expansion spans three to seven weekly windows. Median innovation delays consume 9–30% of the 500-day horizon. These relations create a trade-off between current spending and future operating capacity. Persistence operates at several scales. The retention terms for goodwill, satisfaction, and the mean- reverting macro process imply half-lives of approximately 27.4, 17.0, and 34.3 days. A fixed knowl- edge stock retains about 79.85% after 500 daily depreciation steps without new contributions, but receives no direct terminal liquidation value. Initial cash covers 1,250 days of base-tier overhead alone. Thus fixed overhead does not force exit within the horizon; discretionary spending and pro- curement can still exhaust liquidity. The 690 daily baseline opportunities are measured before the outside option and resource con- straints. They are not guaranteed purchases. This pool was retained when the formal market used eight rather than nine firms; it was not scaled by active-firm count. The nominal pool per firm there- fore rises from 76.67 in the nine-firm search to 86.25 in the eight-firm evaluation, a 12.5% increase before consumer choice. Fewer firms can ease rivalry for this pool, but the outside option, stock- outs, and different opponent policies prevent this ratio from predicting realized sales or Rule-relative Score. Consequently, the separate nine-firm Rule search does not validate the eight-firm competition intensity. Segment weights, outside utilities, seasonal amplitudes, and market-phase transitions de- fine contrasting demand regimes. The available evidence does not identify uniquely correct values for them. Model-free replay verified the 27 main-evaluation markets (13,500 daily transitions; 1,944 world decision windows) and the 26 repeated/baseline markets (13,000 transitions; 1,872 windows). To- gether these checks cover 53 final market states, 424 company scores, 26,500 daily transitions, and 3,816 world decision windows (Table 26; Appendices F.4 and F.6). These checks establish agree- ment between recorded decisions, transitions, and scores. Earlier response diagnostics used a differ- ent scenario and are not evidence of parameter validity for these evaluation runs. The Rule search 34 Preprint Table 18: Terminal-valuation sensitivity. Means over 24 main-evaluation outcomes per CEO (USD thousands). Salvage includes on-hand and paid inbound inventory; ranks are identical under both measures. RankCEO agentOfficial ScoreSalvageCash-only Score 1GPT-6 Astra116.941.09115.85 2GPT-5.6 Sol98.441.3697.08 3Qwen 3.8 Max8.042.635.41 4Rule CEO−47.304.00−51.30 5DeepSeek V4 Flash −114.509.70−124.21 6Gemini 3.8 Flash−119.795.03−124.81 7DeepSeek V4 Pro −159.335.42−164.75 8Claude Sonnet 5−314.3615.86−330.22 9Doubao Seed 1.8−434.8812.02−446.90 selects a baseline policy, not the market coefficients. No parameter sweep in these agent evaluations establishes ranking robustness to demand, liquidity, delays, or noise. Appendix E separately as- sesses terminal-valuation sensitivity by rescoring the recorded outcomes without inventory salvage. This holds decisions and trajectories fixed; robustness under a changed reward rule or horizon would require new agent runs. EBEHAVIOR AND TRAJECTORY EVIDENCE Cases T1–T3 use committed decisions, environment receipts, and daily accounts from the formal campaign; their amounts are in dollars. Selected after observing outcomes, they illustrate operating mechanisms rather than estimate an action’s causal contribution to ranking. T1: Terminal inventory drawdown. Astra and Sol finish with mean cash of $631,851.79 and $613,078.14, but mean inventory salvage of only $1,090.25 and $1,363.75. Their positive Mean Scores therefore persist if terminal salvage is omitted from Equation 2. In Astra’s lower-median outcome (seed 29, company 1), the day-490 decision sets all reorder targets to zero, effective day 491. Over days 491–500, procurement is zero, revenue is $39,775.00, and cash costs are $5,839.83. Cash rises from $549,475.53 to $583,410.70. The remaining 53 units have $817.00 salvage value, with no paid inbound stock; final Score is $68,227.70. Astra also has positive Score in all 11 appearances in markets with no bankrupt firm. Terminal-valuation sensitivity. For each of the 216 main-evaluation outcomes, we remove ter- minal inventory salvage while retaining the recorded decisions and cash flows: S cash i = S i − salvage i,T = C i,T − V i,0 . Table 18 shows that all nine mean-Score ranks and signs are unchanged. Qwen’s cash-only mean remains positive but small ($5.41k). Cash-only rescoring retains all eight market and all 56 directed three-seed mean signs (Table 19). Astra’s market difference remains positive in all three seeds; Sol’s private difference remains pos- itive in all three (mean $134.86k), while its market difference remains negative on average (mean −$48.99k; two negative seeds). This post-hoc check does not test agent behavior under a different reward or horizon; cash from late inventory sales remains counted. T2: Procurement exhausts liquidity. Doubao (seed 11, company 8) sets reorder targets for A and C to 10,000 units on day 105. Applied receipts precede $469,372.00 of procurement on day 106, when cash falls from $467,094.87 to $57.19. Further procurement and operating costs reduce cash to−$0.93 on day 110, before the next planning boundary at day 112. All 16 decisions were sealed. The firm exits despite holding inventory with $151,610.00 terminal salvage value. T3: Recurring budgets exhaust cash. DeepSeek Pro (seed 29, company 2) sets total daily de- velopment to $75,000.00 and total daily marketing, including unchanged allocations, to $63,450.00 on day 35. Both changes take effect on day 36 and are charged on days 36–38. Cash falls from $339,084.34 to $203,460.46, $67,030.87, and −$68, 317.92, respectively. Bankruptcy occurs on day 38, before the next decision boundary at day 42. All six decisions were sealed. 35 Preprint Table 19: Matched differences without terminal salvage. Three-seed market means (USD thou- sands), seed signs in order 11/29/47, and retained directed signs under cash-only rescoring. Each entrant has seven directed means and 21 seed-level contrasts. Market difference Market seed signs Directed sign matches EntrantOfficial Cash-onlyMeanSeed-level GPT-6 Astra+466.68 +475.40 (+, +, +)7/720/21 GPT-5.6 Sol−89.66 −48.99 (−, +,−)7/720/21 Qwen 3.8 Max+156.92 +183.76 (+, +,−)7/721/21 DeepSeek V4 Flash −84.20 −79.99 (+, +,−)7/720/21 Gemini 3.8 Flash+55.78+62.37 (+, +,−)7/720/21 DeepSeek V4 Pro −573.25 −528.21 (−,−,−)7/721/21 Claude Sonnet 5 −458.05 −457.64 (−,−,−)7/721/21 Doubao Seed 1.8 −639.88 −642.88 (−,−,−)7/721/21 Every market at seed 11 cash on hand for the eight firms of each market: the all-LLM reference market, then Rule CEO taking one seat at a time $0 $200k $400k $600k $800k 0250500 All-LLM reference market Day $0 $200k $400k $600k $800k 0250500 Rule replaces GPT-6 Astra Day $0 $200k $400k $600k $800k 0250500 Rule replaces GPT-5.6 Sol Day $0 $200k $400k $600k $800k 0250500 Rule replaces Qwen 3.8 Max Day $0 $200k $400k $600k $800k 0250500 Rule replaces DeepSeek V4 Flash Day $0 $200k $400k $600k $800k 0250500 Rule replaces Gemini 3.8 Flash Day $0 $200k $400k $600k $800k 0250500 Rule replaces DeepSeek V4 Pro Day $0 $200k $400k $600k $800k 0250500 Rule replaces Claude Sonnet 5 Day $0 $200k $400k $600k $800k 0250500 Rule replaces Doubao Seed 1.8 Day GPT-6 AstraGPT-5.6 SolQwen 3.8 MaxDeepSeek V4 Flash Gemini 3.8 FlashDeepSeek V4 ProClaude Sonnet 5Doubao Seed 1.8Rule CEOfirm exits Figure 9: Cash on hand in all nine main-evaluation markets at seed 11. The upper-left panel is the eight-agent reference market; the other panels replace the named agent with Rule CEO. Colors identify agents, the dashed brown line identifies Rule, and skull markers indicate firm exit days. Traceability. anc/data/formal/trajectory_summary.json maps T1–T3 to run IDs, checkpoint and trace paths, SHA-256 hashes, decision windows, receipts, and accounts. The run/agent-trace hash prefixes are ef03e110/5f49ac152640, 665f7924/4e4b8ccde867, and 5ac797e6/4678d3095161. Cash paths are accounting records, not interim official Scores. These cases alone cannot separate the effects of rival exit, market changes, and other firms’ simul- taneous adaptations. Cash paths by market. Figures 9–11 show the 27 main-evaluation markets by seed and lineup. Each panel contains eight firms: the reference market appears first, followed by eight same-seat Rule replacements. Figure 4 groups the same trajectories by agent. These are cash balances, not interim official Scores; the repeated and No Action runs are excluded. 36 Preprint Every market at seed 29 cash on hand for the eight firms of each market: the all-LLM reference market, then Rule CEO taking one seat at a time $0 $200k $400k $600k $800k 0250500 All-LLM reference market Day $0 $200k $400k $600k $800k 0250500 Rule replaces GPT-6 Astra Day $0 $200k $400k $600k $800k 0250500 Rule replaces GPT-5.6 Sol Day $0 $200k $400k $600k $800k 0250500 Rule replaces Qwen 3.8 Max Day $0 $200k $400k $600k $800k 0250500 Rule replaces DeepSeek V4 Flash Day $0 $200k $400k $600k $800k 0250500 Rule replaces Gemini 3.8 Flash Day $0 $200k $400k $600k $800k 0250500 Rule replaces DeepSeek V4 Pro Day $0 $200k $400k $600k $800k 0250500 Rule replaces Claude Sonnet 5 Day $0 $200k $400k $600k $800k 0250500 Rule replaces Doubao Seed 1.8 Day GPT-6 AstraGPT-5.6 SolQwen 3.8 MaxDeepSeek V4 Flash Gemini 3.8 FlashDeepSeek V4 ProClaude Sonnet 5Doubao Seed 1.8Rule CEOfirm exits Figure 10: Cash on hand in all nine main-evaluation markets at seed 29. Panel order and visual encoding match Figure 9: the reference market precedes eight same-seat Rule replacements, and skull markers indicate firm exit days. Cash outlays. Figure 12 summarizes gross cash spending across each candidate’s 24 main- evaluation outcomes. Procurement converts cash into inventory; its contribution to Score depends on subsequent sales and terminal salvage. Behavior and Score. Figure 13 juxtaposes recorded behavior with Mean Score. The panels are descriptive: inquiry and note measures come from one seed-11 reference market, while Score aver- ages 24 outcomes. Reference-market memory. Figures 14 and 15 show selected notes and events in the seed-11 reference lineup. Written plans need not become executed decisions. E.1MECHANISM HYPOTHESES FROM MEMORY AND ACTIONS Scope and procedure. We analyze the 27 main-evaluation markets at seeds 11, 29, and 47. Ac- counting profiles cover all 24 outcomes per agent. The memory audit extracts 2,477 recorded decision-window traces from 36 company trajectories in 12 markets: the three reference markets and the matched Rule replacements of Astra, Sol, and DeepSeek Pro. It covers the three leading agents’ reference trajectories and both affected-agent histories for each of the four pairs in Fig- ure 6D. Qwen loses in all three reference markets, so we also inspect the lower-median positive Qwen outcome within each seed: Rule replaces Pro at seeds 11 and 47, and Astra at seed 29. These selected Scores are $225.78k, $66.69k, and $32.81k, respectively; seed 11 has only one positive Qwen outcome. We recover auto-loaded notes and successful file writes, verify each source trace against its archived SHA-256, and link claim-relevant excerpts to sealed decisions and settlement receipts. MEMORY.md is persistent, agent-written material; its contents are beliefs and plans, not verified descriptions of the simulator. A written plan is counted as executed only when the decision and receipt agree. The four pairs were selected by the preceding stability analysis, not by their fit to a mechanism. The following comparisons generate hypotheses; other agents also adapt, and independent trajectories can diverge before any observed rival response. 37 Preprint Every market at seed 47 cash on hand for the eight firms of each market: the all-LLM reference market, then Rule CEO taking one seat at a time $0 $200k $400k $600k $800k 0250500 All-LLM reference market Day $0 $200k $400k $600k $800k 0250500 Rule replaces GPT-6 Astra Day $0 $200k $400k $600k $800k 0250500 Rule replaces GPT-5.6 Sol Day $0 $200k $400k $600k $800k 0250500 Rule replaces Qwen 3.8 Max Day $0 $200k $400k $600k $800k 0250500 Rule replaces DeepSeek V4 Flash Day $0 $200k $400k $600k $800k 0250500 Rule replaces Gemini 3.8 Flash Day $0 $200k $400k $600k $800k 0250500 Rule replaces DeepSeek V4 Pro Day $0 $200k $400k $600k $800k 0250500 Rule replaces Claude Sonnet 5 Day $0 $200k $400k $600k $800k 0250500 Rule replaces Doubao Seed 1.8 Day GPT-6 AstraGPT-5.6 SolQwen 3.8 MaxDeepSeek V4 Flash Gemini 3.8 FlashDeepSeek V4 ProClaude Sonnet 5Doubao Seed 1.8Rule CEOfirm exits Figure 11: Cash on hand in all nine main-evaluation markets at seed 47. Panel order and visual encoding match Figure 9: the reference market precedes eight same-seat Rule replacements, and skull markers indicate firm exit days. Where the money went share of all cash spent, pooled over each candidate's 24 runs 0%25%50%75%100% GPT-6 Astra 66%17%9% $1333.8k GPT-5.6 Sol 64%16%8% $1739.2k Qwen 3.8 Max 61%20% $1268.2k Rule 55%31% $650.5k DeepSeek V4 Flash 59%18%12% $1523.2k Gemini 3.8 Flash 56%22%8%10% $1144.5k DeepSeek V4 Pro 62%13%8%11% $2818.0k Claude Sonnet 5 56%13%10%10% $2459.6k Doubao Seed 1.8 54%13%11%16% $2143.5k spend per run ProcurementCapacityMarketingProduct and R&DServiceMarket researchHolding and penalties Figure 12: Cash outlays in the main evaluation. Bars pool each CEO agent’s 24 outcomes and show the share spent by category; right-hand values are total spending divided by 24 (USD thou- sands). These are gross cash outflows, not Score contributions. Positive mean returns. Table 20 separates revenue from net cash outlays. The accounts reconcile exactly for all 216 outcomes: final cash equals initial cash plus revenue minus outlays net of refunds. Score then adds terminal salvage and subtracts initial enterprise value; equivalently, in dollars, S = R−P +L− 16,000, where P is net cash outlay and L is terminal salvage. Thus sales volume alone does not explain the ranking. Astra’s notes repeatedly compare inventory positions and capacity with realized demand and the remaining horizon. The corresponding actions reduce replenishment targets, cancel unused capac- ity, and reverse an expansion (Table 21); its mean marketing and capacity-investment outlays are $35.91k and $21.00k. Sol supports higher volume with $81.67k and $49.25k in these categories, but also reverses capacity commitments when utilization weakens. Their positive means are consistent with different balances between sales and persistent commitments, rather than sales maximization alone. Astra is positive in 24/24 outcomes and Sol in 23/24; neither goes bankrupt. 38 Preprint (a) Windows that never closed share of decision windows left unsealed, 24 runs -400k -300k -200k -100k 0 +100k 0%5%10%15%20%25% Windows left unsealed Me an S c o r e ( U S D t h o u s an d s ) (b) Reading before deciding SQL queries per window, seed-11 market -400k -300k -200k -100k 0 +100k 020406080 Queries per window one market, one seed (c) Contingencies written down if-then statements per note, seed-11 market -400k -300k -200k -100k 0 +100k 0.00.51.01.52.0 If-then statements per note one market, one seed Figure 13: Recorded behavior and operating returns. Vertical positions are main-evaluation Mean Scores over 24 outcomes. In (a), horizontal values are the percentage of active decision windows left unsealed across those outcomes. In (b), queries per window come from the seed-11 reference market; (c) counts if–then statements per note for agents with notes in that market. These comparisons do not identify causal effects. Objective: maximise terminal cash + inventory liquidation − initial enterprise value. Horizon day 500; 416 days remain. d84 Day 133 set_product applied 134: B restored ... previous memory ended 126 and missed this decision. d140 Day 211 procurement payment 2,303,100 cents caused the recent cash dip. d217 Days 267–273: 173 units fulfilled, no abandonment; recurring policy expense 55,500/day. d273 A 4,200 test reduced A contribution versus the 4,000 baseline despite the higher unit price. d329 Inventory A0/B53/C0; consumer backlog 2 A; support backlog 1. d385 Tier 1 costs 1,200,000 and takes 21 days: no expansion is justified by current demand. d441 Knowledge and capacity have no separate terminal value; recurring policies persist until changed. d500 (a) GPT-6 Astrasurvives, +$103.1k Day 21 cash 45,904,634c. Value segment ↔ Product A: satisfaction positive and rising. Best-fit pair so far. d21 Day 63 cash 36,142,584c. If day 64–70 data doesn't show recovery, try a price-cut test ... or fresh market research. d63 Day 91 cash 32,193,302c. NO operating changes staged — pure hold. d91 Bid 5,800c/unit on open tender B qty 222 (deadline 119, delivery 126). d112 Day 140 cash 26,582,079c. Bid 5,900c on B tender; started TIER 3. d140 Day 315 cash 1,456,567c. 185 days remain. No open tenders this window. d315 CRITICAL RULE: “projects, knowledge, and capacity have no separate terminal score value.” d336 Day 364 cash 974,105c. Bid on tender dec5d5f8... (B, qty 251). Exits on day 371. d364 (b) Claude Sonnet 5exits day 371, −$487.9k Product A: price 3,800c, promotion 3,500c (valid, < price), reorder target 400. d7 Product A reorder target increased to 500 (from 150) to eliminate stockouts. d21 Products A and B are now stockout (inventory 0) due to strong demand; capacity expansion to Tier 1 completed. d42 Days 113–119 net cash flow −1,547,881c; cash 21,759,485c. d119 Reduced A price to match company-7. Cash 15,229,586c; net cash flow −223,294c. d154 Cash 11,877,708c; A inventory 0, backlog 1; Tier 2 expansion pending (day 203). d182 Cut all non-essential recurring costs: stopped support and mainstream-search budgets to reduce daily burn. d301 Cash falls below zero; the company exits with 196 days still to run. d304 (c) Doubao Seed 1.8exits day 304, −$515.6k 0125250375500 Figure 14: Selected memory and operating events. Astra, Sonnet, and Doubao in the seed-11 reference lineup; badges report terminal Score in USD thousands, while cash values marked c in excerpts are cents. This Doubao firm exits on day 304; case T2 above describes a different seed-11 replacement lineup whose Doubao firm exits on day 110. Qwen’s notes specify thresholds for reverting price or budget changes, and several reversions are executed. Yet this practice is also present in its losing reference trajectories. Its 11 positive out- comes have much higher mean fulfillment than its 13 negative outcomes despite similar marketing expenditure. Conditional planning therefore provides a plausible means of adapting spending and margins, but does not by itself explain the sign of returns. The small pooled surplus reflects uneven realized demand and profitability, not a uniformly successful policy. Why aggregate externalities differ. Table 22 decomposes the matched changes for Astra and Sol. Both attract more consumer orders than the Rule company they replace; Sol’s increment is larger in every seed, alongside a larger increase in marketing outlays ($16.76k/$34.64k/$83.15k 39 Preprint Claude Sonnet 5 A diagnostic ladder Day 63 “If day 64–70 data doesn't show recovery, next window try a price cut” Day 70 “If C price cut shows no effect either, price is likely NOT the constraint” Day 98 “If bids keep losing despite going below consumer price, quality gating rather than price may be the binding constraint” verbatim, seed 11 GPT-5.6 Sol Both branches of a bid priced in advance Day 28 “If won, verify reserved inventory and preserve capacity through day 42; if lost, lower the C target promptly and downsize tier 1” Day 224 “Do not expand: tier 1 would arrive day 252 if ordered then, after the due date” Day 350 “If the bid loses, current C demand should absorb the extra inventory in about 25 days” verbatim, seed 11 Qwen 3.8 Max Falsifiable thresholds Day 28 “If units < 300, the model is wrong — fall back to week-2 economics and cut marketing” Day 42 “If our bid was lowest and we still lost, we FAILED the quality gate” Day 49 “If a segment starts abandoning, capacity is binding ⇒ raise A price first, then B and C” verbatim, seed 11 How often, per note if-then statements ÷ notes written Qwen 3.8 Max 1.70117/69 Claude Sonnet 5 1.1227/24 GPT-6 Astra 0.5027/54 GPT-5.6 Sol 0.3423/68 Doubao Seed 1.8 0.000/18 Doubao wrote 18 notes and no contingency at all, and left the market on day 304. Figure 15: Conditional plans in private notes. Selected seed-11 reference-market excerpts and if– then statements per written note for agents with notes. These counts describe text, not plan execution or decision quality. Table 20: Accounting profiles of the three agents with positive Mean Score. Money is in USD thousands; all entries except n are means. Net outlays subtract recorded refunds. The last two rows condition on Qwen’s realized Score and are descriptive, not separate experimental groups. Agent/sample nScore Revenue Net outlays Fulfilled units Marketing Astra24116.94 1,460.951,329.1023,25235.91 Sol2498.44 1,839.241,726.1629,76681.67 Qwen248.04 1,279.921,258.5220,95485.22 Qwen: S > 0 1179.25 1,501.791,409.4724,24684.46 Qwen: S < 0 13 −52.21 1,092.191,130.7918,17085.86 versus Astra’s $9.52k/$10.34k/$19.20k). This is consistent with stronger demand competition, but aggregate outcomes also depend on rivals’ spending and survival. Averaged across seeds, Astra’s rivals lose $436.89k in revenue but reduce net outlays by $792.84k, yielding a positive Score difference after salvage. Sol’s rivals lose $1,066.09k in revenue and save only $882.24k in net outlays. The mean contrast therefore concerns both revenue pressure and spending responses. It is not a claim that Astra raises every rival’s sales or that one expenditure channel explains every seed: at seed 47, rival revenue increases under both entrants, but outlays rise much more under Sol. Four directed relationships. Table 23 reports all three main-seed decompositions for the four selected pairs. For a rival in a matched comparison, ∆S = ∆R− ∆P + ∆L; initial endowments cancel. We examine both accounting channels and explicit strategy revisions. Pro → Doubao: price-following and margin pressure. Doubao explicitly responds to company 2 (DeepSeek Pro) in all three reference trajectories. At seed 11/day 84, its note says matching company-2 to boost demand; it raises C’s per-unit promotion from $5 to $10, matching Pro’s $120 effective price. At seed 29/day 98, it cuts C from $130 to Pro’s $125. At seed 47/day 28, it matches Pro’s B/C list prices of $72/$130 but also adds $5/$10 promotions, producing lower effective prices of $67/$120. Sealed decisions and receipts confirm these changes. Matching prices, sometimes with additional discounts, is a plausible source of margin pressure. Doubao’s revenue and gross profit are lower in every seed when comparisons stop before either matched firm exits. However, the seed-11 terminal Score difference is mostly a $150.81k salvage difference: the Rule-replacement Doubao exits on day 110 with substantial stock, versus day 304 in the reference market. Price-following does not explain this entire terminal contrast. Sol → Sonnet: costly investment alongside weaker returns. Sonnet spends more on development plus innovation with Sol in each seed. Its notes favor growth through quality and capacity: seed 11/day 42 projects demand above 90 units/day and seals tier-2 expansion; seed 47/day 70 starts another innovation project while gross profit remains below recurring costs. This supports an investment-response hypothesis. Lower full-horizon sales alone are not a sufficient explanation: at seed 11, Sonnet exits on day 371 only in the reference market. Through day 370, it earns $307.10k more revenue, but its additional non-procurement outlays exceed its extra gross profit. At seeds 29 40 Preprint Table 21: Memory statements linked to executed decisions. Entries identify agent, seed, and decision day. All use the reference market except Qwen: Rule replaces Pro at seeds 11/47 and Astra at seed 29. Excerpts omit Markdown emphasis and use cents; action amounts are converted to dollars. Receipts confirm execution, not the action’s causal payoff. Agent; seed/dayNote excerptSealed action and settlement Astra; 11/7Reductions pause future procurementA/B/C reorder targets become 350/160/120 on day 8. Astra; 29/147Tier1 is unnecessary for observed consumer volume Downsize from tier 1 to 0, completed on day 154; C target falls from 400 to 80. Astra; 47/7Base capacity has 25 units/day spareCancel pending tier-1 expansion; the receipt records a $6,600 refund. Sol; 11/126a lead-time hedge for tier-3 innova- tion Expand from tier 1 to 2, due day 161; later downsize on days 182 and 224. Sol; 29/231Don’t keep unused capacity just for terminal value Downsize from tier 1 to 0, completed on day 238; A target falls from 583 to 300. Sol; 47/7Keep all nine segment/channel mar- keting budgets at 5k/day Retain total marketing of $450/day while cancelling pending tier-2 capacity. Qwen; 11/224SUPPORT=0 ALSO REVERSEDRestore marketing and support to $40/day each, effective day 225. Qwen; 29/56slash developmentReduce total development from $95 to $15/day, effective day 57; restore service budgets. Qwen; 47/350A PRICE TEST RESOLVEDReverse A’s price from $40 to $47, effective day 351, after its stated volume threshold is missed. Table 22: Accounting decomposition of aggregate externalities. Each row compares the reference market with the same-seed Rule replacement of the entrant. Rival columns sum the other seven firms; monetary differences are in USD thousands. Net outlays subtract refunds. Own consumerOther seven firms Entrant Seed order difference∆Revenue ∆Net outlays ∆Salvage∆Score Astra11+6, 172 −1, 982.84 −2, 324.11+24.60 +365.87 Astra29+6, 221 −199.92 −749.27 −30.79 +518.56 Astra47+7, 179+872.10+694.86 −10.68 +166.55 Sol11+16, 399 −3, 009.41 −2, 707.13 −30.28 −332.57 Sol29+15, 539 −483.51 −759.61 −79.35 +196.75 Sol47+13, 040+294.66+820.03 −2.47 −527.84 and 47, it instead loses revenue. The negative contrasts therefore combine spending and demand channels. Astra→ Flash: lower investment burden. Flash’s innovation spending is $285k, $150k, and $320k lower with Astra at seeds 11, 29, and 47; total non-procurement outlays net of refunds are also lower in all three. Its seed-29/day-28 note rejects excess capacity at demand of 35 units/day, and a sealed downsize reaches tier 0 on day 35. Spending restraint is consistent with the positive Score contrasts even though revenue declines in two seeds. The evidence does not establish that Astra alone induced that restraint. Astra → Pro: different routes to a positive contrast. At seed 11, Pro’s lower outlays more than offset its lower revenue; its day-21 note rejects tier-2 capacity without demand growth, and that window seals no expansion. At seeds 29 and 47, Pro instead fulfills 25,965 and 20,913 more units, respectively, with revenue gains exceeding net-outlay increases. These observations suggest spend- ing adjustment in one seed and sales gains exceeding added outlays in the other two. They support a positive matched contrast through several operating paths, not a single identified bilateral mecha- nism. 41 Preprint Table 23: Main-evaluation accounting for the four directed pairs. Differences are reference minus matched Rule-replacement outcomes for the affected agent, in USD thousands. These three- seed observations provide the mechanism evidence; the five-observation stability summary remains in Figure 6D. Entrant→ rival Seed∆Revenue ∆Net outlays ∆Salvage∆Score Pro→ Doubao11+549.63+550.01 −150.81 −151.19 29 −984.90 −929.82+0.02 −55.05 47 −999.76 −940.06 −4.92 −64.62 Sol→ Sonnet11 −169.19+198.26+27.30 −340.16 29 −324.02 −245.91+5.09 −73.02 47 −238.33 −66.83 −2.31 −173.80 Astra→ Flash11 −1, 297.37 −1, 372.24+0.14+75.01 29+271.72 −20.88+0.75 +293.35 47 −317.84 −614.25 −6.12 +290.29 Astra→ Pro11 −422.11 −652.91+2.54 +233.33 29+1, 301.46+1, 145.73 −6.21 +149.53 47+1, 636.91+1, 604.48 −2.83+29.61 FEXECUTION AND REPRODUCIBILITY F.1SETTLEMENT AND DURABLE STATE The coordinator collects decisions against one immutable world version, journals accepted re- sponses, and commits a complete barrier before advancing the economy. Standing policies execute daily inside each window. Runs stop at day 500; a surviving company participates in 72 windows. Standalone runs atomically publish JSON snapshots and journal responses in SQLite WAL with synchronous=FULL; campaign runs use transactional SQLite storage. Recovery checks the run, configuration, and window identities before reusing receipts. Locks and revision-checked leases prevent concurrent ownership. Ordinary continuation retains the conversation, draft, private files, and consumed reply/token/time allowances. Automatic retry waits consume active time; queueing and manual pauses do not. A request without a durably received response may still incur provider work or billing. F.2RETRIES AND FRESH ATTEMPTS In the main evaluation, transient provider failures retry within the current attempt’s allowance. Ordi- nary HTTP 400 failures permit three consecutive attempts for one logical request, including the first. The first two failures wait one and two seconds by default, respecting a longer Retry-After. A valid accounting-bearing reply resets the consecutive count. Counters persist across pause/resume; the third consecutive failure stops the affected run. Authentication and identity errors require reso- lution. A separate supervisor may restart a failed, unsealed company-window up to three times beyond its original attempt. It archives the failed conversation, draft, allowance, trace, and timing records; restores private files to their window-start contents from verified receipts; and grants the original allowance. Earlier economic windows and accepted peer decisions remain intact. The count is shared across fault categories and persists through restarts. Budget exhaustion, running or backing- off attempts, and sealed or settled decisions are ineligible. Missing restoration evidence prevents a fresh attempt. The formal campaign records three such operations, each at fresh-attempt index 1: DeepSeek V4 Flash at seed 47, days 315 and 441, and Qwen at seed 47, day 392. Their failed-attempt usage is retained separately. Neither settled usage nor the sum of known settled and discarded usage is a complete provider bill. Table 24 distinguishes decision outcomes from request-level failures and states their economic treatment. 42 Preprint Table 24: Decision outcomes and their economic treatment. StatusEconomic treatmentReporting consequence SealedExecute the validated draft at the barrier; an unchanged-policy seal is valid. Count a successful decision and its reported usage. Budgetex- hausted Discard unsealed changes; standing policies continue and the barrier may settle. Record budget_exhausted, separately from a deliberate unchanged-policy seal. TimeoutExhausted infrastructure recovery applies no unsealed changes; standing policies continue. Record timeout, failure metadata and the infrastructure-timeout count. Protocol errorCorrect recoverable replies in-session; a final invalid Harness result applies no draft. Retain economic consequences and correction diagnostics separately. Provider errorRetry eligible failures; terminal failures stop the affected run at its durable frontier. Apply the bounded supervision rule where eligible; retain all attempt costs. Pause/in- tegrity error Preserve receipts; do not settle an incomplete or inconsistent barrier. Resolve before accepting a completed result. F.3CONCURRENCY AND PROVENANCE In the main evaluation, the seed-priority scheduler admits at most 30 active runs. Earlier eligible seeds have priority; later seeds may fill available slots. Active CEOs within a run decide concur- rently, with no additional global decision or resident-worker cap. Runs share compute but not eco- nomic state. Latency can affect how much work fits within a decision allowance, while settlement waits for the common barrier. The main-evaluation manifest binds the scenario, seeds, seats, lineups, Rule profile, prompt, tool schemas, allowances, memory policy, provider settings, and source fingerprints. The main- evaluation release is http400-three-attempts-20260913: package 0.15.2, economy ceos-market-v3, decision ceos-decision-v9, runtime ceos-runtime-v4, and cam- paign ceos-campaign-plan-v6. Per-run records retain execution provenance. The main- evaluation runtime uses Python 3.13.15, Bun 1.3.14, NumPy 2.5.2, Pydantic 2.13.4, and HTTPX 0.28.1; OpenCode’s revision begins 1be9fd55a932. No Action’s separate implementation is specified in Appendix B.6; it does not change these formal records. F.4COMPLETION AND DECISION AUDIT The 15 September 2026 audit verified 27 completed runs, eight companies per run, and 24 termi- nal outcomes per candidate. Model-free Exact Replay reconstructed 13,500 daily transitions and 1,944 windows and reproduced final states and scores. Checkpoint/trace hashes and exported states matched; SQLite integrity was ok, and the published analysis was recomputed. This checks execu- tion conditional on saved decisions, not the reproducibility of new LLM generations. The 15,078 settled company-window records comprise 14,362 seals, 683 budget exhaustions, 32 timeouts, and one protocol error. Rule contributes 1,728 seals; Decision health for LLM-based agents uses 13,350 active company-windows. The 108 failure-history records are execution events, not distinct failed decisions. Exited firms submit no further decisions but remain in terminal-score denominators. Table 25 reports final decision outcomes separately for the main evaluation, repeated runs, and No Action replacements. Across all 53 runs, 29,748 active company-windows comprise 28,477 seals, 1,215 budget exhaustions, 54 timeouts, and two protocol errors. The unsealed-decision rate divides non-sealed company-windows by all active company-windows. It includes budget exhaustion, timeout, and final protocol-error outcomes. The source field is failure_rate; it is not an API failure rate. 43 Preprint Table 25: Final decision status by CEO agent and experiment group. Active company-windows count decisions while the firm remains active; the status columns sum to that count. All settled decision records are included. CEO agent Active company- windows Sealed Budget exhausted Timeout Protocol error (a) Main evaluation: 27 runs gpt-6-astra1,7281,715940 deepseek-v4-pro1,5561,556000 claude-sonnet-51,6801,273397100 qwen3.8-max1,7281,7072100 gpt-5.6-sol1,7281,69127100 gemini-3.8-flash1,7281,60512300 deepseek-v4-flash1,7051,6841281 doubao-seed-1.81,4971,4039400 Rule CEO1,7281,728000 Group total15,078 14,362683321 (b) Repeated runs: 18 runs gpt-6-astra1,1521,138860 deepseek-v4-pro1,1171,117000 claude-sonnet-51,1521,04310720 qwen3.8-max1,1521,144800 gpt-5.6-sol1,1521,08557100 gemini-3.8-flash1,1521,05210000 deepseek-v4-flash1,1521,144701 doubao-seed-1.89528718100 Rule CEO1,1521,152000 Group total10,1339,746368181 (c) No Action baseline: 8 runs gpt-6-astra504501030 deepseek-v4-pro479477200 claude-sonnet-55044564710 qwen3.8-max504501300 gpt-5.6-sol5044634100 gemini-3.8-flash5044663800 deepseek-v4-flash504499500 doubao-seed-1.84584302800 No Action CEO576576000 Group total4,5374,36916440 All 53 runs29,748 28,4771,215542 Figure 16 summarizes settings written and decision closure in the 27-run main evaluation. Table 26 lists the archived evidence for both evaluation campaigns and their combined verification totals. F.5ANALYSIS SCOPE The main-evaluation results use all 27 completed runs. Repeated and baseline comparisons use the separate weights defined in Appendices A.2 and A.3. Case traces describe recorded behavior; they do not isolate an action’s causal contribution. Parameter robustness requires separate experiments with the same model and harness settings. Replaying old decisions under new parameters would test an unchanged policy sequence, not agents’ adaptation to the new setting. F.6REPEATED RUNS AND BASELINE DATA The 18 repeated active-Rule runs and eight No Action replacements all completed day 500. They yield 208 terminal company scores, for 424 across all 53 evaluation markets. Their separate ver- ification records contain 26 model-free replay receipts, 13,000 daily transitions, and 1,872 world 44 Preprint What each CEO changed, and whether the window closed settings written per decision window, and the outcome of the window itself; 24 runs each 0481216 settings changed per windowwindow outcome GPT-6 Astra3.00.8% unsealed GPT-5.6 Sol2.22.1% unsealed Qwen 3.8 Max3.41.2% unsealed Rule15.00.0% unsealed DeepSeek V4 Flash2.31.2% unsealed Gemini 3.8 Flash0.77.1% unsealed DeepSeek V4 Pro2.70.0% unsealed Claude Sonnet 51.524.2% unsealed Doubao Seed 1.83.96.3% unsealed Price and promotionStock targetsMarketing budgetsProduct and R&DService budgetsCapacityEnterprise bidsMarket research sealedallowance exhaustedtimeout or protocol error Figure 16: Recorded actions and window outcomes. Left: mean settings written per active de- cision window across each CEO’s 24 main-evaluation outcomes, grouped by policy type. Right: sealed versus unsealed windows; unsealed includes budget exhaustion, timeout, and protocol error. Unsealed drafts do not change standing policies. Table 26: Retained evidence for the 53 evaluation runs. Main-evaluation and repeated/baseline records are archived separately; full hashes accompany the research artifacts. ArtifactEvidence Main-evaluation manifestPlan hash prefix 3e34a22e881; scenario hash prefix 893815172625; model controls, seats and economic configuration. Repeated/baseline recordsSeparate campaign records, final checkpoints, terminal analysis and verification receipts for the 18 repeated and eight No Action runs; see Appendix F.6. Main-evaluation replay27 verified run receipts, 216 terminal company scores, 13,500 daily transitions and 1,944 world decision windows. Repeated/baseline replay26 verified run receipts, 208 terminal company scores, 13,000 daily transitions and 1,872 world decision windows. Evaluation total53 verified markets, 424 terminal company scores, 26,500 daily transitions and 3,816 world decision windows. Decision healthPer-agent status counts for all three experiment groups in Table 25; 29,748 active company-windows in total. Recorded workloadSuccessfully sealed decisions with complete reported usage, summarized in Appendix F.7; per-trajectory counts in anc/data/workload. Model identificationRequested/served identifiers and explicit accepted aliases retained in the run records; no independent weight verification. decision windows. Analysis recomputed from terminal results matches the archived output. Ta- ble 25 gives the per-agent outcomes for these runs, and Table 26 combines the verification counts. Replay verifies settlement for recorded decisions; it is not a new agent rollout. These runs used updated recovery for unsealed windows that retained cumulative usage. Cumulative output-token limits were disabled for Astra and Sol; per-response ceilings, economic parameters, and scoring rules were unchanged. The paper’s data directory separates the three-seed main evaluation (anc/data/formal) from the additional results (anc/data/supplementary).The latter contains the 26-run ter- minal analysis and simple-five-observation-results.json, including all 56 five- observation contrasts and the three-control baseline comparisons. Raw amounts are cents; displayed USD thousands divide by 100,000. Figure 6A–B and Tables 2–3 retain the main-evaluation esti- mates. Figure 6C–D use the separate weights defined in Appendix A. F.7EXPERIMENTAL SCALE AND RECORDED WORKLOAD Table 27 accounts for the three evaluation groups and the separate Rule search. A CEO decision window is one opportunity to revise company policies; a surviving firm participates in 72 windows. 45 Preprint Table 27: Experimental scale. Simulation days count market-level transitions. Shared controls are counted once; replay does not add new experimental runs. ExperimentMarkets Firms per market Company trajectories Simulation days CEO decision windows Main evaluation27821613,50015,078 Repeated runs1881449,00010,133 No Action baseline88644,0004,537 Evaluation total53842426,50029,748 Rule parameter search75967537,50048,600 Table 28: Recorded workload across the 53 evaluation runs. Counted windows are successfully sealed decisions with complete reported usage. Model replies per trajectory are the mean± sample SD of the counted replies across each agent’s company trajectories. Token quantities are millions (M). (a) Decisions and interactions CEO agentTrajectories Counted windowsModel repliesTool calls GPT-6 Astra472,37623,97396,763 GPT-5.6 Sol472,26527,133135,060 Qwen 3.8 Max473,31536,60794,260 Claude Sonnet 5472,24430,95988,768 DeepSeek V4 Flash473,27335,343132,440 DeepSeek V4 Pro473,09636,52797,639 Gemini 3.8 Flash473,11679,28586,070 Doubao Seed 1.8472,62655,92268,569 Rule CEO402,88000 No Action CEO857600 Total42425,767325,749799,569 (b) Tokens and interaction depth CEO agentInput (M)Output (M)Total (M) Model replies per trajectory GPT-6 Astra478.1630.74508.90510± 58 GPT-5.6 Sol989.2648.931,038.19577± 85 Qwen 3.8 Max1,533.71183.551,717.26779± 58 Claude Sonnet 51,080.2210.231,090.45659± 181 DeepSeek V4 Flash1,436.89214.381,651.27752± 53 DeepSeek V4 Pro796.8190.52887.32777± 180 Gemini 3.8 Flash565.1013.72578.821, 687± 130 Doubao Seed 1.8733.5939.59773.181, 190± 333 Rule CEO0000 No Action CEO0000 Total7,613.74631.658,245.39— A recorded model reply may contain several tool calls; neither unit is equivalent to a CEO decision window. Table 28 sums successfully sealed decision records with complete reported usage, excluding un- successful or separately archived attempts. These records contain 325,749 model replies, 799,569 tool calls, and 8.25 billion tokens. Input totals include cached tokens; output totals include reported reasoning tokens, without counting either component twice. Rule CEO and No Action CEO make no model calls. Per-trajectory and per-agent counts are provided in anc/data/workload. 46