E-Commerce Bench: Evaluating LLM Agents on Long-Horizon Autonomous Business Operation

• E-Commerce Bench addresses gaps in prior benchmarks by combining real-world market data, deterministic counterpart negotiation, dynamic events, and concurrent multi-store operation over a year-long horizon.The benchmark is grounded in Taobao & Tmall data and supports up to four online stores across a simulated 365-day year.Source: §1 · p. 4
• The benchmark requires agents to research markets, negotiate inventory, formulate sales strategies, fulfill operations, and manage cash flow to maximize end-of-year total assets.Hidden costs and consumer demand force continuous search, negotiation, and market modeling through the benchmark’s e-commerce tools, memory, and context management.Source: §1 · p. 4
• GPT-5.6 Sol grows capital roughly 14-fold but ranks 16th of 18 on fraud avoidance, while Qwen3.8-Max-Preview leads open-weight models and 16 models show no clear price reduction across repeated purchases.Four of 18 models enter bankruptcy on some runs, and Claude Opus 4.7 leads bargaining and fraud avoidance despite mid-pack profit.Source: §1 · p. 4
• Its economy is reproducible: fixed customer purchases and returns and a deterministic negotiation kernel determine pricing, concessions, and accept/reject decisions, while an LLM only renders dialogue.The environment contains 6,886 products and 576 suppliers, with demand shifted by promotions, disasters, and supply-chain shocks; only renderer wording varies between runs.Source: §1 · pp. 4–5
• Evaluating 18 models across seven dimensions reveals capability profiles that total assets alone obscure, because the best earner is neither the best negotiator nor the most efficient operator.The dimensions include negotiation quality, fraud avoidance, cash flow and solvency, operational efficiency, operations execution, learning, and the primary score.Source: §1 · pp. 4–5

2.1 Episodic Long-Horizon Agentic Tasks

• Episodic benchmarks evaluate a terminal deliverable, missing dynamic dependency tracking, failure recovery, and history-based optimization across continuing sub-tasks.Examples span economically meaningful work, real-software interaction, and exploration-driven environments, but each run ends when its goal is met.Source: §2.1 · p. 6

2.2 Continuing Long-Horizon Agentic Tasks

• Continuing benchmarks accumulate rewards during indefinite operation, including portfolio trading, self-evolving agents, and business simulations such as Vending-Bench, YC-Bench, MerchantBench, and RetailBench.Existing business-operation benchmarks trade off negotiation, adversarial suppliers, real data, or reproducibility, leaving no other entry with a negotiating counterpart and reproducible supplier economics.Source: §2.2 · p. 6
• E-Commerce Bench combines all six target properties, using fixed demand and a deterministic negotiation kernel for pricing and accept/reject decisions while an LLM only renders dialogue.This design provides reproducible supplier economics while retaining negotiation within a data-grounded business-operation economy.Source: §2.2 · p. 6

3.1 Overview

• Agents receive ¥100,000 for a 365-day Chinese-marketplace operation spanning up to four stores, with performance scored on six capabilities including negotiation, solvency, execution, and long-horizon learning.The agent researches categories, bargains with suppliers, buys inventory, prices and ships products, handles returns, and withdraws settled revenue.Source: §3.1 · p. 7

3.2 Agent Loop Layer

• The agent loop limits each day to 600 simulated minutes, executes batched tools sequentially, advances through fixed daily settlement, and ends episodes after 4,000 turns or three tool-free turns.Most tools consume minutes, while advancing to the next day triggers costs, demand, cancellations, escrow settlement, deliveries, and reputation updates.Source: §3.2.1 · p. 8, Figure 4 · p. 8

3.2.2 Context Management

• Context eviction begins at 120,000 tokens and releases 60,000 by removing oldest tool-call groups, while a self-curated 20-entry memory store remains outside eviction.Across 90 episodes, 1,495 evictions occurred; transcript loss depended on tool traffic as well as turn count.Source: §3.2.2 · p. 8, Figure 5 · p. 9, §3.2.3 · p. 9

3.4 Sales and Economy Engine

• Sales are deterministic and depend on price, weekends, promotions, seasonality, calendar events, reputation, capacity, and inventory rather than sampled market demand.Capacity can dominate other factors: one accessory’s compounded expectation of 128.2 units was reduced to 9 sold, with the category constraint contributing a 7.5 factor.Source: §3.4 · p. 9, §3.4.3 · p. 10

3.4.2 Three-Account Deferred Settlement

• Cash follows deferred settlement: costs leave the bank immediately, sales pass through escrow and a platform wallet, and revenue becomes spendable only after nine days and explicit withdrawal.Ten consecutive days with a negative bank balance cause bankruptcy, even when sales are profitable but revenue remains unavailable.Source: §3.4.2 · pp. 9–10, Figure 6 · p. 10

3.5.2 Two-Layer Supplier Design

• The negotiation system separates deterministic economic decisions from LLM-rendered dialogue, supports repeated cross-supplier bargaining, and embeds fraud whose key signal is multi-round concession behavior rather than opening price.Among 576 suppliers, 152 are fraudulent; pre-deal scams inflate reservation prices toward 1.5× the honest floor, while post-deal scams underdeliver or provide defective stock.Source: §3.5.2 · p. 12, §3.5.3 · p. 13, §3.5.4 · p. 14

3.5.5 Negotiation Metrics

• Negotiation metrics measure per-reorder price discipline relative to supplier cost floors, but include only honest pairs, weight reorders equally, and exclude early-terminated episodes.The null treats paired cycles as equally difficult, while Figure 8 illustrates trajectories ranging from surrendered gains to progressively lower prices.Source: §3.5.5 · p. 14

3.6 Data Layer

• E-Commerce Bench’s data layer combines desensitized real-platform records with hand-synthesized quantities calibrated against platform logs, supplier price floors, and demand parameters.This construction supplies the benchmark’s product, supplier, and demand inputs while preserving a realistic e-commerce basis.Source: §3.6 · p. 14

3.6.1 Catalog and Suppliers

• The environment contains 6,886 products, 576 suppliers, 12 store types, 10 market events, and 8 promotions derived from desensitized platform data and hand-built synthesis.The catalog includes 424 honest and 152 fraudulent suppliers, with category-balanced fraud coverage and a fixed supplier list; fictional promotion names reduce recognition-based shortcuts.Source: §3.6.1 · p. 15

3.6.2 Systematic Information Asymmetry

• Agents observe catalogs and action outcomes but must infer hidden demand, cost-floor, elasticity, return-rate, and store-capacity parameters through costly exploration and operation.Market and supplier lookups replace quantities with labels or partial information; post-deal fraud is revealed only through arriving goods, while prior supplier prices and natural return rates provide limited learning signals.Source: §3.6.2 · p. 15

3.7 Evaluation and Metrics

• Final assets are the primary score, supplemented by six dimensions: bargaining, fraud avoidance, solvency, operational efficiency, operations execution, and learning.Table 2 ranks 18 models across five episodes using final assets and one metric for each capability; bankrupt episodes remain in pooled means, while operational records are not themselves scores.Source: §3.7 · pp. 15–16, Table 2 · p. 16

4.1 Experimental Setup

• The benchmark evaluated 18 models across five independent 365-day episodes each, using shared tools and deterministic economics, with bankrupt episodes included in mean asset scores.The asset multiplier is end-of-year total assets over the ¥100,000 opening stake; all 90 episodes ended before the 4,000-turn cap.Source: §4.1 · p. 16

4.2 Overall Performance

• A 1,264-fold asset gap separated GPT-5.6 Sol from Qwen3.5-Plus, with proprietary models occupying the top four places and Qwen3.8-Max-Preview fifth overall.All models started from the same ¥100,000 stake and deterministic simulated year; ten of 90 episodes ended insolvent.Source: §4.2 · p. 16

4.3.1 Negotiation Quality

• Negotiation scores ranged from 0.596 for Kimi K2.6 to 0.811 for Claude Opus 4.7, with rankings driven more by captured surplus than agreement frequency.Every model settled at least 96% of concluded honest sessions, while counterpart templates varied less than models overall.Source: §4.3.1 · p. 17

4.3.2 Fraud Avoidance

• Fraud losses varied by over 160-fold, and GPT-5.6 Sol ranked 16th of 18 on BadSpend% despite leading assets, because models differed mainly in which approaches became orders.Defective lots comprised 53.4% of money reaching fraudulent suppliers, while fraudulent suppliers received near-proportional outreach across models.Source: §4.2 · p. 16, §4.3.2 · p. 18

4.3.3 Cash Flow and Solvency

• Ten episodes went bankrupt, while the two biggest earners held the thinnest average cash buffers among solvent models, linking growth with greater solvency risk.No episode had wallet cash before day 15, 49 of 90 bottomed out in January, and all four models with bankruptcies surrendered more than half of peak assets.Source: §4.3.3 · p. 18

4.3.4 Operational Efficiency

• Fable5 generated ¥479 profit per tool call versus GPT-5.6 Sol’s ¥363 while making 59.9% fewer calls, and routine shipping and state-management actions consumed most call budgets.Shipping, advancing the day, withdrawing revenue, and publishing inventory accounted for 71.8% of calls, while pricing revisions represented 0.66%.Source: §4.3.4 · p. 18, Figure 10 · p. 19

4.3.5 Operations Execution

• Execution and learning exposed contrasting policies: GPT-5.6 Sol led assets despite a ¥1.62M refund bill, while Qwen3.8-Max-Preview alone improved surplus per day and beat shuffled prices significantly.Controllable return rates ranged from −0.05 for Qwen3.7-Max to 5.88 for Qwen3.5-Plus; Qwen3.8-Max-Preview scored 0.834 on AnchorRatio.Source: §4.3.5 · p. 19, §4.3.6 · pp. 19–20

4.4 Error Analysis

• Error analysis identified recurring failures across all models, including unsold inventory, fraudulent repeat orders, bargain acceptance, repeated research, and edge-of-year shipping failures.Every model triggered at least five detection rules, while fewer failures did not predict greater wealth and the cleanest profiles belonged to the best bargainers.Source: §4.4 · p. 20

5 Conclusion

• GPT-5.6 Sol turns ¥100,000 into roughly 14 times as much, but ranks 16th of 18 on fraud avoidance and earns less per tool call than Fable5.Claude Opus 4.7 bargains and avoids fraud best while finishing mid-pack financially, whereas Qwen3.8-Max-Preview leads open-weight models at 4.2 times the stake.Source: §5 · p. 21
• E-Commerce Bench simulates a 365-day online retail business using 18 tools, deterministic customer demand and supplier negotiation, and products, suppliers, and events from a real e-commerce platform.The benchmark evaluates 18 models across end-of-year total assets and six underlying capabilities, with each model running the year five times.Source: §5 · p. 21
• Four models go bankrupt in some runs, including two GPT-5.5 episodes that fail as early as day 17 before revenue settles.The benchmark’s deterministic kernel removes evaluation noise in complex multi-party interactions, while multidimensional diagnostics expose diverse failure modes.Source: §5 · p. 21

A Agent Harness and Interface

• The agent sees the environment only through tool schemas and replies, with year-long transcript retention managed inside a 128,000-token context window alongside persistent memory.The interface specification covers reply formatting, price-list presentation, and the eviction algorithm; prompts are reproduced separately in Appendix H.Source: A Agent Harness and Interface · p. 26

A.1 The Interface Contract

• Every tool call returns one string, while structured results remain parseable JSON and day-scoped notices attach to the day-advance result or the latest response.This routing prevents news from being attributed to an unrelated earlier call.Source: §A.1 · p. 26
• At most three of 17 morning-pass fields reach the agent, while 8,301 notices traversed the channel across 90 archived episodes, including 6,722 aggregated daily deliveries.The interface therefore delivers a compressed operational view rather than the full state computed by the environment.Source: §A.1 · p. 26
• Tool replies carry operational warnings such as shipping trade-offs and bankruptcy thresholds, and the context editor records eviction passes when episodes reach the context ceiling.Listing 1 shows a day-boundary notice and a solvency warning preceding a bankruptcy; Figure 12 summarizes eviction frequency over 90 evaluated episodes.Source: §A.1 · pp. 26–27, Figure 12 · p. 27

A.2 Loop Control and Context Editing

• The context loop counts uncleared message tokens, preserves the system and first user turn, protects two recent groups, and clears oldest complete tool-call groups until the release target is met.The threshold is 120,000 tokens with a 60,000-token release floor; plain-text assistant turns and idle warnings are never evicted, preserving valid provider protocols.Source: §A.2 · p. 27

B Data Layer

• The frozen data layer contains 6,886 SKUs, 60 category rows, 576 suppliers, 12 store types, 10 market events, and 8 promotions, identical across models and runs.The released tables expose quantities needed to verify reported numbers, while supplier behavior is specified separately in Appendix D.Source: B Data Layer · p. 28

B.1 Category Parameters

• Category parameters hide price ratios, elasticity, and return-band endpoints from agents, while SKU-level reference prices and natural return rates vary within category bands.Wholesale ratios equal the cost-floor ratio cj/vj across all 60 categories; return sources become readable only after stock moves.Source: §B.1 · p. 28

B.2 Store Types, Seasonality, and Physical Cost

• Agents choose one store type, whose difficulty, operating cost, category span, seasonality, and physical costs determine operations; nine of 12 types peak in November, while Food & Beverage peaks in February.Four types remain within ±0.2 of neutral seasonality, and parcel size scales freight and storage costs while shipping speed multiplies freight and return rate.Source: §B.2 · p. 28, Table 5 · p. 29, Figure 14 · p. 30

B.4 The Supplier List

• The supplier list is frozen and contains 424 honest and 152 fraudulent suppliers across 60 categories, typically seven or eight honest versus two or three fraudulent suppliers per category.Each supplier has an identity, honesty label, behavior template, possible scam type, served category, and bankruptcy threshold; names provide no signal, while email carries one.Source: §B.4 · p. 30

B.3 Event and Promotion Calendar

• The calendar contains eight opt-in promotions and ten unavoidable events: promotions provide advance notice, whereas events announce on their first day and can delay only new purchase orders.The Guangdong factory fire overlaps the Spring Blossom Sale, creating a supply shock without contaminating measured discount responses.Source: §B.3 · p. 29, Figure 15 · p. 30, Table 6 · p. 31

B.5 Properties a Reader Should Check

• Released files understate return rates, leak supplier fraud labels, misstate easy-tier rent, and ignore the Raw Material Shortage cost increase.Twenty categories show “low” or “very low” wording despite numeric bands reaching 20%–35%; supplier email suffixes separate honest from fraudulent suppliers, while the rent column shows ¥45 instead of ¥60 and shortage costs are unread.Source: §B.5 · pp. 31–32

C Economy Engine

• The economy engine hides demand and cost-floor parameters while exposing catalog facts and realized outcomes, making sourcing and pricing depend on inference from consequences.Table 7 characterizes each hidden quantity through simulated minutes, ¥, or irreversible loss; tool-call costs are measured against a 600-minute working day.Source: §B.5 · p. 32, Table 7 · p. 32

C.1 The Daily Settlement Order

• Daily settlement executes thirteen fixed-order steps at 08:00, charges costs before revenue, applies one-day-lagged sales, and finalizes unpaid or unshipped orders deterministically.Solvency is tested at the morning low-water mark; year-end finalization discards unshipped orders, processes seeded returns, and drains escrow to the platform wallet.Source: §C.1 · p. 33, Figure 16 · p. 33

C.2 Demand Model

• Demand combines category volume, store type, price, promotions, seasonality, events, reputation, and capacity saturation, with deep discounts sometimes reducing demand.The quadratic price branch is non-monotone, while seeded rounding and on-shelf inventory make outcomes deterministic and stockouts truncate demand without signaling the agent.Source: §C.2 · p. 34

C.3 Reputation Law

• Reputation is recomputed each settlement from shipped, returned, cancelled, and sold units, with cancellations penalized more heavily than returns and counters decaying by 0.85.Goodwill rewards shipping rather than selling; a newly opened store’s reputation falls from 0.5 to 0.353 at its first settlement.Source: §C.3 · p. 34

C.4 Returns

• Returns combine natural, supplier-quality, price, and shipping factors, are decided at shipment, arrive after seeded delays, refund retail price, and re-enter inventory as fresh stock.Each shipment draws returns at shipment time, assigns a uniform arrival day from 3–7 days, and refunds from the originating escrow batch while freight and commission remain unrecovered.Source: §C.4 · pp. 34–35

C.5 Cost Model

• Costs debit cash through setup, operations, occupancy, storage, freight, procurement, and commission, while liquidation returns only 10% of quantity-weighted purchase cost.Storage rises with product size and age, commissions are 2% of gross revenue and persist on refunds, and closing without liquidation leaves storage accruing.Source: §C.5 · p. 35

C.6 Unit Profit and Break-Even

• A returned unit is resold as stock, so each retained unit funds 1/(1 −θj) shipping attempts and corresponding storage cycles over its Th-day warehouse wait.Break-even volume n⋆ counts the daily retained units required to cover the store tier’s operating cost.Source: §C.6 · p. 35
• Procurement is the largest unit-cost component, and the negotiated closing price affects unit profit more than pricing choices.Figure 18 decomposes retail price into five shares and relates unit profit and daily break-even volume to the price ratio.Source: §C.6 · p. 35, Figure 18 · p. 36
• At the floor price, the Women’s Fashion SKU earns ¥87.29 per retained unit and needs 2 daily retained units, while Snacks & Nuts needs 13.The fashion SKU has a 49.2% natural return rate, uses standard shipping, and waits Th = 14 days; closing at the floor multiplies unit profit by 2.4 and 2.8 for the two comparisons.Source: §C.6 · p. 36

C.7 Constant Reference

• The environment uses a ¥100,000 opening stake, permits at most 4 stores, runs 365 days from 2026-01-01, and declares bankruptcy after 10 consecutive days of negative balance.Orders have a two-day shipping deadline and nine-day escrow window; every model run shares seed 20260122.Source: §C.7 · p. 36

D Deterministic Negotiation Kernel

• The deterministic supplier policy is implemented through four decision functions and five scam scripts, with metrics defined in §E.1 and prompts in §H.It is described as an independent implementation of the counterpart design of Zhang et al. (2026a).Source: D Deterministic Negotiation Kernel · p. 36

D.1 Kernel Decision Functions

• The kernel represents each supplier–SKU bargaining state with a latent type, opening harshness, deadline scale K = 10, price frame, seeded streams, and offer histories.Reservation price rB comes from the data layer, while urgency κB ∈[0, 1] and stance ηB come from supplier templates; offers are normalized by the frame width.Source: §D.1 · pp. 36–37
• Acceptance is evaluated first, walk-away second, and counter-offers third; acceptance uses logistic reciprocity, while counter-offers move monotonically toward the floor without crossing it.Since ρf ≤0 in every template, conceding faster never improves the agent’s acceptance chance; walk-away requires a below-floor offer from midpoint onward, with kwalk = 5.Source: §D.1 · p. 37
• 66.0% of concluded negotiation sessions close in 3 message turns, and no kernel session passes round 7.The deadline scale K shapes negotiation terms but does not truncate sessions; bargaining continues while both sides answer.Source: §D.1 · p. 37

D.2 Data Grounding and Calibration

• The kernel derives SKU-specific cost floors, wholesale quotes, and scam caps from category ratios, with pre-deal scams raising the reservation floor by θ_s = 1.5.The scam floor is re-evaluated before charging, preventing agents from bypassing it by skipping bargaining.Source: §D.2 · p. 38
• The calibration ties wholesale and cost ratios across all 60 categories, making acceptance of the opening quote score exactly 0.5 before rounding against the CSE+ baseline.The scam cap binds most often, producing an average overpayment multiple of 1.218; harshness is solved so mean first quotes equal nominal wholesale prices while stance and urgency affect concessions and acceptance.Source: §D.2 · p. 38

D.3 Behavior Templates

• The five scam templates use the hardest honest template at lower κ_B, while cue policies range from state-dependent signals to constant holds or fixed adversarial pressure.Template assignment is fixed by the released supplier list, and suppliers sharing a template bargain identically because per-supplier urgency does not enter the kernel.Source: §D.3 · p. 39
• Mean negotiation turns barely vary across honest templates, indicating that agents do not spend extra rounds on harder behaviors.The reported pooled standard errors follow the per-session definition and are additionally split by model.Source: §D.3 · p. 39

D.4 Session Protocol and Commit Rules

• Economic actions are parsed from fenced negotiate blocks into SKU-specific offers, acceptances, or rejections, with prose removed before rendering.A purchase enters the metrics ledger only after validation, quote matching, and order settlement; failures return the pair to bargaining and prevent phantom deals.Source: §D.4 · pp. 39–40
• Only suppliers retain complete deal histories across cycles, whereas agents retain prior information only through surviving context or persistent memory.Each new offer on a pair creates a fresh kernel while preserving the supplier’s persistent type and reservation price.Source: §D.4 · p. 40
• 192 supplier closures across 90 archived episodes show that bargaining is constrained by fixed message time and attrition after 10–20 successful orders.Concentrating sourcing on favorable counterparts can therefore cause those suppliers to disappear before the year ends.Source: §D.4 · p. 40

Bargaining behavior

• 0.79%–1.66% of reference price per counter-offer was conceded by pre-deal kernels, versus 3.94%–9.93% for honest templates.These results cover 3,978 archived sessions with non-zero recorded concession rates and support the population claim for pre-deal scams only.Source: §D.5 · pp. 40–41
• 4.37% and 4.44% concessions from post-deal scams fall inside the honest band, so price-based inspection misses their delivery defection.An agent that reads only prices sees the overpayment associated with pre-deal scams, not the post-deal short-shipping or defective stock.Source: §D.5 · pp. 40–41

Fraud timing and pricing

• Five scams affect 152 of 576 suppliers and split into pre-deal price inflation versus post-deal fulfillment defects, with pricing unable to distinguish scam siblings.Every fraudulent supplier uses the Adversarial kernel and exactly one scam; pre-deal reservation prices remain below honest wholesale quotes, while post-deal scams retain the honest floor and defect at delivery.Source: §D.5 · p. 40, Table 9 · p. 41

Rare scam limitation

• The vip_fee pay-first scam is too rare to measure reliably because failed purchase orders leave sessions open after rollback.Table 8 reports single draws; membership fees were paid only three times, totaling ¥3,000, by two models across 90 episodes.Source: §D.5 · p. 41

Primary score

• ¥1? No. End-of-year assets sum bank balance, platform wallet, and escrow after deferred orders and returns settle, with the asset multiplier normalized by the ¥100,000 opening stake.A = b + w + e and M = A/A0, where A0 = ¥100,000; figures average assets over five episodes, including bankrupt years.Source: E Metric Definitions · p. 42

Session eligibility

• Only bargaining sessions ending in agreement or disagreement qualify; open records, failed-order agreements, and fraudulent suppliers are excluded from value-extraction metrics.Honest-supplier sessions use 424 names from the 576-name supplier list, while fraudulent sessions form the remainder.Source: E Metric Definitions · p. 42

Notation and aggregation

• Each session closes at agreed price p_z or f_z = ⊥, while negotiation width is Δ_z = v_j − c^s_j relative to the SKU reference price and supplier floor.Statistics are computed per episode, then averaged unweighted across five episodes.Source: E Metric Definitions · p. 42

E.1 Negotiation Quality

• The negotiation-quality records retain 12,060 honest sessions and exclude 1,163 fraudulent sessions after admission rules, with AGR+ denominated by honest concluded sessions.Deals closed totals 13,128 agreements over 90 episodes, while rounds to deal counts exchange turns and includes fraudulent counterparts.Source: §E.1 · pp. 42–43
• SE+ factorizes into AGR+ · CSE+, but CSE+ drives rankings because every model settles at least 96% of concluded honest sessions.The agreement factor remains useful for detecting policies that achieve excellent deals by abandoning sessions; surplus can be negative when agreements exceed buyer value.Source: §E.1 · p. 42

E.2 Fraud Avoidance

• BadSpend% ranks fraud avoidance using cleared purchase-order cash, including membership fees, across five scam types; lower values are better.Diagnostic measures include scam-specific spending, membership-fee payments, and distinct fraudulent suppliers contacted or ordered from, while refusal behavior is unobservable.Source: §E.2 · pp. 43–44

E.3 Cash Flow and Solvency

• Cash-flow statistics use daily bank, wallet, warehouse, escrow, and total-asset records, but terminal-row handling, unequal episode lengths, and pooled bankrupt runs constrain comparisons.The terminal row would spuriously raise pooled overdrawn mornings from 177 to 187, while idle-wallet values represent money left at 08:00 before withdrawals.Source: §E.3 · pp. 44–45
• Drawdown ranks the unweighted mean peak-to-trough total-asset ratio across five episodes, with ratios above 1 indicating assets fell beyond the entire peak into debt.Reconstructing the numerator from recorded rows leaves the 18-model ordering almost unchanged but moves Gemini 3.5 Flash from 0.450 to 0.502.Source: §E.3 · p. 44

E.4 Operational Efficiency

• Operational efficiency ranks profit per tool call using ratio-of-sums across five episodes, avoiding domination by episodes that end near zero after few calls.For GPT-5.5, two day-17 bankruptcies produced −552 and −491 per call; averaging per-episode ratios would move it from fourth to seventeenth.Source: §E.4 · p. 45
• Efficiency diagnostics partition calls into eight activity bands and measure turns, context evictions, and memory traffic, but calls capture neither monetary cost nor simulated working time.Waiting costs a call without simulated minutes, empty turns lower calls per turn, and truncated episodes reduce every operational-efficiency mean.Source: §E.4 · pp. 45–46

Measurement scope

• The dimension ranks controllable return rate, where lower is better, while also reporting realized returns, return channels, refund costs, on-time share, cancellations, shipping tiers, and store counts.Controllable return rate captures the pricing contribution to expected returns; the other diagnostics provide operational context.Source: §E.5 · p. 46

Return accounting

• The return model separates natural, defective, pricing, and shipping-tier channels, accrues them when parcels dispatch, and averages normalized rates over five episodes.Year-end totals divide by max(1, shipped units), while realized returns are counted when parcels arrive and delayed returns are finalized.Source: §E.5 · pp. 46–47

Shipping-tier effects

• 89 of 90 episodes record a zero shipping-tier channel, leaving controllable return rate equal to the pricing channel, while freight speed still affects underlying return rates and is reported as fast-shipment share.Fast, standard, and slow freight use factors of 0.75, 1.00, and 1.30, respectively.Source: §E.5 · p. 47

Fulfillment and store churn

• Orders receive two days for dispatch; missed deadlines cancel shipments and revenue, while fulfillment reports on-time share and cancellation counts, and store activity tracks openings and reopens across four concurrent slots.Each opening costs ¥500, and reopened store types restart at the reputation floor.Source: §E.5 · p. 47

Limitations

• The metrics are bounded by expectation-based returns, zero-valued no-shipment episodes, sales-dependent on-time means, scale-dependent gross costs, and supplier-pooled defective rates that measure fraud avoidance rather than execution.The defective channel aggregates cumulative SKU deliveries, while refund loss and return freight are gross ¥ totals tied to sales volume.Source: §E.5 · p. 47

E.6 Learning over the Horizon

• AnchorRatio evaluates whether later repeat purchases improve over earlier agreements by comparing each re-order with the agent’s previously won price.The analysis also reports normalized overpayment, new-low frequency, permutation z-scores, and half-year comparisons ordered by session conclusion time.Source: §E.6 · p. 47
• The anchoring analysis retains honest supplier–SKU pairs transacted at least twice, pooling repeat deals within episodes and weighting each re-order equally per unit.The filters exclude fraudulent pairs, missing conclusion dates, nonpositive price ranges, and one-time pairs; episodes without repeat pairs drop from model means.Source: §E.6 · p. 48
• The null model independently reshuffles each repeat pair’s positions, and per-episode statistics combine with Stouffer’s rule so positive new-low values indicate improvement over shuffled ordering.Regret reverses the subtraction because lower regret is favorable, while the model-level AnchorRatio averages episode ratios rather than pooled means.Source: §E.6 · p. 48
• The repeat-purchase filters leave 2,230 supplier–SKU pairs across 90 episodes, with model-specific re-order counts reported in Table 16.Half-year comparisons require at least four honest sessions, split sessions by conclusion day, and score disagreements as zero surplus.Source: §E.6 · p. 48
• AnchorRatio ranks models because it separates 53 of 153 model pairs by more than two standard errors, versus 15 for surplus half-lift.The measures have important scope limits: anchoring uses population spread, other archive spreads use sample spread, and half-lifts are unscaled differences of within-episode means.Source: §E.6 · p. 49

F Per-Dimension Results

• Per-dimension tables cover all 18 models, retain consistent row ordering by mean end-of-year total assets, and archive formulas and field names in Appendix E.Section 4 reports only selected headlines, whereas this section collects the underlying measurements for every dimension.Source: F Per-Dimension Results · p. 49

F.1 Per-Episode Asset Trajectories

• Episodes finish above and below the ¥100,000 opening stake in 8 of 18 panels, while surviving models’ best-to-worst episode ratio has median 2.76.Among 14 models that never went bankrupt, the median ratio ranges from 57.0 for Qwen3.7-Max to 1.55 for Kimi K2.6; once episodes exit, GPT-5.5’s mean curve becomes survival-conditioned.Source: §F.1 · p. 49

F.2 Negotiation Quality

• Within models, surplus varies by a median 0.093 across counterpart templates, exceeding the 0.051 difficulty gap between the easiest and hardest presets.The within-model span remains below the 0.215 best-to-worst per-model CSE+ gap, and fraudulent counterparts produce lower measured surplus for 16 of 18 models.Source: §F.2 · pp. 50–51

F.3 Fraud Avoidance

• The widest honest supplier search reaches only 14.3% of 424 names, while the narrowest reaches 2.4%, leaving screening decisions dependent on each model’s selected supplier slice.Fraudulent suppliers are analyzed separately from honest counterpart templates because they share the Adversarial label but use different concession parameters.Source: §F.2 · p. 50, §F.3 · p. 51

F.4 Cash Flow and Solvency; F.5 Operational Efficiency; F.6 Operations Execution

• The appendix evaluates solvency through normalized and raw drawdowns, troughs, overdrawn mornings, idle cash, warehouse size, and bankrupt runs, alongside operational-call efficiency and return execution.Efficiency records turns, profit per call, eviction passes, memory traffic, and activity-band allocation; operations records refunds, shipping tiers, store counts, and rates over shipped units.Source: §F.4 · p. 51, §F.5 · p. 52, §F.6 · p. 52

F.6 Operations Execution

• Operations execution rarely fills the four-store cap: only GPT-5.6 Sol, GLM 5.2 (high), and Qwen3.8-Max-Preview average more than 3.5 stores trading daily.Most models still open stores beyond the cap, replacing existing stores rather than expanding capacity.Source: §F.6 · p. 52

F.7 Learning over the Horizon

• Learning analysis compares supplier-SKU anchoring and first-versus-second-half behavior, while procurement templates show the hardest counterpart receives 4.8% of spend against 11.3% of names.The easiest template receives 18.5% of spend against 24.1% of names, with fraudulent spending excluded from both panels.Source: §F.7 · p. 52, Figure 24 · p. 53

F.8 Profiles and Rank Crossings Across Dimensions

• No model ranks in the top five across all seven dimensions; Qwen3.8-Max-Preview comes closest, while operations execution shifts asset rank by 4.2 places on average.Across 153 model pairs, 311 rank orderings reverse; operational efficiency shifts rank by only 0.7 places.Source: §F.8 · p. 53

G Failure Case Studies and Detection Rules

• Failure case studies trace membership-fee, short-delivery, and defective-lot scams, including annotated transcripts, a bankruptcy timeline, and rules underlying the incidence matrix.The section preserves prices, dates, wording, and episode records while explaining manufactured urgency and promised future discounts.Source: G Failure Case Studies and Detection… · p. 53

G.1 Membership Fee Upsell; G.2 Short Delivery

• GPT-5.5 declined every membership-fee pitch, while a short-delivery supplier delivered 103 of 150 confirmed ibuprofen units, leaving a 47-unit shortfall the agent never checked against inventory.The supplier demanded a one-time premium-program charge and refused product orders until paid; the pitch arrived twice, nine months apart. GPT-5.5 sent 17 orders to short-delivery suppliers over the year.Source: §G.1 · p. 54, §G.2 · p. 55

• Defective suppliers reveal nothing through conversation or quotes, while every delivered unit raises the SKU’s return rate to the defective blend.Returns arrive days after the causative shipment, so early sales can appear clean even when inventory is defective.Source: §G.3 · p. 55
• Figure 26 counts fraudulent suppliers contacted and ordered from across 18 models and five episodes, including conversion shares and fraudulent spend.The figure’s panels compare contacts with orders, order conversion shares with fraudulent spend, and contacts with fraudulent spend.Source: Figure 26 · p. 56
• On the morning of the re-order, the SKU showed zero returns across twelve delivered units, while four sold units remained within the three-to-seven-day lag; ten days later, half the lot’s sold units were returning.Source: §G.3 · p. 56
• 35 orders over the year made defective stock the episode’s costliest scam.In the cited episode, delayed returns caused half the units sold from the reordered lot to come back ten days later.Source: §G.3 · p. 56

Collapse trajectory

• 586 orders cleared 1,428 units in 130 days, failing to catch up with inventory bought for a full year.Four stores opened on day one, and the subsequent buying converted three fifths of the stake into stock.Source: §G.4 · p. 56
• 12 of 24 procurement orders went to fraudulent suppliers at the raised pre-deal floor, deepening the cash-flow hole.Source: §G.4 · p. 56

Solvency failure

• The episode became unrecoverable after the first negative close on 2026-05-02, because the platform wallet held at most ¥1,597.98 and liquidation recovered too little.No withdrawal could have saved the episode, distinguishing this failure from wallet-coverable closes in §4.4.Source: §G.4 · pp. 56–57

G.5 Detection Rules Behind the Incidence Matrix

• Opening-quote acceptance was flagged for 636 of 11,987 honest agreements, although suppliers near their own floor leave little room for improvement.The rule therefore makes part of the measured negotiation shortfall definitional rather than necessarily evidence of poor bargaining.Source: §G.5 · p. 57

H Prompts

• The prompt configuration is fixed across episodes, with the system message preserving turn-loop rules, eviction thresholds, and the task brief as guaranteed context.The first user turn is a one-sentence business instruction, and the assembled system message contains 7,371 characters across the archived episodes.Source: §H.1 · pp. 59, 62, 63

H.1 Agent Task Brief

• The shared task brief makes the agent maximize combined bank and wallet balance over {max_days} days while researching, sourcing, negotiating, selling, fulfilling, handling returns, and growing stores.The brief permits up to four store types and describes delayed settlement, upfront costs, working-capital management, and bankruptcy after ten consecutive negative-bank days.Source: H Prompts · p. 58, §H.1 · p. 59

G.5 Detection Rules Behind the Incidence Matrix

• Table 18 defines ten recurring failure modes using one mechanical test per episode, while incidence counts models and episodes without ranking severity by frequency.The matrix includes 18 models and 90 episodes; truncated bankrupt episodes enter unadjusted, and upward drift uses 88 episodes with a per-episode permutation test.Source: §G.5 · p. 57, Table 18 · p. 61

H.2 Supplier Renderer Contract

• The supplier renderer receives recent supplier history and served catalog prices, but a kernel decision determines the quoted prices and decisions while the renderer verbalizes them.The renderer is explicitly told whether the supplier is honest or fraudulent; fraudulent suppliers use one of five scam scripts, while purchases still execute at the kernel-agreed price even if the rendered message differs.Source: §H.2 · p. 64

H.3 The Fenced negotiate Grammar

• The agent must express economic negotiation actions in fenced negotiate JSON blocks supporting offer, accept, and reject operations with SKU, price, and quantity fields as required.The grammar appears both in the task brief and schema, and validation is specified in §D.4.Source: §H.3 · p. 64
• Text outside negotiate fences is conversational only, so a message arguing for a price without a structured block does not move money.This separates natural-language negotiation from executable economic actions.Source: §H.3 · p. 64