ZORO

RETAIL MANAGEMENT · RL EVALUATION

Can an agent manage a supermarket well enough to survive and profit over 180 days?

Zoro is an Ethara.AI initiative that builds a structured evaluation benchmark on top of RetailBench, a simulation environment where an AI agent manages a single-store supermarket over 180 simulated days. The benchmark turns the open-ended RetailBench simulator into a rigorous, reproducible task suite that measures how well frontier language models handle complex, multi-step business decisions over time. Running a store is not a single decision — it requires coordinating inventory, pricing, supplier selection, customer reviews, news events, and cash position simultaneously, where a bad stocking decision in week one can cascade into bankruptcy by week eight. The oracle policy — with privileged access to supplier quality scores — achieves $131,510 net worth and 267,998 units sold. GPT-5.5, the strongest model tested, reaches just 18.6% of that ceiling.

  • Long-horizon planning: decisions made on day 5 directly affect outcomes on day 80
  • Evidence-grounded reasoning: product quality data must be gathered before acting, not assumed
  • Consistent strategy execution: the same decision logic applied repeatedly across 180 cycles
  • Temporal follow-up: orders placed must be tracked; quality issues discovered must trigger corrective action
  • Oracle-verified difficulty: only achievable tasks enter the benchmark — trivial and broken configs filtered out

Three numbers that define the scope of the Zoro evaluation harness, grounded in the RetailBench simulator.

Target Evaluation Tasks

10,000

2k hard · 6k middle · 2k easy

Simulation Horizon

180

days per evaluation task

Oracle Net Worth

$131,510

reference ceiling (not a target)

Four components define Zoro's evaluation framework, from task construction to failure attribution.

Component 01

Task Suite Design

  • ~10,000 versioned tasks, each pinning all 27 RetailBench environment parameters
  • Four presets × five demand elasticity profiles × five economic tiers × three news intensity levels
  • Tasks stored in a stable dataset.json schema with preset, scenario, seed, and oracle verdict

Component 02

Oracle Verification

  • Oracle policy runs on every candidate task with privileged access to supplier quality scores
  • Tasks where oracle pass_threshold_ratio is in [0.2, 0.8] are marked achievable and enter the benchmark
  • Trivial (> 0.8) and impossible (< 0.2) configurations are excluded automatically

Component 03

Evaluation Harness

  • Five agent scaffolds: ReAct, Plan-and-Act, Reflection, Step-Reflection, Exec-Strategy
  • Tool calls communicated via XML parsing — works identically across any model regardless of native function-calling support
  • Every tool call logged to tool_calls.jsonl across all 180 days

Component 04

Diagnostic Metrics

  • Eight stage-wise metrics: QDepth, Completeness, PriceCloser, SupplierQual, QualityFirst, PriceFirst, FollowUp, Resolved
  • Each metric maps to one of three identified failure modes: incomplete evidence, surface-level decisions, inconsistent policy
  • Computed post-hoc from tool_calls.jsonl — not from simulation state

The method

Three phases turn an environment configuration into a scored, reproducible reinforcement-learning signal.

Phase 01

Instance Preparation

  • Assign a preset, scenario overlay, and random seed to pin all 27 environment parameters
  • Run the oracle policy to verify the task is achievable — not trivially easy, not structurally broken
  • Record the oracle net worth as the reference ceiling for that task

Phase 02

Agent Evaluation

  • Run the agent under one of five scaffolds across all 180 simulated days
  • Agent reads state via read-only tools, then issues action tools to order, price, and manage listings
  • Every tool call and response is logged to tool_calls.jsonl

Phase 03

Scoring & Diagnostics

  • Primary metric: net worth at day 180 expressed as % of oracle ceiling
  • Secondary: eight diagnostic metrics computed from tool_calls.jsonl to attribute failures
  • Multiple seeds run per task to estimate variance from stochastic demand

Seven frontier models were benchmarked on the dynamic_hard preset under the ReAct scaffold across 180 simulated days (results from the RetailBench paper; arXiv:2603.16453).

GPT-5.5 achieved 18.6% of the oracle's $131,510 net worth — the strongest result of seven models tested. DeepSeek-V4-Pro reached 7.7%. Five of seven models bankrupted before day 180, ending with negative net worth.

Chart 1 placeholder
Fig. 1: Estimated mean GPT-5.5 score (% of oracle) by demand elasticity profile. Inelastic demand tasks yield higher scores; mixed-demand tasks present the greatest challenge.

Score by environment preset, and planned task distribution across presets.

Chart 2 placeholder
Fig. 2: Estimated GPT-5.5 mean score by environment preset. dynamic_hard is the most challenging; still_middle removes news and customer reviews for a cleaner inventory signal.
Chart 3 placeholder
Fig. 3: Planned task distribution by environment preset, stacked by demand elasticity type. 10,000 total tasks across four presets.
1 ABOVE 50% ORACLE 6.3%
12 PARTIAL 1–49% 75.0%
3 BANKRUPT / UNGRADED 18.8%
18.6% GPT-5.5 Best Score (% of Oracle, dynamic_hard)

Dataset Scenarios

Presets

SKUs (hard preset)

GPT-5.5 Mean Score

Loading...
Scenario Preset Difficulty GPT-5.5 SKUs

Four presets define Zoro's difficulty landscape. Dynamic presets activate news events and customer reviews, introducing information asymmetry. Still presets use static pricing for a cleaner inventory signal.

No preset data available yet.

Need a long-horizon agent evaluation harness for your domain?

Ethara builds held-out, reproducible agent benchmarks grounded in real domain data for frontier research teams.

Get in touch