Update README.md

README.md

@@ -955,18 +955,64 @@ configs:
     path: data/ZX_TAXES_TL/*.parquet
 ---
 
-![Argo-Bench City](assets/argo-bench-city.jpg)
 
-# Argo-Bench: evaluation warehouse
 
-This is the data warehouse for [Argo-Bench](https://argo-bench.com), including synthetic food-delivery company's
-operational and financial records for calendar 2024, in an Oracle E-Business
-Suite–style schema (GL, AR, AP, OM, XLA, HZ, ...) plus custom `XX_` operational tables.
+# Argo-Bench
 
-- **235 tables, 7,538,158,687 rows, 1,219 Parquet files, 71.3 GiB** (ZSTD).
-- The records are generated: customers, couriers, employees and transactions are synthetic.
-- The warehouse only: no answer keys, grader state or generator labels are included.
-- The questions, harness and grader: [github.com/TextQLLabs/Argo-Bench](https://github.com/TextQLLabs/Argo-Bench).
+A frontier benchmark for enterprise-scale data science. This repository holds its warehouse:
+one simulated year of a New York City food-delivery company, exported to an Oracle E-Business
+Suite 12.2 schema of 235 tables and 7.54 billion rows.
+
+**[Leaderboard](https://argo-bench.com)** · **[How it works](https://argo-bench.com/overview)** ·
+**[Paper](https://arxiv.org/abs/XXXX.XXXXX)** · **[Code](https://github.com/TextQLLabs/Argo-Bench)**
+
+## An enterprise ERP you can download
+
+The data that enterprise analytics runs on is rarely public. Large companies keep their orders,
+payouts, ledgers and customer records in ERP systems such as Oracle E-Business Suite and SAP,
+extended with custom tables of their own, and access to them is tightly restricted. Benchmarks
+therefore run on public datasets, which tend to be collections of loosely related tables rather
+than one system whose modules have to agree, and their answer keys are only as good as the
+people who wrote them. On real data, fraud that was never caught has no label, and a decision
+that was never taken has no outcome.
+
+Argo-Bench simulates the business instead. The warehouse can be released whole because no one
+in it is real, and every answer key comes from the simulator's own state rather than from an
+annotator.
+
+## What's in the warehouse
+
+- **A calibrated company.** 81 million orders from 17,902 real New York restaurants, delivered by
+  81,116 couriers over calendar 2024. The world is built from 34 public datasets and reports, and
+  calibrated to the delivery-app economics New York City publishes each quarter and to the
+  platforms' public filings. It absorbs a real shock: the city raised the minimum pay for
+  delivery workers to $19.56 an hour on April 1, 2024.
+- **One system.** 159 standard EBS tables across GL, AR, AP, OM, XLA,
+  HZ, OKC/OKS, ZX and more, where every standard table and column exists in Oracle's EBS 12.2
+  data dictionary, and 76 custom `XX_` tables for dispatch, courier pay, incentives, promotions,
+  sessions and support. An order resolves into dispatch decisions, courier pay, merchant payouts
+  and balanced general-ledger journals, and the modules reconcile. The schema was designed with
+  three ERP consultants.
+- **Fraud and abuse where public evidence says it happens.** Couriers who steal orders or spoof
+  their GPS, promotion-farming rings, account takeovers, shell storefronts and diverted payouts.
+  Each pattern is calibrated to how often honest customers share a device, an address or a card,
+  which is what makes them hard to tell apart.
+- **As-of views.** Eleven month schemas cut every table to what existed at the end of each month,
+  so a task can be posed as of any month-end and graded on the months after it.
+
+## Running the benchmark
+
+The 210 tasks cover trust and safety, marketplace operations, FP&A, accounting and growth. Agents
+query the warehouse, analyze it in a sandboxed Python environment, and file bans, forecasts,
+budgets, reported figures and dashboard data sources. The grader scores each filing against the
+simulation's hidden state. In the paper, the strongest of 12 models, Claude Opus 5.5, averages
+59.5 points and scores 95 or more on 34.8% of tasks.
+
+The tasks, the reference agent and its sandboxes are at
+[github.com/TextQLLabs/Argo-Bench](https://github.com/TextQLLabs/Argo-Bench), which also explains
+how to export a run for official scoring.
+
+Argo-Bench is built by [TextQL](https://textql.com).
 
 ## Layout