Real websites become resettable agent training grounds: 244 sites, 12,984 tasks, and a 12-2-0 blind-audit win

A fully automated pipeline now turns real websites into executable, resettable local environments where visual web agents can explore, change state and recover from errors. Earlier controllable environments relied on human-assisted per-site mirroring, on generation from specifications that can diverge from a specified site, or on capture-and-replay whose interaction coverage is bounded by what was captured. The resource pool holds 244 website artifacts across 14 application domains and 53 categories and supports 12,984 executed and accepted tasks; in the 85-site historical construction records, no per-site manual implementation, debugging or repair is recorded after initial setup and authorization.

Construction and acceptance are split: an Explorer–Builder reconstructs pages from browser-observable structures, resources and before/after interaction states, and an independent verifier re-executes checks on the frozen candidate. Builder self-checks can overlook source-inconsistent defects, and a discrepancy may reflect either faulty implementation or insufficient observation, which call for different actions. The verifier cannot edit candidate code and routes discrepancies to local repair or targeted recapture depending on whether evidence establishes the expected outcome; of 229 historical Builder–Verifier report pairs, 219 pass Builder self-checks and independent verification finds evidence-supported discrepancies in 39 pairs across 22 sites.

The reconstructed environments supply training data that transfers to external web-agent benchmarks. What was missing was a supply of real-website environments that keep source correspondence while allowing repeatable execution. Fine-tuning Qwen3.5-9B on 5,856 interaction trajectories raises REAL success from 22/91 to 31/91, or 9.9 percentage points, on 91 jointly valid tasks; Odysseys-Easy Rubric Avg and Task Avg rise by 6.25 and 6.32 percentage points, and WebArena-Verified Overall rises 13.33 percentage points under the fixed harness.

The demonstrated scope is 244 website artifacts across 14 domains and 53 categories supporting 12,984 executed and accepted tasks; quality comparisons cover five WebHarbor sites with 14 historical states and five VeriEnv artifacts, and training validation covers Qwen3.5-9B before and after fine-tuning on REAL, Odysseys-Easy and WebArena-Verified.

Training experiments mainly concern the 9B parameter scale and do not systematically evaluate larger foundation models, other post-training methods or larger training budgets; the system reconstructs and verifies source states at particular times without systematically evaluating mirror timeliness or long-term maintenance under continuing source change; high-fidelity reconstruction also involves content authorization and compliance constraints.