100 executable manufacturing ERP workflow tasks with deterministic FactoryScore grading

100 deterministic long-horizon DevOps/SRE agent tasks over an executable NovaCart world with 32-67 call reference trajectories and state-diff vcode verifiers.

100 high-level sales operations requests with evidence-determined multi-system trajectories.

100 high-level legal-agent tasks with 97-asset multi-provider evidence rooms, distinct deep trajectories, and native causal state verification.

100 causal corporate-finance workflows over a closed ERP sandbox: provider-shaped MCP tools, distributed evidence, exact state transitions, and no LLM judge.

Six independently authored synthetic CRM workflows on one mocked CLI/web/REST/MCP world per episode. Partial Arc-inspired coverage, not source-row reproduction.

24 deeply realistic stateful MCP tasks across six enterprise domains

WebBench: browser-only correlated SaaS work over deterministic populated Blobfish worlds — 159 tasks graded executably against the applications backends (WebScore, no LLM judge)

HubBench: one oracle-proven, deterministically graded Blobfish benchmark family per Harbor Hub professional-domain cluster — 104 stateful multi-system employee-decision tasks over isolated SQLite worlds (HubScore, no LLM judge)

100 executable investment-banking workflows with deterministic DealScore grading

100 executable Oracle-Fusion-shaped ERP workflows with deterministic ERPScore grading

Executable litigation environments for agent training and evaluation. 100 long-horizon tasks, a median of 108 verified tool-calling steps each, worked across eight systems of record (Clio Manage, iManage Work, CM/ECF, Google Calendar, Docusign, Relativity, CourtListener and nine litigation sub-systems). Graded by deterministic state-diff and trace assertions — no LLM judge on the reward path.

100 executable semiconductor exception workflows with deterministic SemiOpsScore grading

100 executable manufacturing ERP workflow tasks with deterministic FactoryScore grading

100 deterministic long-horizon DevOps/SRE agent tasks over an executable NovaCart world with 32-67 call reference trajectories and state-diff vcode verifiers.

100 high-level sales operations requests with evidence-determined multi-system trajectories.

100 high-level legal-agent tasks with 97-asset multi-provider evidence rooms, distinct deep trajectories, and native causal state verification.

100 causal corporate-finance workflows over a closed ERP sandbox: provider-shaped MCP tools, distributed evidence, exact state transitions, and no LLM judge.

Six independently authored synthetic CRM workflows on one mocked CLI/web/REST/MCP world per episode. Partial Arc-inspired coverage, not source-row reproduction.

24 deeply realistic stateful MCP tasks across six enterprise domains

WebBench: browser-only correlated SaaS work over deterministic populated Blobfish worlds — 159 tasks graded executably against the applications backends (WebScore, no LLM judge)

HubBench: one oracle-proven, deterministically graded Blobfish benchmark family per Harbor Hub professional-domain cluster — 104 stateful multi-system employee-decision tasks over isolated SQLite worlds (HubScore, no LLM judge)

100 executable investment-banking workflows with deterministic DealScore grading

100 executable Oracle-Fusion-shaped ERP workflows with deterministic ERPScore grading

Executable litigation environments for agent training and evaluation. 100 long-horizon tasks, a median of 108 verified tool-calling steps each, worked across eight systems of record (Clio Manage, iManage Work, CM/ECF, Google Calendar, Docusign, Relativity, CourtListener and nine litigation sub-systems). Graded by deterministic state-diff and trace assertions — no LLM judge on the reward path.

100 executable semiconductor exception workflows with deterministic SemiOpsScore grading