Four-Factor Integrity Breakdown & Primary-Source Evidence
Access Model: Strictly Private
"To maintain evaluation integrity, all 91 tasks across the four AA-Briefcase project scenarios are private, including task instructions, project input files and grading rubrics."
Originality
Weight 25%
10/10
Reuse Profile: All Original (0% reuse)
"AA-Briefcase is a new benchmark for testing models on realistic knowledge work tasks in complex projects built by industry experts. Models are evaluated on multi-week knowledge work projects, each with many linked tasks and thousands of input source files."
Freshness
Weight 20%
10/10
Initial Release: 2026-06Latest Refresh: 2026-09Rolling Updates: No
"Version 4.3.2 September 2026 to October 2026 GDPval-AA v2.1: the Elo scale is now anchored to DeepSeek V4.1 Flash (max) at 1600, and ratings are fitted with a Crowd-BT model. AA-Briefcase v1.1: the anchor is unchanged at GPT-5.5 (medium) at 1000, but each pairwise scope is now fitted with a Crowd-BT model."
Integrity: Independent Offline (Leak-Verified)Sandbox: offline isolated
"Model submissions are run with Stirrup in a week-scoped E2B sandbox. Turns: Agents run for up to 500 turns per task. Tools: The agent is given a single code-execution tool that runs shell commands and code inside the sandbox, plus the finish tools below (and a view-image tool when the model supports vision). The sandbox has no internet access, so the agent can only use the provided source files."
Frontier Model Standings on AA-Briefcase v1.1
Top 15 active models tracked
Benchmark maintainers can embed this live SVG badge in GitHub READMEs or documentation:
[](https://benchaudit.org/benchmarks/aa-briefcase-v1-1/)