Back to Project
D

DAYJOB Tax

Project Oct 28, 2026

Surge AI's benchmark of 50 professional tax assignments, testing whether agents investigate records and deliver complete work under detailed rubrics.

Visit primary link 2 sources attached
Surge AI's DAYJOB Tax leaderboard shows Sonnet 5.5 at 10.8 percent pass at one under a 95 percent rubric threshold.
Surge AI's published evaluation, using pass@1 at a 95% rubric threshold. This is a full-task pass rate, not general tax-question accuracy, and has not been reproduced by AITrending. Source: Surge AI
agent-ecosystemai-agentsverification

What happened lately

DAYJOB Tax timeline

DAYJOB Tax evaluates complete tax engagements rather than isolated tax questions. Surge AI describes 50 expert-created tasks and 1,789 workspace files, with assignments spanning eight practice areas. An attempt passes only when it satisfies at least 95% of its task’s rubric criteria. The author’s published results are specific to that evaluation setup.

The public Hugging Face repository contains two sample tasks; access to the full 50-task benchmark requires contacting Surge. Tasks use Harbor format and a Linux document-processing environment, with network access restricted to allowed model API hosts. The dataset card describes synthetic documents modeled on professional records. Public sample scores alone do not constitute a score on the full benchmark.