FIELD NOTES / AI ENGINEERING

AI application evaluations

Beyond a polished answer: what did the agent actually deliver, which boundaries held, and what survived independent review?

Available in Chinese and English.

1 published case

CASE 001 / LONG-HORIZON ENGINEERING

A long-task agent: strong execution, incomplete delivery

Two implementation stages, one delivery under review. Passing engineering checks coexisted with critical boundary failures and incomplete user journeys.

dsh DeepSeek V4.1 Flash

September 11, 2026Read the technical evaluation

Delivery performance60/ 100Blocked

A weighted expert score, not a completion percentage.

Real work

Existing projects and approved implementation plans, not isolated demo prompts.

Checkable evidence

Separate implementation, passing checks, user journeys and release decisions.

Explicit limits

Disclose identity sources, missing observations and what a score cannot establish.