Hard and independently authored, each with its own oracle and tests.
Analysis
100 independently authored Terminal-Bench-style tasks.
Applied Work Labs publishes a selected set of 100 hard tasks from a private inventory of 448. Each row is an authored world: instruction, Docker environment, oracle, and tests. The analysis surface follows a corpus → trajectories → reward → optimisation path. Solve chips on this site are simulated binary solve counts for Model A and Model B, shown to demonstrate the analysis views, not to report model performance.
Authored tasks held back from this preview. The 100 above are selected from it.
- Domains
- 8
- Languages
- 20
- Format
- Terminal-Bench 2.0
- Reward
- solve rate
- Arms
- Model A · Model B
Browse the corpus
One hundred unique tasks, searchable by domain and language, each with a binary solve count for two arms.
Open the tableRead a trajectory
Open a task, read the instruction, and inspect the five trials on each arm. No harness step log is claimed.
Open the inspectorInspect the reward
Binary solve counts only. See the five trials on each arm and the n/5 summary beside the strip.
Open the signalScientific computing 24 · Security 20 · System administration 14 · Build 11 · Data processing 11 · Machine learning 10 · Games 9 · Debugging 1
Method pages remain inspectable.
Project A defines how a Terminal-Bench-style package earns release. Project B is a planned synthetic repair case. Those pages do not change because this gallery exists.