Long-Horizon Terminal Tasks

Purpose & provenance

The v2.1 task packages connect explicit success conditions, environment assets, inspectable implementations and separate grading. The preview highlights ledger recovery, feedback control and mechanism clearance; the download includes every task in this release. Per-task attribution, change records and component provenance remain in the package.

Public sample

Download dataset ↗

Schema / fields & files

Field and type excerpts; consult the pinned source revision and original documentation.

instruction.md

UTF-8 string: task objectives, required outputs and success conditions; objectives vary by task.

task.toml

TOML object: runtime, resource and verifier configuration; artifacts, when present, is array[string]. Keys and schema versions vary by task; consult the original file.

environment/

Environment build files and input assets; runnable images are not guaranteed or bundled. See RUNTIME.md for compatibility requirements.

solution/ / tests/

Reference implementations, independent verifiers and test assets; isolate them from evaluated agents. Public answers do not form a confidential test.