Long-horizon computer use, on offline copies of real sites
Where the tasks come from
Every task starts from human-written requests collected on Turing's Apollo platform: long, multi-site web tasks written by contributors, each with its own step list and rubrics. Odysseys Offline chains four of them into one longer request and makes it run on offline copies of the sites, with deterministic grading.
- 1 · ApolloContributors write long-horizon web requests with step lists and rubricsHuman contributors (10 behind this sample)
- 2 · ScreeningAutomated feasibility and rubric-alignment checkApollo pre-QC
- 3 · Chaining4 related requests combined into one first-person request; every obligation mappedAuthor agent, blind audit
- 4 · Offline taskSites, checks and rubric built; solved end to end in the VM; 8 runs per configurationAuthor agent, auditor, VM solve
The 20 source requests (4 per task, 4–24 rubrics each) and how each obligation maps into the combined request are under “Built from” on every task and in each task's mapping.md. Contributor identities are not published.
Trajectory examples
Every recorded run, step by step. Pick a configuration and a run; the frame, the action list and the checks beside it follow the run as it plays.
The evidence
The tasks where the two configurations disagree most. Open one for the request, a replay of every run, and how each check was graded.
The sites
Offline clones of the real sites, seeded with captured public records. Open a site for more screens.