For enterprises & AI labs
Tovellan is building human-created tasks and verifiable environments for frontier-agent teams. Our first fictional recovery workflow is a local engineering demonstration. GST-Bench v1.0.0 remains a separate India-specific research release; a hosted lab pilot and live-model evidence do not exist yet.
Or check the numbers first - the GST-Bench leaderboard.
What you get
The cards distinguish a fictional engineering exercise from the released benchmark and operating controls still needed for a real pilot.
The first fictional transaction-recovery environment runs locally with bounded tools and a verifier. A customer evaluation service is not yet operating.
A local test-account journey connects a scoped buyer brief, contributor submission, separate review and authorized delivery. Real participant intake remains closed.
Covered product paths include purpose-bound privacy safeguards and recorded actions. Retention, rights operations and grievance provisioning remain launch requirements.
GST-Bench v1.0.0 is frozen and content-hashed. Its deterministic retrieval-icl result proves delivery and held-out exclusion, not model generalisation.
Product areas
A local exercise connects roles and evidence today. A customer-hosted service and expert operations require separate launch work.
The fictional transaction-recovery environment resets, executes typed actions and records outcome evidence locally. There is no hosted lab run.
local engineering demonstration
The local test-account path covers project scope, contribution, independent review, versioned delivery and buyer comparison. Public participation remains closed.
connected local pilot · no paid work
GST-Bench v1.0.0 is an India-specific frozen release. Its recorded reference result is deterministic delivery proof with protected held-out membership.
released evidence · narrow claim
Engagement boundary
The public site provides the current release record. Any future commercial work requires a separate written scope, security review and agreement.