FAQ
Straight answers on the frontier-agent direction, local exercise, released benchmark, current availability and data safeguards.
Tovellan is building human-created tasks and verifiable environments for frontier agents. Evaluation and post-training teams are the first buyer hypothesis, with US-first discovery. Demand has not been validated.
A fictional purchasing recovery environment is implemented locally. The buyer, contributor, reviewer and delivery journey is undergoing acceptance repair. They are scripted test-account exercises, not a hosted customer service or live-model evaluation.
No. Evaluation, development, training, redistribution and publication require separate rights. The fictional local exercise permits local demonstration only; protected challenge or GST held-out material is not reused for training.
GST-Bench is an owned GST and accounting benchmark. The released v1.0.0 corpus contains 48 staff-authored tasks, split 34 train and 14 held-out. The release is frozen and content-hashed, while held-out membership and content remain private. See the release record.
The deterministic retrieval-icl reference moved from 0.0000 to 0.7083, a delta of +0.7083. It proves dataset delivery and held-out exclusion. It is not evidence of live-model generalisation.
Held-out content, identifiers and membership remain private. Internal operating details and any additional methodology disclosure are also kept off the public site.
The public site provides GST-Bench v1.0.0 release evidence and a candid description of the fictional local engineering preview. Managed pilots, custom data delivery and live-model evaluation remain unavailable.
No buyer price table or active contributor rate card is published. Any future engagement requires a separate written scope and agreement; this site is not an offer or payment promise.
Customer workspaces are designed around tenant isolation, encryption and controlled access. Detailed architecture is available only through an authorised security review. See the security page.
Covered product paths include purpose-bound privacy safeguards and recorded actions. The public preview has no accounts or personal-data forms; retention, rights operations and grievance provisioning remain launch requirements. See the privacy notice.
Review the released evidence and the current availability status. The pre-launch domain does not provide an enquiry channel.