Research · methodological protocol · 2026
Toll Bench research
Can AI systems deliver real-world human wants?
Explore the WorkThe protocol
Toll Bench: Can AI Systems Deliver Real-World Human Wants? sets out a method for testing whether AI agents deliver stated human wants, using real-world outcomes and supporting evidence rather than claims of completion alone. The PDF protocol is version 1.7, dated October 6, 2026, at 39 pages. The HTML presentation includes amendments through October 17, 2026. This is a public project protocol; the separate arXiv manuscript is described as in preparation.
A July 2026 version 1.4 PDF is an earlier version of the same protocol, not a separate research work.
A separate founding thesis
The Want-Have Collapse: The Human Side of the Singularity is a speculative founding thesis. It is presented separately from this methodological protocol. The benchmark, data, verifier and harness are supporting project infrastructure.
What the protocol measures
Steven’s protocol begins with a frozen finish line, budget, timeline and approval cadence. The accepted attempt also records the agent-system configuration so that later changes do not blur which system attempted the work. Milestone evidence, outcome approval and, for paid successes, settled payment contribute to the result record.
The method separates official successes and recorded failures from active attempts and excluded or unresolved results. It reports completion, cost and time within defined bands. The Toll combines retained cost and the signed timeline; the published verifier checks whether the public inputs reproduce the reported calculations.
Those calculations describe recorded attempts. The protocol treats model comparisons as observational until controlled assignment is carried out, and reproducible arithmetic does not replace evidence about whether a real outcome occurred.