WorldAuth Operator · machine-operability

Make one workflow easier for machines to read.

AI agents transact by reading screens designed for human eyes. On our own measured benchmark, that costs 12,036 observation tokens where a machine-native surface costs 235. The public beta offers structured page queries, form plans and synthetic escrow demonstrations. No real money moves.

WorldAuth publishes the Operator Standard — the open standard underneath all of this. The standard is free forever. Start with free evaluation or enquire about a scoped feasibility engagement. Production settlement and commercial certification are not available.

6.0×Cost, screenshot vs machineModelled from measured tokens
5.1–5.9×Energy, same comparisonLow to high conversion band
10%Interface taxShare of a screenshot agent’s input tokens spent reading the page rather than doing the deal
0Pixels sent to the modelThe browser is a transport, never a camera. The structured path still costs tokens and the calculator charges them

Measured, not asserted

A recorded comparison, with explicit limits.

Our benchmark ran on a lean site we built ourselves, which is the most attackable thing we own. So we measured 35 named freight and logistics sites, and then put ten vision agents on the same pages to see what looking really costs. They read 67 screenshots and answered 17 of 40 questions. Bounded queries answered 22 for 49.5× fewer reported tokens. These vision observations are hand-recorded, ten sites and one run; the full structured-read comparison is 3.7×.

The same page also carries a correction: our first attempt blamed the industry for a barrier we had largely created ourselves. That is on the page too.

Sites read structurally30 of 35

Bounded query, median67 tokens

vs the accessibility tree48.1×

Quote, track, contact and book all reachable4 of 30

Routed to a feed or an API, no browser10 of 35

In plain terms

A robot shopping in a store built for humans.

That is what agent commerce is today. The agent opens pages, looks at screens, reads labels and guesses at buttons. It works. Almost all of what it spends goes on reading the store rather than doing the deal.

The Operator Standard describes a machine interface for discovery, value and authority. This site implements an evaluation subset: discovery, bounded page queries, form plans and a synthetic escrow state machine. Live settlement and chartered production operations need further implementation.

Discover an interface. Inspect a plan. Measure the result.

The same task, three ways
Screenshot117,810 tok · $0.0821
Looks at the website the way a person does. Renders pixels, reads them back with a vision model, clicks and types.
DOM / a11y tree50,860 tok · $0.0441
Drives the real website through the accessibility tree instead of pixels. Cheaper than vision, still reading a surface built for people.
Machine surface7,902 tok · $0.0137
Asks the service directly what it sells and commits to a deal. No page, no pixels, no parsing. Three tools in the benchmark, small responses. The live surface has six today and publishes its own definition size so you can check it has not drifted.

Input tokens per task, median of 3 runs, measured 2026-09-16. One synthetic task on one lean site — see the stated limits.

Before and after

Proposed production design — not shipped guarantees.

DimensionHuman web & EDI todayOperator Standard
Cost per agent transaction$0.08 in tokens, on a deliberately lean page$0.014, and no page at all
Freight document cost$0.10–1.00 per EDI document, 10+ documents per loadOne signed intent, cents
Interface timeSeconds of reading per taskMilliseconds — loopback only, not end-to-end latency
DisputesEmails, adjusters, roughly 30 daysSelf-executing terms against sealed evidence
TrustCAPTCHAs, screens, hopeSignatures, escrow, a named charter
Who is liable when the machine errsUnclear, contestedA named legal entity, with caps enforced in code

What you can touch today

Three things. Nothing more is needed to start.

01 — Free forever

The protocol and the kit

The standards, the reference implementation and a conformance kit you can run in one command — 45 Level 1 vectors today, 267 specified across three levels. Anyone can implement it, and anyone can check anyone else against it. The standard is the distribution; there is no lock-in to object to.

Read the standards
02 — Evaluate one workflow

The pilot

A scoped feasibility enquiry for one workflow, followed by agreed measurements, safety boundaries and a written result. No live settlement or production deployment is included in this beta.

Discuss feasibility
03 — Build in five minutes

The sandbox

A key in one minute and your first machine-native intent call in five. Watch the live demo market demonstrate synthetic escrow state transitions with a counter showing the tokens any AI burned to do it: zero.

Get a key

The differentiator

We publish the bugs we found in our own work.

The register lists 69 findings, 58 of them against our own prior builds or our own test harness — including the two that left the first version's trust layer non-functional, and a critical liveness defect that would have let a referee who simply went quiet freeze funds forever.

A conformance authority that only publishes its successes is marketing. Our own Lightning rail fails level two of our own kit, and that failure is on the public matrix.

Findings published69

Found against ourselves58

Conformance vectors you can run45

External auditNot yet complete

Start with one measurable workflow.

Discovery, escrow, chartered liability and unattended dispute resolution, in one specification you can read and one kit you can run: discovery, trust, negotiation, escrowed value, chartered liability, and liveness when the guardians go dark.