01What a voyage is
One claim, in plain words. The agent is given the claim and an allowlist, and from there it decides where to go. Each thing it does is a leg, numbered in order, written to the database as it happens. A voyage that crashes halfway leaves a truthful partial record rather than nothing.
02The three states
A settled voyage is confirmed, contradicted, or unsettled. The first two require a source and a quote. The third requires a reason. There is no fourth state and no blend, and a row that broke that rule is shown on the register marked unbacked rather than passed off as a finding.
03The quote is checked, not trusted
This is the part that matters. For a confirmed or contradicted verdict the agent must copy the exact sentence that decides it, and that sentence is searched for in the text this voyage actually fetched. Whitespace, quote marks and dashes are normalised; nothing else is. If the line is not there, the verdict is demoted to unsettled and the invented quote is recorded.
It catches real failures. On the first run with this check in place the pilot answered confirmed and produced a fluent sentence about Python's sort() that appears on no page it had opened. That is exactly the failure a language model makes most convincingly, and it is now the cheapest thing in the system to detect.
04The browser is real
A headless Chrome driven over the DevTools Protocol, running on the same machine as everything else here. Not a fetch() with a browser-shaped header, and not a hosted service you would then have to take on trust. It can navigate and read visible text. It cannot click, type or submit a form — an agent that can fill in a form on a stranger's site is a much larger promise, and this one does not make it.
05The boundary
The agent may reach the hosts on the allowlist and nothing else. Hosts are matched exactly. Every refusal is a leg and every refusal is counted publicly. An agent that hides what it was refused looks more capable than it is.
06Limits, by design
A voyage stops after 40 legs or 8 minutes, whichever comes first, and a forced landing is recorded as unsettled with ran out of legs or ran out of time as the reason. Running out of budget is not evidence, and it is not dressed up as any.
07What it costs
Tokens billed are recorded per voyage and multiplied by the published rate for the model in use. The figure on each voyage is derived from what the provider actually reported, not from an estimate of what it should have been.
08What this cannot do
It cannot settle a claim that no allowlisted host answers, and it will say so rather than reach. It cannot read a page behind a login or a paywall. It cannot judge a claim that is a matter of opinion, and phrasing one as a fact will produce an unsettled verdict rather than a view.
And the agent's prose is the agent's prose. Where a voyage carries a written note, the interface labels it as the agent's own words rather than as a finding — the finding is the quote and the source beside it.