What We’re Building Next

Most of PROBE's sources answer automatically. Some do not: either they are manual-only links, or they keep returning challenges and unclear pages instead of a decision. Rather than guessing why, we measured it.

A census of every unanswered source

On 15 October we ran a census of 80 sources: every manual-only source plus every automatic source that produced no decision in our latest full test run. Each was checked with known existing accounts and random names that should not exist, three times from a home connection and once from the network PROBE is hosted on. We also recorded alternative public endpoints and the platform's terms. The result is that every one of the 80 has a recorded reason; none are left unclassified.

• 35 sit behind anti-bot protection that challenges plain requests even from a home connection.
• 15 stay manual because of their terms, even where a check would technically work. See Honest by Default.
• 8 need an official API key, one of which is already being set up.
• 7 show the same login page or shell to existing and missing users alike, so only an account session could tell them apart.
• 6 depend on the network: 3 answer from a home connection but not from a data centre, 2 only resolve in a real browser, and 1 rate-limits a single address immediately.
• 3 cannot be answered from a username alone or are waiting on someone else: two need more than a username to identify one account (a full game ID, for example), and one awaits written permission from the platform.
• 1 appears to be down from every network we tried.
• 5 are fixable for free. One had only failed transiently. The other four are covered below.

Four sources nearly ready

Booth, Patreon, BIGO Live and Treehouse each produced a clear difference between existing and missing accounts, from both networks, using ordinary identified requests. Every check against known accounts and random names came back correct, and the ambiguous cases, such as inactive or private profiles, return unknown rather than a guess. They go live after the same release checks as any other change.

Answers that hold still

A blocker seen once might be noise. The census now runs every night from both networks, and a source's recorded reason only changes after three consecutive nights agree. That keeps us from flipping a source back and forth on a single bad response.

Canaries that reach a person

Every night, each automatic source is also checked against known existing accounts and random names that should not exist. A false found on a missing name is the costliest error a username tool can make, so it raises an alert immediately. A known account reported missing, or a source that stops deciding, raises an alert if it happens two nights running. Alerts go to a person by email, not into a log nobody reads.

Then: infrastructure, carefully

For the sources that need it, we are evaluating paid network infrastructure such as residential connections and managed browsers. The census tells us exactly which sources need which tier, which turns a vague wish into a costed plan. The rules do not change: identified requests, unknown when the evidence is unclear, and no infrastructure for sources whose terms say no. Some limits are not technical at all. A login wall or a terms restriction cannot be bought around.

And measurement

Two pieces of measurement work come next. The first is a fresh accuracy campaign on the current classifier, since the figures in earlier posts come from earlier runs. The second is a comparative benchmark against other username tools, using matched sources, usernames and timing. We will not make comparative claims until that benchmark exists.

We are not attaching dates to any of this. When each piece lands, we will publish what we measured, including the parts that did not improve.