Home / Checking what it gives back
Reopen every source before believing it
A research agent reports what an index shows, not what a live page contains today, and presents a dead link with the same confidence as a live one: any source meant to support a conclusion gets reopened before being cited.
A research agent answers with the same confident tone regardless of the actual state of what it reports. It presents a dead link and a live link in the same way, because from its point of view both are only a match found in an index, never a page actually reopened and read.
What a raw finding does not prove
An identity search sweep run by eleven agents in parallel returned 103 findings, each accompanied by an address presented as evidence. Reopening one by one the 94 that carried an address changed the picture: 33 were dead or redirected, and 9 pointed to a namesake of the actual subject rather than the subject itself. Nearly half the batch was unusable, even though every item had been presented with the same confident tone as those that held up. The figure following this paragraph gives the exact numbers for this gap: the raw count, the count of addresses dead or redirected, and the count of those pointing to a namesake, on this same sweep.
The same risk exists for a verbatim quotation, presented in quotation marks as proof that an agent really read the source. A quotation that looks like work already done is more dangerous than a bare address, precisely because it inspires more confidence. Two signals often give it away: a quotation written in a different language from the page it claims to quote, and two quotations attributed to sources that look interchangeable. The same refutation mandate applies to this case: hand the quotation to a different agent, tasked with finding the exact passage in the page actually reopened, never with confirming that it sounds right, exactly the principle detailed in a green result proves nothing until it has itself been proven.
Reopening under a refutation mandate, not a confirmation one
The fix is a separate pass, handed to an agent different from the one who found the source in the first place. Its mission is not to confirm the finding but to try to refute it: does the page really respond, does the content match what was claimed, is this really the right subject rather than a namesake. A single request is already enough to settle a good part of the batch:
$ curl -s -o /dev/null -w "%{http_code}" https://example.com
200
$ curl -s -o /dev/null -w "%{http_code}" https://example.com/page-inexistante-000
404
$ curl -s -o /dev/null -w "%{http_code}" http://ce-domaine-n-existe-pas-du-tout-12345.test
000
Three addresses queried, only one code 200. Code 404 signals a page that responds but no longer exists at that address, code 000 signals a domain that did not respond at all: two different ways to fail, and already two of these three addresses ruled out without reading a single line of content. On an identity or attribution search in particular, the disambiguation criterion, a specific employer, a specific date, a specific place, is spelled out aloud before starting, and the absence of that criterion sends the verdict back to undetermined, never to confirmed by default.
Reporting a raw count without the surviving count amounts to reporting nothing at all. A zero gap between the two, on a batch of several dozen findings, is not proof of a flawless sweep. It is the sign that reopening was never really attempted.
A real sweep, reopened address by address
Reopening under a refutation mandate
A research agent queries an index and extracts a list of addresses from it, each accompanied by a short excerpt judged relevant. It presents this list as proof of its research, built entirely from the content returned by that index.
Write, in one sentence, what this result establishes, and in one sentence what it does not establish.
What this establishes: This list establishes that an index contains a match for each of the addresses cited.
What this does not establish: It does not establish that these addresses still respond today, nor that they point to the right subject rather than a namesake, since none was reopened.
The three most common miscalibrations
- Too broad This list establishes that every address cited genuinely leads to an active page that confirms the match.
- Too narrow This list establishes nothing at all, an index never being able to contain useful information.
- Beside the point This list shows that the agent spent more time on this search than on the previous one.
- A research agent treats a dead link and a live link with exactly the same confidence, since it only reports an index match, never an actual reading of the page.
- On a real sweep, reopening a batch of addresses knocked out nearly half the batch, dead, redirected or pointing to a namesake of the actual subject.
- A verbatim quotation in quotation marks inspires more confidence than a bare address, which makes it more dangerous when it is fabricated rather than genuinely read.
- A source meant to support a conclusion is handed, for reopening, to an agent different from the one who found it, with an explicit mandate to refute it.
- A raw count identical to the surviving count, on a batch of real size, signals a reopening that never really happened rather than a flawless sweep.
Take three sources an agent has just cited to you, reopen them yourself or hand them to a different agent tasked with refuting them, and note how many genuinely survive being reopened.