Case studies · 19 October 2026 · 4 min read
Five reports can still be one account
I split 36 FORGE pins into 216 facts. The useful result was not a confidence score, but a clearer account of who said what, and what remains unknown.
In my first audit, the problem was a report that joined facts the sources did not join. In the follow-up, I had to correct my own checking. This time I asked a narrower question: when a map pin cites several reports, what does each fact actually rest on?
Five links can lead back to one police statement. A community voice might confirm a death without knowing its cause. One report might give the date while another supplies a street. An overall label cannot carry all those differences.
What I checked
The private audit froze 36 pins across three provinces and split each into six claims: deaths, date and time, place, method, reported motive and legal status. That made 216 fact entries. The newer Retreat pin was outside that frozen set and remains labelled “not yet audited” in this layer.
The unit was the origin of a claim, not the website carrying it. Several outlets repeating one spokesperson did not become several independent confirmations. A statement without an identified origin did not become first-hand reporting just because it sounded definite.
The audit kept quotations, archived links, conflicting accounts and differences in precision. Agreement on a suburb did not establish a particular street. A discovery date did not establish the time of death. A reported motive stayed attributed and uncertain.
Two different judgments
Source reliability and claim credibility answer different questions. Have I documented this origin’s track record? And what supports this particular claim?
FORGE used a private source-and-information grading rubric, adapted for news reports rather than presented as an official certification. Every recorded source-grade entry remained F: reliability could not be judged from a documented FORGE track record. That was not a verdict that every source was unreliable. It meant the required history was absent.
A claim could still have stronger or weaker support. The private credibility totals, rebuilt from the files for this draft, were:
| Private category | Fact entries |
|---|---|
| Confirmed by independent origins | 1 |
| Probably true | 130 |
| Possibly true | 66 |
| Doubtful because reports conflict | 2 |
| Improbable | 0 |
| Cannot be judged | 17 |
These are judgments under an adapted rubric, not measured probabilities. The 221 source-grade entries are repeated entries attached to claims, not 221 distinct sources.
What changed on the map
On 32 of the 36 pins, at least one fact’s basis differed from the overall label through a disagreement, a non-police basis or a narrower detail. This count excludes missing or undetermined facts. It is not a count of 32 false incidents.
The live FORGE map now has a drawer section called “What each fact rests on”. It shows plain-language bases, source and archive links, and recorded conflicts. When police are a party to an incident, the drawer says so; their account is not presented as an independent finding.
The wording corrections were small but important: a date was no longer attributed to a police statement that did not supply it; a suspected motive stayed suspected; a community member was no longer described as family; and a school post quoted by an outlet was recorded without pretending I had obtained it separately.
What this does not establish
There was one main grader: Maya generated the grades, and I approved them. The comparison sheet was not an independently graded sample. Its apparent agreement figures cannot validate the method. Keeping that correction visible matters more than keeping a neat success story.
The sources also have limits. News selection is not a complete incident census. Police accounts can dominate coverage. Some locations are approximate, some dates are discovery dates, and working translations have not all had independent language review. More careful attribution does not remove those gaps.
The software checks answer a different question again. The saved browser results record 111 successful drawer checks: 37 pins at each of three emulated screen sizes. The full code suite recorded 257 passed and two skipped. Those results support the interface implementation, not the truth of the incidents. They are not physical-device certification or a comprehensive accessibility audit.
Practise the distinction
Exercise 006: Five reports, one origin gives you five invented articles about one invented incident. Your task is not to catch an automated report making a bad link. It is to trace each claim backwards and decide which reports add an origin, which repeat one, and which leave you unable to judge.
The map should not ask readers to trust a number. It should make the route back to the evidence easier to inspect.
Evidence and review note
This draft draws on the owner’s frozen private audit, its rebuilt counts and correction queue, the published FORGE implementation, and the saved browser and test-suite records described in the project’s provenance register. The private records are not publicly linked here; the totals are therefore not fully reproducible by readers from this post alone. Before publication, the author should review the numbers and wording and decide whether a redacted audit export can be released. No victim or suspect is named in this post.
Related posts
Checking the check: two mistakes in my own work
I audited a machine's report, then rebuilt it with stricter checks. My audit and my rebuild each held a mistake. Here is what they were and how to correct work in public.
A polished report is not a checked report
I let my own OSINT engine write a brief on an arrest, then checked it by hand. The arrest was real, but much around it did not hold up. This is what the human step found.
Same handle, different person: the trouble with username matches
Username search tools tell you a page exists, not who owns it. Here is how to weigh handle matches, spot false positives and false negatives, and avoid naming the wrong person.