In the wild
What happens when the verifier meets real posts from DappGo News and Threads?
Read the findingsDevelopment probes. Post-hoc labels. Not a held-out benchmark.
Caught in the wild
Displayed numbers that conflict with primary sources. These are independent reference checks, separate from the agent’s outputs below.
A displayed unit error does not prove who caused it. Auto-translation is recorded where present; original-language causation was not verified.
Claims without a resolved reporting period remain unscored. They are excluded from these conflict cards.
How much could it check?
Abstentions, skips, failures and budget stops stay visible. A supported output is not a measure of full-post coverage.
Threads posts came from a selected Top-search sample. Reliable engagement counts were not recorded, so this is not a measured high-engagement population.
Only independently resolved primary-source items enter the truth denominator. Expected abstention on out-of-scope items is a policy check; unresolved items are not financial truth labels.
Latency is observed wall time, including retrieval, caches and retries. Provider workers and some tests overlapped; this is not a controlled speed comparison. Budget stops have no latency sample.
Costs are recorded token-rate estimates and retained reservations, not invoices. Cost per item uses every sampled item, including skips and failures.
News limitations include issuer confusion between Hon Precision and Honeywell, plus retained paraphrase errors. Zero false support would not exclude erroneous contradictions.
What we fixed after this test
A recheck of known failures is post-hoc development evidence. The original results stay above.
No completed fix recheck is included in the available page data.
Follow the receipts
The page is compiled offline from retained inputs, reference labels, all result rows and accounting summaries. No model runs happen here.