How often does an uncorroborated claim ever get confirmed?

Pre-registered 2026-09-13 · readout no earlier than Mon Mar 15 2027 00:00:00 GMT+0000 (Coordinated Universal Time) · status: collecting

The question

When a strange story first reaches the wire carrying only a specialist outlet or a single witness, how often does a mainstream or primary source ever pick it up?

Hypothesis, stated in advance

Corroboration is rare and falls sharply with the opening tier: stories opening at tier 3 are corroborated substantially more often than those opening at tier 4.

What would count as support

Tier-4 openers are corroborated at under 10%. Tier-3 openers are corroborated at a rate at least twice that of tier-4 openers.

Method

Every story cluster is timestamped at the moment we first ingest any item for it, and its opening sourcing tier recorded. A cluster counts as corroborated the first time an item from a previously unseen publisher at tier 1 or 2 joins it. Rates are reported per opening tier with Wilson 95% confidence intervals. Clusters that existed before instrumentation (flagged `backfilled`) are excluded. Reported as an upper bound — see confounds.

Data

thestaticwire.clusters
thestaticwire.cluster_events

Known confounds

Written down before the result, so they cannot be discovered conveniently late.

Syndication is indistinguishable from corroboration without body-text comparison, which we do not store. Seventeen outlets carrying one AP wire story counts as one publisher only if the publisher strings normalise to the same value, which they will not. This inflates measured corroboration, so all rates are upper bounds.
Corroboration can only be observed for stories our source list would catch. A claim confirmed exclusively by an outlet we do not read is recorded as uncorroborated.
Right-censoring: stories opened near the readout date have had less time to be corroborated. Analysis is restricted to clusters opened at least 30 days before readout.

Result

Not yet. Data is still accumulating and the readout date has not been reached. Whatever it says when it arrives — including nothing at all — will be published here.

Amendments

Changes made after registration. Appended, never edited over.

2026-09-13 — Corroboration now requires a RECOGNISED publisher, not merely one at tier 1-2. Before this change any outlet Google News surfaced inherited the keyword wire's tier of 3, so local television newsrooms (KLAS, FOX 5 DC, KARE11, News Center Maine) and DVIDS could never satisfy the corroboration test, while celebrity rewrite sites were presented as vetted specialists. 127 of 139 publishers were affected. Tier assignment moved to pattern rules in src/lib/publishers.ts and items now carry publisher_rated; an unrecognised outlet is ranked conservatively, shown to readers as Unrated, and cannot corroborate. Discovered before any data collection had begun, so no collected observations are affected.
2026-09-13 — Reclustering on 2026-09-13 recreated all cluster rows, which cascade-deleted the small number of cluster_events recorded to that point and left the new rows defaulting to backfilled=false. Affected rows carried no opened_at so none entered the ledger. All clusters have been re-flagged backfilled and their history reconstructed from items.fetched_at. recluster now refuses to run when observed history exists unless passed --discard-research-history, and reconstructClusterHistory will not touch instrumented rows.

Data access

/api/v1/analyses returns this register as JSON, including every entry that found nothing.