Iran says it seized a second US underwater drone

What we read: 4,373 press articles, 25 Aug to 25 Oct.

• Consensus measures reporting agreement, not reality — a uniformly repeated error passes any count.
• Chains are collapsed by syndication tracing; residual shared-upstream risk remains.
• Retroactive: computed 2026-10-25 over the archive, not tracked live.
• Receipt gate: 74 extracted claim(s) excluded — no receipt whose own text supports the claim (the extractor's row binding is verified mechanically, never trusted).
• Relative-anchor supersession: 2 claim(s) that described themselves as the most recent/last-known fact were retired because a claim from the SAME reporting thread, naming the same participants and the same measured quantity, describes a later instance of that subject — see 'retired'. Subject-matched, same-chain detection only: an anchor claim updated by a DIFFERENT outlet is not auto-retired, though its own firstSeen/lastSeen still render so a reader can judge its age.
• Relative-anchor supersession: 9445 later same-thread claim(s) were NOT treated as replacing an anchored claim: 6597 are about different participants, a different measured quantity or a forecast, 2844 could not be read role for role or period for period, 0 compare against another time, 1 characterize or restate the same value and 3 report an earlier or the same period. A claim is replaced here only by a later report on the same participants and the same measured quantity, for a later period; every other pair keeps both claims standing.
• Questions: 12 extracted sentence(s) were questions (a headline asking, e.g. "Trump asked on TruthSocial 'When are the Iranian people going to rise up and fight?'"), not assertions, and take no claim standing in any tier.
• Literal-span counting (#1036): 485 extracted claim(s) were receipted by sentences that do not all assert the same proposition, and were split into 1143 separate claims, each counted only on the sources whose own sentence asserts it. Every count, disagreement and replacement here is decided on the sources' own words; the extractor's sentence is kept as a labelled paraphrase and decides nothing.
• Domain-competence gate: 3386 kernel claim(s) and 441 frame(s) excluded — every supporting chain came from outlets confidently placed outside this subject's coverage domains (aerospace_defense, auto, energy, europe_politics, india_south_asia, iran_regime, …). An outlet counts as competent when its measured home beat intersects the domains the subject's own corpus lives in; outlets with too few observations are never branded and always count.
• Semantic-dome membership: 9223 kernel claim(s) and 2199 frame(s) were source-faithful but failed typed-signature membership and are excluded — a claim enters this ledger only if its signature names a family entity (party, place, or authority) AND either its action class sits in the family's geopolitical profile or it names a family place outright.
• An item is published only when it is RECEIPTED: the outlet's own source link AND the source's exact words (verbatim: a literal span of the stored document, offsets into it, content-addressed by docId). The claim text is the extracted proposition and is never a quote. Anything else is UNRECEIPTED: retained, counted here by reason, never shown; every published count is computed from receipted items only.
• Disagreement pairing: 10 candidate pair(s) with overlapping words and different numbers were NOT shown as disagreements: 4 report separate incidents or readings (another named date or time, another vessel, or reported on another day), 6 assert different propositions (another party or quantity) and 0 could not be matched role for role. Two reports disagree here only when they give different values for the same actor, counterparty, quantity and incident: the same named date, or the same report day when no date is named.
• Event grouping (event@2): 2984 claim(s) are source sentences; 283 were grouped into 120 multi-sentence event(s) and 1733 matched no other claim, giving 1853 event(s). 968 claim(s) could not be matched (untyped action class, no participant read, or no report day), so their count stays per sentence: could not match is not single source. The rule is mechanical and conservative (same typed action, shared participants and wording, no different named person or market, first seen within 1 day, no conflicting date, time, vessel, means (mine, missile, drone, boarding, gunfire) or value, every member matched to one anchor), so it under-groups: it measures reporting agreement about one incident, not verification.