Raw headlines in. Verified claims out.

Every gate the hourly hunt runs an article through, in order, illustrated with real items pulled from the live archive — not invented examples. Most examples below follow one tracked subject, Ilia Topuria (Subject C, the shorthand label used throughout docs/architecture-overview.html) — named in full here, as everywhere else in the repo: the watchlist is committed, so there is nothing for a label to hide. Everyone else named below is real.

This page covers the funnel only — discovery through claims. For the whole system (the Telegram service, operations, cost, the autonomous improvement loop) see docs/architecture-overview.html in the repo; for how it's tested, see docs/test-suite-overview.html.

Each stage below opens with a grey In code strip naming the files that implement it. One thing to hold onto while reading them: §1 through §7 all happen inside a single function, huntSubject() in hunter.js, which runs once per tracked subject per hour. The gates are not separate services or separate passes — they are consecutive if statements in one loop over at most five articles. Line numbers drift with edits; the file names don't.

§1Discovery — two rails, one window

Every hour, for every tracked subject: Google News RSS, queried once per language edition the subject gets covered in (English always; Ukrainian and Spanish where it applies), plus six direct publisher feeds — UFC.com, MMA Fighting, Bloody Elbow, Sherdog, Sport.ua, Marca — fetched once per run and filtered per subject by matching surname stems in both scripts. Both rails only look at the last 24 hours.

One real run: 6 outlets fetched, 180 items pooled across all three subjects — of which 5 were unseen for Topuria. — docs/checkin-log.md, 2026-08-09

The 24-hour window overlapping the hourly cadence ~24× is deliberate: a single failed run loses nothing, because the next run re-covers the same ground.

§2Gate 1 — Have we already seen this URL?

Exact match on url or its decoded resolved_url against everything ever stored → dropped silently, no database write at all. This is the only gate that costs nothing — no embedding, no fetch, no LLM call.

The database has no record of this gate — dropped items are never written. Cloud Logging does: across every scheduled run since 2026-08-07, 1,248 URL checks → 90 new (93% already known, dropped — 1,158 of 1,248). See Fig. 1.

Survivors are capped at 5 unseen items per subject per run, newest first, applied after this gate — so on a busy day the newest 5 unseen items go through now, and whatever's left over is simply picked up next run. Nothing is lost, only delayed.

§3Gate 2 — Same story, different link?

The survivor's headline is embedded and compared against the subject's recent items. Cosine similarity ≥ 0.80 → held back, never posted — but not discarded: it inherits whatever claim its nearest neighbour already supports, recorded as corroboration for free, no LLM spent. The threshold is measured, not guessed: a real translated pair of the same story scored 0.841; unrelated same-subject pairs never rose above 0.702.

40 of 95 items were held here — the single biggest gate in the funnel.

Official sources (UFC.com) are exempt from this gate specifically — a confirmation headline is, by construction, near-identical to the rumor it confirms, so holding it here would swallow the exact transition the claims layer exists to catch. And inheritance itself is checked: if an item actually reads closer to some other claim than the one it's about to be linked to, it stays held but unlinked rather than becoming a wrong piece of evidence. Both mechanisms are detailed in docs/architecture-overview.html §4.

§4The body step — reading past the headline

Only for items that survived both free gates. Google's wrapped URL is decoded to the real article address, then the body is extracted through a ladder — cheapest and most reliable first — and every attempt is logged, success or failure:

Those 27 do not add up to the 55 items that reach this step, and the gap is honest rather than missing data: the body_via column that records which rung won only shipped partway through the archive's life. Items processed before it exist with a body and no rung — 49 of them got their text later from a one-off backfill script, which is also why the mention counts in §6 could be recomputed for articles nobody re-fetched. Telemetry starts when you add it, not when the behaviour starts.

Fixed by dropping a paragraph identical to one already kept, before anything downstream counts mentions — both at rung 2 and at rung 0, where a publisher's own feed ships the whole article and would otherwise be flattened to one line before any duplicate could be spotted.

§5Gate 3 — What is this article actually about?

One Haiku call per survivor: the headline, source, a body excerpt when there is one, and every currently active claim for the subject. Forced tool use means the answer is always one of five verdicts — never free text to parse. The same call answers a second, independent question in the same breath: how prominent is the subject in this article — central, supporting, or passing. The verdict says what the article asserts; the role says who it is about. §6 is where the role earns its keep.

Two things the tables are easy to misread. "Goes to §6" is not "posts with its headline intact" — §6 decides that, and may drop the headline entirely. And central is not a promotion: only passing changes any outcome. The role is nonetheless recorded on every article the matcher sees, including the ones it drops and holds, so its agreement with the count rule can be measured on real data later.

The same real thread — Ali Abdelaziz predicting Usman Nurmagomedov beats Topuria — supplies most of the examples below: it produced a NEW, two MATCHes, and the corroboration stack §7 ends on, across five days and four outlets in two languages.

NEW — a claim-worthy fact not seen before

MATCH — the same fact, sighted again

Both are linked as evidence for claim #13 and neither is posted — the group never sees a repeat of a fact it already has. Worth noticing what is not in this pair: two further articles in the same thread never reached the matcher at all, because Gate 2 recognised them as near-duplicates one step earlier and they inherited the claim for free (§3). Identical corroboration, one gate sooner, one Haiku call cheaper — the funnel doing exactly what it is shaped to do. §7 shows the finished stack, with each source labelled by the gate that caught it.

WRONG_SUBJECT — not about this subject at all

Never posted, never a claim. The Eurosport item is the instructive one: the matcher read this correctly on the first pass — the exact case §6 is about is the English version of the same underlying Salkilld story slipping past on a technicality one gate later.

NO_CLAIM & UNSURE — posts, but no fact recorded

NO_CLAIM: about the subject's world, but asserts nothing claim-worthy — division context, reactions, lifestyle. UNSURE: the matcher genuinely can't tell, and abstains rather than guess. The Ferran Torres item is the archive's only UNSURE so far, and it earns the label — a viral nightclub clip is lifestyle noise the system ignores by design, except that Topuria's own manager posted a mocking video reply, which is a quote of a sort. The matcher declined to decide whether that counted. Both verdicts post through the normal digest exactly as if no claims layer existed at all — which is where §6 picks up, with the role judgment riding along.

§6Digest tier — does this earn its own headline?

An article that opened a claim keeps its own headline, full stop — whatever fed a claim has earned a line. Everything else that posts — NO_CLAIM, UNSURE, and the NEW verdicts whose claim type the digest ignores — goes down a three-rung ladder. The rungs run in order and the first one that answers wins; nothing below it is consulted.

There are only two answers. The article keeps a bullet of its own, or it is folded — the code's word is tangential — meaning it is merged with every other demoted article from that run into one shared line naming nothing but their outlets. Folding is a demotion, not a rejection: the article still posts, and the reader can still reach it. It just stops taking up a headline.

Reading that ladder answers the question people ask first: the two rules are not weighed against each other. A central or supporting role does not answer rung 2, so the article simply drops to rung 3 and the counting rule decides it exactly as it did before the role existed. The role can only ever add a demotion, never cancel one. That asymmetry is deliberate: the count threshold was read off real archived articles, while the role is one model's opinion and has not been measured yet, so it is allowed to accuse but not to acquit. When the matcher is off or its answer is missing, the role is simply absent, rung 2 stays silent, and the ladder is byte-for-byte the rule that shipped a day earlier.

Both outcomes post. The difference is how much of the article the group is shown. Here is the real Kutateladze item from rung 2, rendered each way inside the same digest message:

A bullet is the whole article as the reader meets it: the publisher's headline verbatim — translated into English first if it isn't — the outlet as a link, and how long ago it ran. Folding keeps only the link. The headline is thrown away, the timestamp with it, and what remains is the outlet's name on one shared line at the bottom of the message, which every folded article that run shares. Two folded articles become Bloody Elbow · MMA Junkie on that same line. So the reader can still get to the story in one tap; what they no longer get is a headline competing for attention with news that is actually about the subject.

One outlet with two folded stories in the same run gets both links, numbered newest-first: Bloody Elbow (1) · Bloody Elbow (2). Until 2026-08-10 the line showed only the newest, which read cleanly and quietly cost something — the message is the only place a folded article is ever offered, and Gate 1 guarantees no later run offers it again, so the older story became unreachable. A bare repeated outlet name reads as a bug; the index is what earns the second link its place. The numbers appear only where an outlet actually has more than one, since a lone Sherdog (1) would imply a sibling that isn't there. Nothing was ever lost from the record either way: every folded article had its own archived row, with its own URL, before the message was built — and a folded article can never be a claim source, because anything that opened a claim skips the ladder entirely.

There is a third outcome the ladder doesn't decide. If everything a run found for a subject folded, there is no bullet for that shared line to sit under, and a message of nothing but a header and a row of outlet names is precisely the noise this stage exists to remove — so nothing is sent for that subject at all.

One clarification the two cards above invite. The ladder's answer is stored as digest_tier, and it has exactly two values: tangential, which always means folded, and main, which means only not folded — not "shows its headline." Anything that opened a claim is set to main without consulting the ladder, and may then leave by a route that discards the headline anyway, because once the system has extracted a fact it trusts its own sentence more than the publisher's wording. Five shapes reach the group in all:

And the ways an article appears nowhere: MATCH held as evidence (§5), WRONG_SUBJECT dropped (§5), a Gate 2 near-duplicate (§3), and the all-folded run above — which also writes the suppression back to the row it had already marked as posted, so a later measurement doesn't count as broadcast something the group never saw.

Both of this stage's real cases were decided at rung 3, which is the only rung that existed until yesterday. The tier is new enough (2026-08-09) that it has fired just 6 times in the live archive: 5 kept their headline (including the Salkilld item above, pre-fix) and one folded — a Mark Zuckerberg viral-video story naming Topuria once in passing, which was the whole of that run's Topuria news and so triggered the third outcome above: nothing was sent for him at all. Rung 2 is newer still (2026-08-10), and it exists because of the one case rung 3 could never have caught: an article about Guram Kutateladze that named Topuria twice — once as a teammate, once as a man in someone else's corner, pure background color both times — and so cleared the ≤1 threshold honestly. That was the first item the tier got wrong on its own terms rather than on a body handed to it broken.

§7Claims — the distilled output

Everything the funnel does exists to feed this table. A claim is one row per real-world fact: a canonical English sentence, a type, a status that starts as rumor and moves to confirmed only when an official source asserts it. Every corroborating sighting above — however it was caught, by the matcher, by Gate 2's inheritance, or by a retroactive backfill — is recorded against exactly one such row.

Two of those statuses are all the system has ever written. The schema (schema.sql, the claims.status column) reserves five, and it is worth being plain about which three are vocabulary rather than behaviour:

No code path writes the last three: the matcher writes only rumor, and the one promotion that exists moves rumor → confirmed. Their only live consequence is already wired, though — the query that assembles a subject's active claims selects status IN ('rumor','confirmed'), so writing any of the other three would quietly drop that claim out of the candidate list the matcher is shown, and the fact would stop being matchable. That is the retirement mechanism, built but never fired. Nothing yet decides when a claim has been denied, gone stale, or been replaced — that judgment is the unbuilt part.

Read the right-hand column and the whole funnel is visible in one stack. One article opened the claim. Two were caught by Gate 2 as near-duplicates and inherited it for free, no LLM spent. Two reached Gate 3 and the matcher recognised the fact it already had. One — the Spanish article — predates the claims layer and was linked when the archive was replayed through the matcher after the fact. Six sightings, one row.

6 sources · 4 outlets · 2 languages · 5 days · still rumor — no official source has weighed in on a prediction, and structurally none likely will.

Across the whole archive: 10 claims, 46 corroborating links, zero confirmed yet. The rumor → confirmed transition is real and wired — one threaded "✅ Confirmed" reply the moment UFC.com covers a pending announcement — it just hasn't fired live. That state machine, the ceremony posts, and everything downstream of a claim (§8 onward) is the other document's territory: docs/architecture-overview.html §6–§8.