2026-10-27

• Show claims tied to stories. What the shows said now reaches the story pages ("🧩 How it's being told") and the front page's cards ("🧩 told as"). A show claim is tied to the story on the front pages it is nearest in meaning to (mxbai embeddings), among the stories running from two days before its episode to a day after, only when it is at least 0.75 alike and 0.04 closer than the next story; otherwise it is tied to none. Calibrated on the 392 Bluesky narratives the story model had tied: the nearest story agreed with the model's 85% of the time, and fell in the same saga 92%. About 1 in 14 show claims clears the bar, so the panel shows mostly on the bigger stories, and a tie can be wrong.
• "Came up on", not "told by". Where the site counts the left, center and right shows behind a motif, a show counts when one of the motif's claims was heard there: told, reported or argued against. Tellings don't yet record which, so these counts say where a story came up, not who believes it.
• Arguing against a story. The site used to say a motif "is the same whether people tell the story or argue against it". Not quite: an answer often has a shape of its own (Voter fraud is answered by Voter suppression, Lawfare by No one is above the law) and is filed under the answer's motif. Where no answer has a motif yet, the arguments against a story are filed with its tellings. The site's notes now say so.
• Each show's signature. A show's page lists the motifs it tells far more than the other shows: weighted log-odds with an informative prior (Monroe, Colaresi and Quinn 2008), the prior a weight of 10 tellings spread over the motifs by their share of all tellings. Its nearest shows are by cosine similarity of their motif counts, whatever their lean.
• The motifs' words in headlines and on TV. Motif names and the tellers' phrases (two words or more) are looked for word for word in the last three days' headlines and captions. Plots and everyday talk are left out: a plot's name is what a plain report says ("found dead", "mass shooting"), where an argument's or a theory's ("rigged election", "radical left") is the telling itself. A headline can still use a motif's words to report on it.
• The motif map is the aisle. The force-directed map is replaced by a chart: each motif placed left to right by the share of its shows that lean left or right, (right − left) / (right + left + 2) so a motif few shows told sits nearer the middle, and up by its tellers on a log scale; answers are arcs. A motif rises when told at least 3 times in the last 7 days and at least twice as often as in the 7 before; it fades when told a third as often or less.
• Out of the public index: watch lists and donation appeals. Straight news and any other watch list (sinks for plain reports, kept so the models learn what isn't a story) no longer appear among the public motifs or their rankings. Folklore leaves out retold "claims" that are non-political appeals for money: on Oct 27 the top card was a pet-charity appeal posted once each by 285 accounts with no replies, a spam wave's shape.
• One side's wording. "D.C." was read as the words "d" and "c", so the left's "D.C. golf" and the right's "DC golf" each looked like a phrase only that side used. Abbreviations with dots are now read as one word, and light verbs ("puts", "takes", "makes") count as headline grammar, not framing.
• Saga names. The naming prompt now asks for neutral words true of every part. Two names were corrected by hand: "California noncitizen voting scandal" (now "California noncitizen ballot dispute") and "Trump attacks El-Sayed and Minaj", whose first part was Trump praising Minaj (now "Trump's Michigan rally").

2026-10-26

• Show claims grouped around a centre. A claim told on two or more shows is now grouped with only the claims close to it and to each other (mutual neighbours at 0.84), around its most central telling, rather than chained through any pair alike enough: at 0.78 one chain had joined 4,436 tellings. The words a claim was filed under are kept when its group is redrawn, so a filed claim isn't filed again as a new one.

2026-10-25

• Show claims need to be 0.84 alike to count as one claim retold (from 0.78). Judged by hand on 70 pairs: at 0.78 to 0.80 about 4 in 14 pairs were one claim, at 0.84 to 0.86 about 10 in 14, above that nearly all. At 0.78 about half the two-show claims were two claims glued together.
• Claims worded as a dispute are split, one claim a side. "Arguments exist over whether..." isn't a claim anyone tells; such a claim is now split into each side's own claim, and the claim writers are told not to word one that way.
• The Washington Post and The Washington Times are read again, the Post through Google News and the Times through its headlines feed, after both stopped coming in. A scraper coming back after failed runs no longer stops the hourly run.

2026-10-24

• One outlet's failure no longer fails the hourly run; its headlines are just missing for that hour.
• Visits are counted with GoatCounter: no cookies, nothing personal kept.

2026-10-20

• Transcripts without Whisper's repetition loops. Whisper decoded each stretch of audio with the text before it as context, its default, and a repeated phrase fed itself: "it is all happening, it is all happening…". 951 of 2,016 transcripts held a loop of five or more repeats, a median 12% of their text, and the speech under a loop was lost. From Oct 20 each stretch is decoded on its own (condition_on_previous_text off), and transcripts carry the label "no-carryover". Claims from shows before then may miss what was said under a loop.

2026-10-18

• News briefings' presenters aren't tellers. In the outlets' daily news briefings (BBC Global News, Reuters World News, WSJ What's News, Up First, CNN 5 Things, Morning Wire, NBC's Top Story), what the presenter reads out is the news, so it no longer counts toward a show claim being retold, and doesn't reach the motif index. The clips they play and the guests they talk to still count, as on any show. The hourly newscasts were already left out.
• The checker's quick tabs: ✅ sure claims and 🚫 likely misfits. The claims to judge listed every sure filing in one fold above the claim, 418 on Oct 18, and passing one filing of a claim left the claim to come back for the rest. Now two tabs of their own list them, a claim a row, a page at a time (10-100 rows a page). ✅ sure has only claims whose every waiting filing is past the sure line (277 of 582), so passing a row finishes the claim. 🚫 misfits has filings the confidence model puts under 30%, none ticked. The model is never sure of a no: held out, the person took out 74% of the filings under 30%, and 86% even under 10%. So the person ticks what to take out; what's left on the page either waits for its card or is checked as fitting. The model's held-out test now records how often the filings under that line were taken out.
• ⏸ Hold. A claim can be put on hold, to look at later: it leaves the claims to judge, ✅ sure and 🚫 misfits for a ⏸ held tab, its filings still waiting, until it is taken off hold. Before, a sure claim left unticked came back each time the list loaded. A hold is no verdict: the confidence model doesn't read it.

2026-10-15

• A saga's parts in the order of their latest news. Parts were numbered in the order they broke, but a part's title tells its latest news (a live story's title moves on with it), so a long-running part came first under today's title: the Fort Hood saga's part 1, "Trump cancels livestream of Fort Hood shooter's execution", was its newest development, after part 5. Saga pages, story pages and the front page's saga cards now order parts by when each was last on the front pages; each still shows when it broke and how long it ran.

2026-10-14

• Flabby motifs opened into their tellings. Pandora's drafts (Gemma's reading of each of the person's motifs of 10+ claims: the distinct stories told inside it, each with a note and its claims) are in the checker's 🧩 tellings tab. A telling the person makes a motif rests on the broader one and takes its ticked claims, which leave the broader one (or stay in both); this is not read as "not this motif" by the confidence model, since the claims still tell the broader shape, only more exactly. Each decision is also Pandora's verdict.
• A story is titled once more when it leaves the front pages. Titles were written only while a story was live, from its newest headlines, and stood for good once it left; but headlines can move to another story afterwards, so a part could keep a title for news it no longer held (saga 48: "Senators Collins and Jackson debate in Maine" stood over ten Georgia headlines). Now, two hours after a story leaves, it is titled once from the headlines it ended with, saga parts first, the past week's stories over the next day or so. The title prompt now also says to give people only the roles the headlines give them (the titles had made Rep. Mike Collins a senator).
• The confidence model no longer learns from the answer. One of its signals counts the motifs a claim is in. In training it was counted as the index stands now, so a filing the person had taken out was already missing from the count, and the model could partly read the person's verdict off it (the Danaids, Oct 12: worth about 0.01 of AUC; the sure line held at 95% right). Training now counts the claim's motifs with the judged one among them, as they stood when it was judged; at scoring an unchecked filing counts itself the same way. The model retrains on it daily.

2026-10-13

• Who's in the news, today only. The names page shows the people, places and groups in today's stories (from midnight ET), those in two or more stories or carried by 15 or more outlets, widest first, each with its first three stories and the rest folded. A title before a full name is dropped when counting ("Rep. Ilhan Omar" is Ilhan Omar; "General Motors" stays), and a name that is a common noun in WordNet and never written with a capital mid-sentence in our headlines ("president", "senate") is not a name.
• A story split by one hour's clustering is joined again. While both parts are on the front pages, the bigger keeps the story and the smaller starts a new one; when the two clusters joined again, the joined cluster went to whichever part held more of its headlines, often the young one, and the older story ended there (Christa Pike: eleven days ended at 7 AM on Oct 13, and a part split off at 1 AM, holding the same discharge headlines from Oct 10, carried on). Now a cluster holding a fair share of headlines from two stories both live lately is put to the model (the same event or not, as for a story's return); the same event, and the older story continues with the younger folded in. A new development (a resignation, a ruling) is still a story of its own in the saga.
• Carrying words. A filing can now carry the tellers' own words that carry its motif ("religion of peace", said sarcastically, under a crowd claim), written by the person in the checker (one phrase per ';'). Each motif gathers them into a phrase dictionary, counted by filing. They are evidence of how a motif is told in the tellers' own terms; they are kept, not yet read by any model. A phrase can belong to several motifs and means different things said sincerely or sarcastically, so it will be a finder and a tracker, never the filer.
• The claim organizer: a claim from posts can be split into the tellings it holds. Claims from posts sit in 3+ motifs 30.5% of the time (Oct 13; 10-15% for the other sources), and the Chimera found their paraphrase groups stitch tellings together. A second round now asks Gemma which of a claim's posts tell each of its motifs; when the motifs fall on different posts (a motif across the whole group counts as a layer, not a telling), it drafts one child claim per telling, each one side's assertion in its own voice, as a proposal for the person; no split is kept with its reason. Nothing changes until the person approves. An approved split takes the claim out of its motifs under a new curation-log action, split, which the confidence model doesn't learn from (an unfile would have read as "not this motif"); the claim stays in the index as the telling (tellings), and its children are filed as new claims with its source, story and a split_from link. From the first applied split, a claim from posts can leave the 3+ count without any filing judged wrong, and the filings count gains its children. Psyche (in the experiments' queue) runs it alongside the filing: on the Chimera's 160 claims against the person's filings, and forward on posts claims since Oct 6, its splits going to the proposals for the checker. Watched against these baselines: posts claims in 3+ motifs 30.5%, posts claims the person also filed elsewhere by hand 45.9%, sure share 44.7%.

2026-10-12

• Bloomberg from its own feeds. Its front page has answered every request with a captcha since Oct 11, 20:04, and we don't get around bot checks; its official section feeds (politics, markets, economics) still answer, so its headlines come from those, articles only. Their order is the feeds' (newest first), not the front page's, so anything weighted by a headline's place on the page treats Bloomberg's as a list from Oct 12 on.
• Posts are tied to news stories by the links they share, where the model finds none. A paraphrase group whose posts share an article in one of our stories is tied to that story when the model's own tie (embeddings, then Gemma) finds nothing; where both tie, they agree by story id 62 times in 65, and the model was the better of the two in the rest. On two weeks of reports this tied 27 more of 855 groups (389 in all). Guardian links now match whether or not they carry the edition (/us/world/... as we read it, /world/... as people share it): 60 of 71 shared Guardian links once missed now match, which also feeds a group's shared-side lean. Past filings backfilled: 258 of 783 posts filings now carry a story.
• The shortlist's held-out test is honest now, and its training too. Held out, a claim that was its motif's only one leaves that motif empty, a look nothing in real filing has; the weights learned to find exactly those, so the held-out top 12 read 86.7% (Thumbelina). Both trainings (the plain weights and the learned shortlist) now leave such rows out, and the test counts them as misses: about 73% at top 12 over all filings, about 85% on motifs of two or more claims. The dashboard's shortlist line steps down on this date for that reason, not because filing got worse.
• Plain news is watched and filed again. A claim judged plain news, in a motif the person marks as watched (Straight news: the checker's watch mark) or with no motif (a person's "tells no story" or the filing model's none), is no longer final: it is on a watch list (motifs/watch_list.json: 126 claims with no motif on Oct 12, and Straight news's 4 once marked; 3 claims with no motif have no words left anywhere to file) and filed again weekly against the index as it has grown, and at once when a narrative thread tells it again on a later day, in the 3 AM runs (at most 60 claims or 5 minutes a night). Filing again uses filing's own shortlist and judge; each pass is logged with the day the claim was reported and the motifs picked (motifs/refile_log.jsonl), so "a report at t0, told as motif X at t1" is on record. A real motif picked becomes an unchecked filing, marked as refiled, for the person to judge; a no puts the claim back with no motif. A watched motif picked means still plain; a new name is logged, never made. A show retelling the claim isn't detected yet (it needs claims compared by meaning); the weekly pass stands in.
• A filing keeps the story its claim was told about. A narrative's claim carries the news story the report tied it to (the model's tie, "same event" or "same issue"), a fact-check's the story whose card showed it; the Focus Group's and the shows' claims carry none (an episode covers several stories, and nothing cheap says which). So which motifs a story is told in, in time order, is a join (motif_index.story_motifs), with the plain report (Straight news) as its baseline: the person's decision that Straight news is a sink and a watch list. A story id is the site database's story, kept across hourly runs; a story folded into another is followed to it. Filings from before today were given theirs from the reports they came in: 212 of 768 narrative filings and 25 fact-check filings, over 81 stories.
• Claims are written short and plain, by the rules read from the person's corrections. Every prompt that writes a claim (posts, shows, the Focus Group, fact-checks) now carries, after its own instruction, the rules read from the person's corrections, a rule to write one short plain sentence (usually under 20 words, what is claimed, not who reports it), and 8 of the person's corrections of claims from the same kind of source as examples (fact-checks had the 5 latest of any source). This is the way the Brothers Grimm found closest to the person's wording on Oct 10 (0.784 alike against 0.758, and the one the blind judge preferred), never deployed until now. Claims written from today may be shorter and worded differently from earlier ones of the same kind; measured weekly (Cordelia).
• The checker's "might fit better" suggestions come from filing's own shortlist. For a weakly filed claim, the 12 motifs the confidence model weighs as alternatives were ranked by the plain six-fact weights, while filing has used the learned shortlist since Oct 8. They are now ranked by the learned shortlist's weights whenever filing uses them (the plain ones otherwise), the claim held out of its motifs as before. On 488 motifs the person added a claim to by hand, the 12 candidates held the person's motif 68.4% of the time against 62.9% (Castor and Pollux); 4.7 of the 12 change for a claim. Procrustes and Isis take their candidates from the same function.
• The taught reranker is kept only if it judges better than the untaught one. Each weekly teaching of Qwen3-Reranker-0.6B on the person's decisions now also scores the same decisions, on the same text, with the untaught model, and keeps the taught layers only if their held-out AUC is higher; otherwise the untaught reranker is used (and the confidence model retrains on its scores) until the next teaching. Both AUCs are kept with the teaching. Measured first (the Magic Mirror): on Oct 8's teaching's own decisions the untaught one scored 0.776 against the taught one's held-out 0.758, and on 838 decisions made since, 0.738 against 0.739: no sign the teaching helps, nor that it hurts.
• Motifs count their occasions. Each motif's kept claims are joined into occasions (alike by mxbai at 0.72, from one fact-check or show claim, or sharing a name found in at most 25 of the index's claims and not a common word); two or more occasions is attested, one is placed. Counted hourly after filing; the map draws placed motifs faint. On Oct 12: 288 of 616 motifs attested. An estimate, from Fable's reading of the index against the praxis.
• A motif can say what it isn't (not_this), read by the filing model, the judge and the element drafter; a related link can carry the person's reason (related_notes). Neither changes anything until the person writes one.
• Seeds can be filed into. A seed (a motif the person makes with a name and note and no claims) was never offered to the filing model: every shortlist was built from motifs with claims, so the praxis's reason for seeding didn't hold (found in Fable's review). Now up to 2 of the person's seeds most like a new claim by name and note (at least 0.45 alike) join its shortlist, outside the learned weights, which lift motifs with no claims unearned (Jack). A filing into a seed is never passed as sure. Measured weekly (Persephone).
• What each new claim is shown is logged (motifs/shortlist_log.jsonl): the motifs shown, the seeds among them, the pick, and the next motif number then, scored weekly against the person's final word (Prometheus). The shortlist's held-out recall (86.5% at 12) is measured on filings it had already found; forward, on the Oct 8 weighting, it was about 60%, mostly motifs the person made after the claim arrived.
• Stories get a second pass: cut where their words fall apart. The person's idea (Oct 11), tested as Atropos and Rumpelstiltskin: stories are still formed by meaning (mxbai embeddings, 0.80), then each story of 12+ outlets is compared by its headlines' words (TF-IDF fitted on the story's own processed headlines) and split at the cleanest normalized cut if it scores under 0.6, each side at least 6 outlets, up to three cuts deep; but only when Gemma 4 26B, asked two ways, says the two sides report two events. On a week of stories, cuts under 0.6 confirmed that way were two events 18 of 18 times; on the last day's stories the score alone would have cut 9, three of them one event, so the confirmation stays. A story split this way keeps its id on the side holding most of its headlines.

2026-10-11

• The motif index is all rights reserved. On the site, in its Markdown and SKOS exports, and in the footer of every page of the daily archive (the build gives kept pages today's footer; the 346 stored copies were rewritten to match): reusing it now needs our permission. It was CC BY 4.0 from October 8 to 10 and CC BY-NC-SA 4.0 from October 10 to 11; copies taken then keep those licenses. The motif casebook (the person's rulings with their reasons) left the public repository with it, kept beside the index in the data folder; its earlier versions remain in the repository's history. The other docs stay CC BY 4.0 and the code AGPL. The Wayback Machine holds no copy of the site from after June 2024.
• Story titles keep up with the story. A story's title was written again only when its outlet count doubled, but stories are kept for days and their news moves on: "Tennessee inmate survives botched execution attempt", written Oct 2, still headed nine days of Christa Pike coverage that had moved on to her leaving hospital. Now a live story's title is written again once it's 12 hours old and new headlines have come in, from its 30 newest headlines, stating the latest development. Titles already written are refreshed over the next runs, 25 an hour at most. Until the afternoon of Oct 11 this didn't happen: the rule read each headline's time after the front page had made it display text, and as text "Oct 9" sorts after "Oct 11", so no story looked to have news since its title. The Trump–Russia diesel deal's part kept its Oct 9 title over two days of Trump calling for Zelensky to be replaced.
• Saga charts drop to nothing between sightings. A saga's spread chart drew each part from snapshot to snapshot, so where no part was on a front page for days it sloped across the gap: a part gone on Oct 4 was drawn until Oct 9. Now each part falls to nothing an hour after it was last seen.

2026-10-10

• Licensing narrowed. The motif index (on the site and in its Markdown and SKOS exports) is now CC BY-NC-SA 4.0 with credit to bignews.day, and commercial use needs permission; the site's own writing and readings are all rights reserved. They had been CC BY 4.0 since Oct 8. Third-party data keeps its own licenses (AllSides CC BY-NC, Wikipedia CC BY-SA, VoteHub CC BY). The code is open source again, under the GNU AGPL v3 or later (it was MIT until Oct 8, then proprietary), and the repository is public; the docs (methods, praxis, casebook, experiments) are CC BY 4.0.
• New motifs are named the way the person names them. The filer and the proposer used to name a new motif in "three to seven words, terse like a folklorist's label", and the person renamed or rejected every such name. They now follow the praxis's naming rules (docs/motif-praxis.md): an idiom, a catchphrase or the tellers' own words, or a known story whose plot is the shape; no abstract or topic labels; one to five words. Tested on 101 motifs the person had renamed (P2): abstract labels went from 51% of names to 10%, five words to three; the person judged the names far better. Motifs already named keep their names.
• One model fewer. Whether a newly named motif is one the index already has is now asked of Gemma 4 26B, not Qwen3 30B, which did only this: on the person's 33 merges and 44 "not the same" verdicts Gemma found 20 merges to the 30B's 18 and joined the same 4 not-same pairs, faster (M1). The two no longer trade places on the card mid-filing.
• Gemma 26B writes the claims from posts and fact-checks. The nightly report's narratives (each group's claim and labels, and the check of which posts are about it) and the fact-check labels were written by Qwen3 8B; Gemma 4 26B now. On the person's corrected claims Gemma's wordings came closer to theirs (W1: posts 0.799 against 0.775, fact-checks 0.791 against 0.746 alike), and Qwen3 8B looped or invented details. Labels already made stay; the headline scorer stays on Qwen3 8B, so the news, loaded and emotion series don't change.

2026-10-09

• The genre drafter is told what each genre is. It used to see only five of the person's motifs in each genre; now it also gets each genre's line from docs/motif-praxis.md (or the person's own 📝 description of it) and the tests for telling close genres apart. On the person's 453 genre calls it was right 53% of the time before and 66% after (Theories 38% to 72%, Plots 57% to 73%: G1 in docs/filing-experiments.md). Proposed genres are drafts, as before.
• The judge checks each motif's must-have element. Each motif now has one sentence saying what a telling must contain to be an instance of it (drafted by Gemma 26B from its name, genre and note, again when they change), and Nimble 9B is asked whether the claim, as told, contains it. On the person's 96 rulings of Oct 9 (the casebook's hard cases) the judge's AUC rose from 0.54 to 0.79; on all 897 decisions from 0.801 to 0.811 (J2 in docs/filing-experiments.md). The confidence model retrains on the new answers; fits from the old ones stand until then.
• The check list goes least sure first. The ✅ check tab used to list the filings the confidence model was surest the person keeps first; now those it can least tell (a fit nearest even), and 🎯 today gives 13 of its 20 a day to such torn filings (10 before, in a wider band: fits from 20% to 80%). A simulation on the person's own decisions (E6 in docs/filing-experiments.md) found the model learns as much from 150 checks in this order as from ~550 at random, and more slowly still in the old order. The sure ones still pass in one step from their own fold.
• A new judge in the confidence model. Its yes-or-no on each filing now comes from Nimble 9B, a decision model (a calibrated probability in one pass), instead of Gemma 4 26B: on the person's 817 decided filings it judged as well (AUC 0.824 alone against 0.818; 0.895 in the model against 0.894) in a fifth of the time. (It doesn't quite fit the card: 2 GB of its 9.5 run on the CPU; the 4-bit build that fits judged worse, 0.797 against 0.818.) The model retrains on Nimble's answers; fits from the old one stand until then.
• Two more embedders in the filing shortlist. Qwen3-Embedding 8B and Snowflake Arctic Embed 2 each compare a claim with every motif's name and note and with its nearest claim (E11: of twelve embedders tried, this pair helped most). The learned shortlist, retrained with them on the person's 470 confirmed claims, puts their motif in the top 12 for 85.7% held out (83.9% before; the plain weights 80%). The confidence model weighs them too.
• Nimble judges the link proposals too. The 💡 related and rests-on proposals' chance of approval (best fit) now weighs Nimble's yes or no instead of Gemma's 0-10 score: on the person's 301 decided proposals it ranked approvals above rejections better (AUC 0.772 against 0.755, in 20 of 20 test splits). The judge's reason no longer shows.
• Claims without their teller. The Focus Group and show readers now write the claim itself ("Defunding the police isn't a federal issue"), not "The voter believes..." (it had named a guest's analysis a voter's). Who said it stays with the quote.

2026-10-08

• "Kind of" is now "rests on". The one link that ranks motifs, until now "X is a kind of Y", now reads "X rests on Y": X only makes sense given Y, as a kind of it, or as a case, argument or figure by which it's told ("Magic money tree", an argument, rests on "Politicians' empty promises", a theory). Every kind-of link still holds as one; a survey of the 83 found about 18 that had only ever been right this way (an archetype or argument resting on a theory). The rule of Oct 7 that the link stays within one genre is gone, and the links it took away are proposed back. The proposer, its judge, the workbench, both maps and the export say "rests on"; the data is unchanged.
• How sure a filing is. Each claim the model files under a motif now gets a fit: the chance the person keeps it, from a small model trained on their own decisions (723 so far: 572 filings kept, 151 taken out). It weighs how alike the claim and motif are (the filing shortlist's measures, the claim held out), how the motif ranks among all motifs for the claim, how many motifs the claim is in, where it came from, and whether the person has curated the motif. Tested with each claim held out: AUC 0.83 (the shortlist's score alone: 0.74); at a fit of 85% or more it passes 64% of the filings the person kept, and 95% of what it passes they kept. The check list goes likeliest first, and the filings over that line can be passed in one step, except a filing in a motif made for that claim alone, or from a source with fewer than 30 of the person's decisions (the shows, for now). Not used on the public site: what's public is still only what the person has checked. Later the same morning, on the person's word: only filings in their own motifs (marked done) get a fit or are learned from; a motif the model made and nobody has shaped measures nothing (a claim had scored 100% in the motif just named from it). Retrained so: AUC 0.82, sure from 95%, where 96% of passes were kept. By midday the fit weighs many more kinds of evidence (app/analysis/motif_signals.py): the motif's note alone, its keywords, the claim rewritten as the bare shape of its story, the motifs of the most similar filed claims, a second embedding model, the person's groups, a cross-encoder reranker (Qwen3-Reranker-0.6B) and the filing model's own yes or no with its probability. Tested on 749 of the person's decisions held out by claim: AUC 0.885; at a fit of 85% or more it passes 76% of the filings they kept, 95% of what it passes kept (29% at that precision the morning before). Boosted trees did no better than the plain weighting, and both level off by half the decisions: more checking alone won't move it much. Fits are worked out in the hourly run and kept on the filings. The person's confirmed filings get one too: the lowest of the model's own, a second look for a hurried yes. In the afternoon, from the first experiments (docs/filing-experiments.md), one more: the claim against each motif's note and nearest claim with the day's news taken out. The main directions the latest 6,000 headlines vary along are mostly topic (a country, a storm, a trial); projected away from both sides, what's left leans to how a story is told. Held out, it brought 1.3 points more of the person's motifs into the top 12, with nothing of their filings in it to leak. The directions are worked out once a day.
• New motifs come with the model's note, genre and groups. A motif the model makes for a claim no motif fits used to arrive bare: no note until it held three claims, no genre or group until the person gave one. Now the hourly run drafts all three straight away: the note from its name (the claims only to see how it's told), the genre judged against examples of the person's own motifs in each genre, and up to two groups, each the model's yes among the three groups whose motifs are nearest. The genre and groups are marked as the model's guesses (🤖 in the checker) until the person marks the motif done or sets them by hand; till then nothing learns from them (genre examples, group centers, the group signal of the filing model and the public pages read only the person's), and a merge leaves them behind.
• The public motif map, readable on a phone. On a phone the map came after a screen and a half of explanation and filters, and then drew all 350 names at once, an unreadable heap. Now the explainer and the key fold away on small screens (open on wide ones), so the map comes first; names show only where they fit at the current zoom, the most telling first (the most claims, what others rest on), the picked motif and its neighbours always, more as you zoom in, the groups' names placed first; related lines stay faint until their motif is picked; and lines between motifs that only share a claim are off unless asked for (told together isn't related).
• 🎯 Today's most useful decisions. The person can't confirm motifs for hours a day, and needn't: the confidence model learned nearly as well from half their decisions as from all of them. What teaches it most per minute is a few decisions of the right kind, so the workbench's 🎯 today tab gathers about twenty: the filings it's torn on (a fit nearest 50%, one per claim), the confirmed filings it now doubts that have a better motif to offer, the motifs it made that nobody has shaped yet (keep as drafted, or open and change), and the likeliest proposals for new motifs and groups. The rest still gets filed and scored; the public site shows only what they've checked.
• A dashboard of everything (the checker's 📈 Dashboard, app/dashboard.py): whether the machine keeps up (each hourly run's minutes against the 45-minute limit, feeds or outlets not coming in, the card's heat and load, models loaded into memory each hour), what came in each day (headlines, outlets, side-feed items by kind, transcripts and hours of audio, fact-checks and whether their pieces were read), what was made of it (claims filed, what waits for the person, proposals, episodes by kind), the motif index's shape and growth, the models' own held-out tests and the latest experiments, and the person's curation by day and kind. Read live, each part again once it's a minute old, in the background.
• The hourly run keeps to its time. The 3 PM run was stopped by its 45-minute limit before it filed any claims, and the 5 PM run took 44 minutes: the day's additions and the backlog of podcast episodes newly let in added up. Each step after the site is published now gets what's left of 40 minutes at most, ten of them kept back for filing claims and their fits; a step with no time left waits for the next run and says so in the log.
• Episodes labeled news or evergreen. Some shows spend an episode on history or a life story. Each new episode of the podcasts, call-ins and video channels is now read from its title and blurb as on the week's news, evergreen, or a mix (ten at a time, once each). A label, not a filter: evergreen episodes are still transcribed and their claims still read and filed (motifs live in history talk too), but the news goes first in the transcription queue, and a claim counts as retold this week only when shows on the news tell it.
• The shortlist weighs every signal. The motifs the filing model chooses among for a new claim were picked by six likenesses from one embedding model. They're now picked by every signal of the filing experiments (keywords, the note alone, the claim's bare shape and its layers, its nearest filed claims, a second embedding model, the person's groups, the news taken out), weighed on the person's confirmed claims and retrained daily. Held out, 83.9% of the person's motifs are in the top 12, against 79.8% before: about one more of their motifs in every twenty reaches the filing model. It's used only while its daily test beats the plain weighing. Each new claim costs two more questions of the filing model (its bare shape, its layers), asked once and kept.
• The reranker taught the person's taste, in production. Experiment E9 (docs/filing-experiments.md) found the cross-encoder, its top six layers trained on the person's decisions, judged held-out claims better (AUC 0.765 to 0.796). Weekly at night it is taught again on the CPU, in a process of its own: three models that each leave a third of the claims out give every decision an honest score for training the confidence model (a reranker that learned those very decisions would look better on them than it is), then one model learns from them all and scores new filings. The confidence model retrains whenever a newly taught reranker appears; each reranker's scores are kept apart.
• Embeddings on the CPU. Ollama's log showed the card spending about nine minutes an hour loading models: each embedding model (mxbai, nomic) loaded beside the filing model pushed it out of the card's memory, and it was loaded again a moment later (242 times in six hours). Embeddings now run on the CPU (24 cores: 27 texts a second for mxbai, 61 for nomic), so the filing model stays on the card.
• Shows' own transcripts before Whisper. Seven of the 66 podcast and call-in feeds publish a transcript of each episode in the feed itself (Podcasting 2.0's podcast:transcript: Breaking Points, Verdict with Ted Cruz, Clay Travis & Buck Sexton, the Jesse Kelly and Michael Berry shows, Bloomberg News Now, Jubilee's Surrounded; mostly iHeart's, timed WebVTT with speakers). Those episodes are now taken from the show: their cues joined into each speaker's sentences, the speaker kept, sponsor reads trimmed as before. No GPU, so they come in even when the model holds it, and Whisper's ten minutes an hour go to the shows that publish nothing.
• The proposer suggests new motifs. The person expects to need many more motifs and wants the model to propose them. Claims that fit nowhere (no motif, but those the person said tell no story; or filed weakly with nothing better on offer) are each read as the bare shape of its story; where shapes cluster, two or more claims telling one story, the model sees them with the five nearest motifs and, reading as a folklorist would, either names a new motif (a terse label, a scope note in the tellers' terms, its genre, the claims that are of it) or says a motif covers them already. One 💡 proposal per new motif; approving makes it with its note and genre and files its claims, in one undoable step. A single claim isn't a story told again, so it isn't proposed (the nightly filing still names those).
• The proposer suggests new groups. On the person's ask ("can it run cluster analysis?"), the nightly proposer clusters their done motifs by what each covers (name and scope note: with their claims, motifs clustered by the week's topics, a plane scare with a Reddit sleuthing motif), keeps the tight clusters of 3 to 12 that no group already mostly holds, and asks the model whether each is a family of stories the way their groups are; if so it names it in their groups' style and leaves out what doesn't belong. One 💡 proposal per family; approving it makes the group with its motifs in one undoable step. A family the model turns down is asked again only once its motifs change.
• Fact-check claims read from the piece, not the headline. The person often opened a fact-check to see what was actually claimed, so their filing rested on more than the one line the models saw. Each fact-check's own text is now kept (app/analysis/factcheck_text.py): from the feed where it carries the whole piece (PolitiFact, FactCheck.org, Lead Stories, NewsGuard), otherwise read from the page, 20 seconds apart per site, at most 24 pages a run, robots.txt obeyed (read under our own name: the sites block named AI crawlers, not readers like us). The labeler reads the piece's opening and gives the claim as its tellers state it, who says it and where, plus two or three sentences of what the piece shows; the person's own corrections of claims (46 so far) are shown to it as examples. A claim already filed keeps its wording (the reading from the piece is kept beside it), so nothing filed changes under the person. The checker's 'where it came from' shows the piece's reading and its opening.
• Five podcasts' episodes were being dropped since Oct 4; six video channels had moved. A health check of the side feeds found The Daily, the Brian Lehrer Show, the Megyn Kelly Show, WSJ What's News and NBC's Top Story with nothing new in four days: their feeds had begun giving every episode the show's own page as its link, so each new episode looked already stored. An episode whose link another shares is now told apart by its guid. The YouTube feeds of the Young Turks, Timcast IRL, Benny Johnson, Matt Walsh, Candace Owens and Destiny pointed at channels gone quiet (some for a year); each now follows the channel its handle names. Dan Bongino's and Flagrant's channels have posted nothing since 2025 and early 2024: dropped, their podcasts carry them. The 145 outlets' front pages and every hourly run of the last day were fine. So it doesn't happen silently again, each hourly run now checks that every outlet was seen on its front page in the last six hours and that every feed that posts regularly has something new within four times its usual gap (three days at least), and that no feed is stuck with a handful of items; what isn't is a warning in the log and a 🩺 box on the checker's 📊 tab (app/health.py).
• Weak fits come with where else they might go. The person's time is better spent correcting weak filings than hunting for motifs: for each claim with a filing the confidence model gives under 50% (one nobody has checked, or one they confirmed that's up for a second look), the hourly run scores the twelve of their motifs the shortlist ranks highest that the claim isn't in, and keeps the best three from 30% up. The checker shows them under the claim ("might fit better") and under each second-look row: ＋ files it there too, ✕ says it isn't that motif (never suggested again).
• Top stories as full cards again. The front page's nine top stories (ranked as before) are the stories page's own cards again, every outlet's headline on each, the lighter cards of Oct 6 having read as boring. Each keeps a bar of who carries it by lean, now in five bands (left, leans left, center, leans right, right, plus unrated) instead of three, so a story carried by the far left reads apart from one carried by the center-left.
• Folklore and Rumors are one page. Folklore and rumors (folklore.html) is now one catalog of what's told, from three places, each marked on its card and filterable: retold online (Bluesky and Mastodon, as before), told on the shows (a claim told on two or more of the shows we transcribe, or by two or more callers: which shows, how many callers, the tellings by the shows' lean; never the quotes or who called), and fact-checked (the old Rumors page; a check already linked from a narrative's card isn't repeated). The model's built-in genre (conspiracy theory, contemporary legend…) is gone from the cards: a card's genres are now those a person gave the motifs it's filed under. rumors.html sends visitors to the fact-checks.

2026-10-07

• Claims from the shows we transcribe. What people post (Bluesky and Mastodon) leans left, so the folklore page heard little of what's told on the right. Every podcast, call-in show and video show we transcribe (about 400 episodes a week: 157 right, 132 center, 105 left) is now read for the claims told on it, by the same model and checks as the focus groups: each claim's quote must be in the transcript and support it, and each says who told it (host, guest, caller, a played clip). A claim reaches the motif index only once it's told on two shows or by two callers: one host's take isn't folklore; told again elsewhere, it is. Quotes and callers stay off the site.
• A better shortlist for filing claims. Filing a claim, the model chooses among the motifs most like it; when the motif a person would have chosen wasn't among them, they added it by hand (over 200 times). The shortlist was the 8 closest by one measure. Now several measures of likeness (the motif's description, its nearest claim, the center of its claims, its size, its neighbours in the kind-of tree) are weighed by a small model trained each day on the person's own confirmed filings, and a shortlisted motif brings its broader motif and its kinds along (about 12 shown). Re-filing all 372 confirmed claims, each kept out of what it was tested against: the model found 375 of the 763 motifs the person had them in (322 before), 47 of the 206 they had added by hand (33), with no more picks the person had taken out (16 both ways).
• Proposals sorted by how likely they're approved. About half the model's kind-of and related proposals were rejected. A second read by the same model, asked how likely the person was to accept each one, was little help alone: it gave most approved and rejected proposals the same 2 or 3 (and a first prompt that called the person strict rejected nearly everything). Weighed together with plain facts about the two motifs by a small model trained on the person's 144 decisions, it does better: tested on decisions it hadn't seen, the least likely tenth of the approved marked a line below which fell 29 of 65 rejected. The checklist is sorted by that chance, and those below the line are folded at the bottom, ticked, for the person to skim and reject at once; nothing is hidden.
• The public motif map looks like the working one. The site's map was drawn once (Oct 6) and missed a day of layout work on the checker's map. Both now draw through one script (static/motif-map-layout.js): each group placed on a ring of its own, links out of a group pulling weakly, a group drawn as a blob around its members with non-members kept out, a dot's fill its genre and its outline its place in the kinds, and one color per group and genre on every page. The site shows genres only once a checked motif has one. Editing stays on the checker.
• A motif a person made and wrote a note for is done. Marking it done was a second step after writing the note, and 92 motifs the person had made (or renamed) and described sat unmarked, off the site. Now saving a note of theirs (or keeping the model's draft) on such a motif marks it done, with the claims it holds then; one they unmark stays unmarked. The 92 were marked done: 332 motifs on the site, 741 claims (each claim still read for named accusations first). A model's motif with no person's name on it still waits for a person.
• A merge keeps everything. Merging one motif into another moved its claims, related links and kinds, but lost its own broader motifs, its groups, its genre, its note and the person's "not the same" verdicts, and its claims seen when it was marked done showed as new again. Now all of it comes along (a broader motif only where it can't make the motif a kind of itself; the note and genre only where the motif kept has none).
• A claim a person checked counts as seen. A motif reaches the site once a person marks it done, with the claims it held then; claims filed since waited, marked new, until it was marked done again, even ones the person had checked ✓ one by one in the meantime. A claim checked ✓ or filed by hand now counts as seen: no longer new, and shown with its motif. The model's own filings since still wait. 554 claims show on the site instead of 431; motifs marked new went from 99 to 34. (The checker's organizer and singles pages were dropped too; the workbench does their work.)
• A correction proposer for the motif index. gemma4:26b reads over the motifs and proposes small fixes, each one change a person approves or rejects in the workbench's 💡 proposals checklist: a typo in a name or scope note (only a small edit counts, not a rewrite), one motif a kind of another, two related, two the same motif (merge), a motif into one of the person's groups. Candidates are found cheaply first (motifs alike in meaning, motifs sharing a claim, motifs near a group's members) and only those put to the model. Approving runs the person's own action (logged, undoable); a decision is kept, so a rejected proposal never comes back, and one overtaken by hand drops out. It runs once a day from the 5 AM hourly run on (never in the run that makes the narrative report, which nears its 45-minute limit), at most eight minutes, going on in the next run if cut short. The model's own "no" and its typo reading are kept with a mark of the motifs as they were (name, note, about how many claims), so a motif that changes is looked at again. It also reviews the links already made (related, kind of): one the model calls unrelated comes up as a proposal to take it away, and one it keeps isn't read again until either motif changes.
• No more villain. The folklore labeller named who each narrative casts as the villain (Propp's part), and its picks were too often wrong to show. It no longer asks, and the Folklore page, the checker and the reports leave the part out, old reports included; the victim and the hero stay. Fact-checks' labels still carry a villain field, unused and never shown (rewording that prompt would relabel every fact-check and the claims filed from them).
• A story back on the front pages stays one story. A news cluster continues the saved story that holds most of its headlines; one that left the front pages and came back hours later under fresh headlines shared none, and became a new story. The physics Nobel was on one noon run on Oct 6, gone, and back at 5 PM as a second story: a saga of two near-identical parts. Now a cluster that continues no story is compared with the stories that left the front pages in the last 48 hours; one sharing two of its frequent words is put to gemma4:26b: the same news event told again, or a different one (the next game of a series, a new ruling, an arrest after a crime)? The same event continues the old story. On the 42 pairs the rule would have asked about in the week before, it joined only the Nobel pair (the first wording refused it over one stray headline about the medicine prize; it now goes by most headlines). The two Nobel stories were merged by hand; a merged story's page redirects to the one kept, and TV, radio and narrative links to it follow.
• Narratives linked to news stories by the bigger model. Which front-page story a retold narrative is about (its card on Folklore, and where it shows on story and saga pages) was asked of qwen3:8b, which put "Jacob Geller is a rapist…" on the Cornell rape case and the threats against the Cornell accuser on a different part of the saga. Asked again of gemma4:26b, the 346 links it had made differed on 42; read by hand, Gemma was right on about 25 of them and the small model on about 11, and Gemma joined fewer wrongly (5 to 7) and missed far fewer (3 to about 20). From tonight's report on; earlier reports keep their links.

2026-10-06

• Named people accused of crimes. An outside reader found the motif index, a story page and a saga page showing "Jacob Geller is a rapist who has avoided accountability…", retold by eight people online: an unverified accusation of a serious crime against someone named who holds no public office. Now every claim the site shows from people online or from voters (motifs, folklore, story and saga pages) is first read by gemma4:26b: does it say a person it names committed a crime or abuse? Officials, candidates and household names stay named, and so does anyone our outlets' front-page headlines have named in three or more articles (the model didn't know Howard Lutnick is in the cabinet). Anyone else is replaced by who they are ("A YouTuber is a rapist…"), and their posts, cast and earlier tellings aren't shown with it; a claim whose name can't be swapped out cleanly, or that hasn't been read yet, isn't shown. Fact-checkers' claims are shown as written. Two hand-kept lists override the model: names never shown (never_named.json) and names always shown (always_named.json). Story and saga pages also show a claim found twice in one night's report, in nearly the same words, once.
• A lighter front page. The front page had become a dashboard, a corpus and a reader at once (an outside reader counted 36 links on one story card). The story cards, every outlet's headline on each, moved to their own page, Every story (stories.html). The front page keeps the cloud and puts the nine hottest stories (ranked as before) in short cards, each with three things: how its coverage spreads across the lean scale, one framing contrast (a phrase only one side uses, skipping phrases made only of the story's title words, or the most-quoted phrase), and one sign from beyond the front pages (TV, radio, people retelling it, a fact-check). Sagas follow, then the trend boxes. Each saga has a page of its own (saga-.html) following it across its parts.
• Sagas: no more joining by persistence. An outside reader found three sagas holding a part from another storyline. All three were the old judge's yeses, and one shows the flaw: the Supreme Court climate case and an overview of the court's new term were judged separate 25 times, then joined on the 26th, because each hour the stories had new headlines, so a new sample to ask about, and one yes was kept for good. Now a verdict is kept for the pair of stories (saga_pairs.json) and never asked again; the judge sees each story's most central headlines, not ones spread evenly across it (stray headlines from a neighbouring storyline had stood in for a story); and a newcomer is judged against the saga's part nearest to it. The judge's wording now says that different angles on one event, its consequences and reactions are one story, and that a different focus is no reason for "separate": on the 46 labeled pairs it found 7 of the 11 true running stories (5 before) and still joined none wrongly. The three wrong parts were detached (the Nebraska rally from the diesel order, the court's new term from the climate case, two Iranians charged over a plot against Jews from the RAF Fairford bombers story); two sagas ended. A check of every saga against this judge would also have split four borderline ones (Alito's two stories, the Saudi-Houthi fighting, Paxton's campaign, measles in two states) and one clear one (two headlines on one Ohio rally); those were left as they are, for a person to judge.
• Each narrative keeps all its posts. The nightly report now keeps the private keys of every post in each narrative, not only ten sample texts, so a person checking a claim sees everything it was found in. The texts stay in the post store under its 30-day limit (a deleted post is gone from it too); the reports of Oct 4-6 had their keys rebuilt by grouping their posts again (361 of 367 narratives matched).
• Jubilee's Surrounded, transcribed. Jubilee's weekly Surrounded (one guest against 20-25 people who disagree with them) joins The Focus Group and the call-ins as ordinary people speaking for themselves: archived from its podcast feed and transcribed on our GPU in the same tier, for research. Its Middle Ground has no audio feed, and its YouTube feed is now mostly short clips of its commentary show, so neither is followed.
• The nav follows a story downstream. Two links (every story, sagas) and four menus in the order a story travels: 📰 front pages (every headline, outlets, headline changes, emotions, who's in the news, the Supreme Court, the archive), 📡 on air (TV, radio, shows and newsletters), 🧶 online (folklore, rumors, the motif index) and 🛠️ how it works (now with the methods log, which wasn't in the nav). The old grab-bag "trackers" menu is gone.
• Scope-note drafts by Gemma 4 26B. A motif's draft note is now written by gemma4:26b, not qwen3:30b-a3b, after drafting notes for the 49 motifs whose notes the person wrote and comparing them with theirs: Gemma's were closer in meaning (0.73 against 0.71) and told the kind of story where the 30B restated a single claim, and the 30B gave up on 10. A draft that opens with the motif's own name ("Breaking ranks occurs when...", half of Gemma's) is asked again.
• Narrative threads judged by Gemma 4 26B. Whether a narrative told today is one told on an earlier day is now asked of gemma4:26b, not qwen3:30b-a3b, after a test on the 32 pairs the live code had asked about (labeled by Claude): Gemma found 18 of 21 retellings in both runs and wrongly joined 1 of 9; the 30B found 15 and 17, missing plain repeats (the same Nobel prize, Sánchez's snap election). The threads were rebuilt from the three daily reports with it: 17 narratives told on two or more days (was 16).
• New motif names matched more often to the motifs already there. When the filing judge names a new motif, its name is first checked against existing ones, but only those at least 0.7 alike by meaning were offered to the model. Tested against the person's merges: the motif a duplicate was later merged into reached the model for only 7 of 31. At 0.6 it reaches it for 20, and the model found 14 (against 7), joining 2 of 26 pairs the person had said were not the same (against 1 of 11). The floor is now 0.6; the model stays (gemma4:26b, gemma4:12b and qwen3:8b did no better).
• Widely seen Bluesky posts, embedded. Until now no post was ever shown, only the model's summary of what a narrative's posts share. Now each narrative can show up to two of its own Bluesky posts that were already widely seen: 25+ likes, reposts, quotes and replies together, or 3+ from an account with 5,000+ followers (Bluesky gives no view counts). They're embedded with Bluesky's own embed, never copied, so a post its author deletes disappears from our pages, and shown posts are looked up again every six hours. Never shown: anyone who has asked Bluesky not to show their posts to logged-out visitors (one in five of the posts in a test sample), accounts labeled bots, posts or accounts with a moderation label, and anything from Mastodon. To do this the sampler now keeps each Bluesky post's address (at://...) with its text, under the same 30-day limit; the nightly report keeps the 20 posts closest to each narrative's claim; each hourly build looks a few batches up on Bluesky's public API (at most 8 calls, 10 s apart). The thresholds are a first guess, to check against the first nights' counts.
• Story pages: a week of retellings, later fact-checks, and voters. A story page showed retellings only from the latest nightly report, tied by the story's label (which gets reworded), and fact-checks only if they were found while the story was on a front-page card: 3 of 160 pages had any. Now each day's last report of the week is read; its story links are tied by id (saved from Oct 6 on), or by the same label, or by the label of the story around then that shares at least half its words; a narrative told on several days shows once, on its latest day (51 stories have retellings, up from 37). Fact-checks are matched for every story of the week at build time, in their own cache, so one that came after the story left the front page still appears. And a new card, "the same shape, told by voters", lists what Focus Group voters told that's filed under the same verified motifs as the story's retellings, in our words, with the episode: the first place motifs connect a story to other material.
• The motif index can be exported. The workbench's 📦 export (and scripts/export_motifs.py) writes the index as a readable catalog (Markdown), a spreadsheet (a row per claim), JSON, or SKOS in Turtle for the planned mapping to prior catalogs (docs/motif-catalogs.md). By default only verified motifs, with the claims they held when verified and only a person's notes, the same as the site; "everything" adds unverified motifs and marks draft notes.
• One-sided wording: the story's own words don't count. "Only the right says 'bakker dead'" was on the front page while the left wrote "Bakker dies": a phrase one side used, but not a framing. A phrase now counts only if at least one of its words (stemmed, with dead/death/dies folded together) never appears in the other side's headlines on the story. On three days of stories this dropped 39 of 101 phrases, mostly vocabulary ("federal judge", "former rep", "terror plot" against "terrorism plot"); "killed" and "died" are kept apart, since that choice is framing.
• Word cloud: no made-up words. The cloud folds plurals to their singular, and the folding turned "dies" into "dy" (WordNet's -ies rule), so on Oct 6 a word "dy" gathered every death story of the day (Jim Bakker, Dennis Hastert, the Kenya Ebola case). Now the cloud picks nouns before folding, never folds to a word of two letters or fewer, leaves words in -ics alone, and drops "dies". On the day's 10,477 headlines this also took out verbs that had passed as nouns ("calls", "plans", "vows", "warns") and two more mangled words ("leaves" as "leaf", "discusses" as "discus"); the top 25 words didn't change. Also: the top word is sized against the runner-up, not the smallest word, so it no longer dwarfs the cloud, and on a phone it's no longer dropped for lack of room.
• One claim, told several ways. A person can now say two claims in the motif index are the same claim: the one folds into the other in every motif either was in, its words kept as a variant of the kept claim, and a later telling in exactly those words is filed as the kept claim, not as a new one. A motif's claim count then counts claims, not wordings.
• Outlet chips: the face is the headline's feeling. Each outlet's chip on a story card showed 😠 for any headline whose mood was below zero, a leftover from the 2024 site, so nearly every chip looked angry, even on obituaries whose headlines were mostly sad. It now shows the headline's own strongest feeling (😱 fear, 🤬 anger, 😭 sadness…; none when neutral). Its mood is compared with the other outlets' headlines on the story (it had included itself), and when they agree the tooltip says "same mood as the other outlets" instead of "+0.00". Most outlets do agree: mood is a five-step scale, and the same event usually gets the same step. Found by an outside reader.
• Only verified motifs on the site. The motif index grows with the model's filings, but the site now shows a motif only once a person has looked over its claims and marked it done, and shows only the claims they saw then; claims filed since wait until they mark it done again. Unchecked motifs stay in the index and in the curation tools. On Oct 6, 234 of 261 motifs had been marked done.
• Sagas: the judge is Gemma 4 26B. On 46 pairs of stories labeled by hand (11 of them one running story), it agreed on 40, joined none wrongly and took 2 seconds a pair; Qwen3 30B-A3B agreed on 41 but joined one pair wrongly and took 33 seconds. Gemma finds a few fewer true parts (5 of the 11 against 7), so a saga may gain a part an hour or two later than before, but a wrong join, the kind a person has had to veto, is rarer. Verdicts already cached are kept.
• Focus Group: a claim must be what its quote says. Since Oct 5 a voter claim's quote had to be in the transcript, but nothing checked that the quote said the claim: "Politicians cheat on their spouses but still get elected" was kept with a quote about politicians telling you what you want to hear, and filed in the motif index. Now Gemma 4 26B is asked whether the quote is evidence for the claim (the voter says it, or a central part of it, in any words), and a claim it isn't evidence for is dropped (kept on record with the reason). A first wording ("does it support the claim as worded?") passed only 6 of 16 sound claims; the one in use passed 14 of 16, turning down two that added what the voter didn't say, and still caught the bad ones. Claims read before were checked the same way, in the wording a person had corrected them to: 26 of 231 were dropped (one said a voter praised a governor the voter had said was "not my favorite") and taken out of the motif index.
• Transcripts: names Whisper mishears. 35 of 46 mentions of AIPAC in our podcast and Focus Group transcripts had come out as "APEC" or "APAC", and Hegseth and Whatley were misspelled too. Whisper now gets a list of names and terms to expect (fixed ones such as AIPAC, Hegseth, MSNOW, and the 50 most-named people and groups in recent stories), and known mishearings are corrected after transcription, sentence by sentence and only where it's safe: APEC is left alone in a sentence about the Asia-Pacific summit, and "a PAC" is never touched. The same corrections were applied to the transcripts already made (14 changed) and to the Focus Group claims drawn from them (10), five of which were already in the motif index and were corrected there through its record of corrected wording.
• Focus Group and radio running order on Gemma 4 26B. Each tested on its own work: on 10 newscasts, Gemma's stories passed the first-words check 96% of the time (Qwen3 30B-A3B 81%) with nearly the same front-page picks; on 12 Focus Group transcript parts every model's quotes were found in the transcript, but Gemma's claims were the voters' own, each with the voter's side, where Qwen3 30B-A3B took a host's analysis for a voter's.
• Story pages: following a story downstream. Every saved story seen in the last week has a page, and its card on the front page links to it. The page follows the story as it spread and changed: the first front page to carry it (a headline first seen before the story began is left out: it joined later), its peak, its first minute on TV and on the radio, the day people were found retelling it; then the front pages over time with the aggregators linking it, the phrases outlets quoted and the wording only one side used; TV minutes by channel and the captions; the radio newscasts and its place in each; the shows, investigators and satire on it; what people retell and the names in it; the fact-checks; and the saga it belongs to. Each run's findings around a story are kept (story_extras.json), so its page keeps them after it leaves the front page. Every link is made by our tools and can be wrong. Charts: how many front pages carried it at each hourly run, by lean (an outlet counts while one of its headlines on the story was on its front page, within half an hour of the run); and who carried it when, a bar per headline from when we first saw it to when we last did, with its TV captions, radio newscasts and the day it was retold online beneath.
• TV and radio on the story cards. Each card shows its minutes on screen on each news channel and the hourly newscasts that carried it in the last day (and how many led with it).
• On TV. The captions along the bottom of the screen (chyrons) on CNN, Fox News, MSNOW and BBC News, as the Internet Archive's TV News Archive reads them by OCR about once a minute (its "Third Eye" service, running since 2017), with how many seconds each stayed up. One request a run; each day is kept as sent. The OCR is noisy ("NICOI I E WAL ACE DRECENTS" for the program title Nicolle Wallace Presents), and a line often runs a headline and a guest's name together, so lines are split, the clock and logos set aside, near-identical lines grouped, and a local model (Gemma 4 12B) says what each is (headline, name, promo, ad, junk) and corrects headlines and names for OCR errors only, with the program and the day's story labels as context. On a test batch Gemma 4 12B invented nothing; Gemma 4 26B "corrected" a steel plant in Iowa to Ohio and Qwen3 8B kept garbled promos as headlines. Each headline is matched by meaning to the stories on the front pages at the nearest run; the page shows each story's minutes on each channel beside its front-page rank, and each hour's longest-running captions.
• Motif filing upgraded. Claims are now filed by Gemma 4 26B judging the 8 closest motifs side by side (each by its name, scope note and three of its claims): which of them, none to three, is the claim clearly an instance of? Only when none fits is it asked, separately, whether the claim tells a recurring story worth a new motif; a plain report gets none. Tested on 60 of a person's settled filings, each held out: the old method (pick or name from the 15 closest, Qwen3 30B-A3B) chose their motif for 37%; the new one 53% on the current index (65% on the bake-off's candidates, with 66% of its picks theirs). Asking for a new name in the same call as the judging cost several points, so the two are separate questions. Which model does which job, and how each was tested: docs/models.md.

2026-10-05

• Choosing motifs, tested against a person's own filings. The 160 claims a person had settled on the motif board (each held out of the index in turn) were filed by the live method (the 15 closest motifs by name and claims, one pick-or-name call to Qwen3 30B-A3B) and by a new one (three de-named "shapes" per claim, candidates found by shape, a yes/no check on each of six). The live method picked the person's motif for 37% of claims, 52% of its picks theirs; the new one 47%, but only 28% of its picks theirs. Finding candidates by shape was the weak part: the person's motif was among the 6 closest by shape for 58% of claims, by the claim's own words 74% (82% of the closest 10). Shapes stay out. A third design, the 8 closest motifs by the claim's words judged side by side in one call ("which of these, none to three, is the claim clearly an instance of?"), was then run with seven local models on 60 of the claims (the person's motif among the 8 for 46 of them, so 77% at most): gpt-oss 120B picked it for 62% (56% of its picks the person's; 29 s a claim), Qwen3.5 9B 60% (48%; 1 s), Qwen3.5 35B 58% (50%; 3 s), Gemma 4 12B 58% (46%; 2 s), Gemma 4 26B 57% (54%; 2.4 s), Qwen3 8B 53% (42%), and the live Qwen3 30B-A3B 52% (52%; 4.6 s). Differences of a few points are within what 60 claims can tell apart; every model beat the live method's 37%. Qwen3.5 under Ollama returned an empty list whenever the schema asked for a list of fixed strings, so its answers are asked for as numbers.
• A claim may have no motif. Until now a claim whose last motif a person removed was filed again next run, and the model was asked for one to three motifs per claim, so claims that only report an event were forced into one. Now the model may give none ("a claim that only reports an event or states a fact, with no story told around it, gets none"), a claim whose last motif a person removes keeps none and isn't filed again, and on the motif board a person can mark a claim "no motif: not a story", which takes it out of every motif for good.
• Thresholds rechecked on the bigger embedding model (mxbai-embed-large). Three measures still matched headlines with the small static model (potion-base-8M) at thresholds set by eye. Each was tested on the day's data:
  • Sagas: 118 pairs of saved stories sharing a name, judged by the 30B ("one running story?"): 8 were. Both models separate them equally well (AUC 0.965 and 0.963), so sagas stay on potion; the gate to the judge drops from 0.58 to 0.54, which lets all 8 through instead of 6 (20 questions to the judge instead of 17).
  • Wire share: of a week's 339 outlet headlines sharing 70% of their words with an AP or Reuters headline, all of them copies when read, potion at 0.9 kept 258 (dropping copies such as "White Sox beat Guardians 3-0 in Game 1 of ALDS"); mxbai at 0.85 keeps all 339. Now on mxbai at 0.85, after the word rule: 242 copies in the week's news headlines instead of 201.
  • Curators: 200 aggregator headlines (Google News, Drudge, RCP, Political Wire), each with its closest saved story, judged by the 30B ("same event?"): potion at 0.7 made 60 links, 46 right; mxbai at 0.8 makes 75, 60 right. The judge was strict (it called "Rasmussen no-hits Yankees… Rays win ALDS opener" a different event from the Rays' ALDS opener), so both are better than these numbers. If Ollama can't embed, wire share and curators fall back to potion at their old thresholds.
• Narratives across days. A narrative in one day's report (posts by many people telling one thing) is now compared with the narratives of the last 14 days: the three closest by meaning (mxbai, at least 0.7) are put to the bigger local model (Qwen3 30B-A3B), which says whether it is the same narrative told again (the same claim about the same people or event) or a different one. Linked days make a thread; Folklore cards show the earlier days a narrative was told and by how many people. The first two days (Oct 4, a six-hour report, and Oct 5) gave 3 threads out of 7 questions: the Nantucket air-ambulance search, Saquon Barkley's injury and a disputed Bills call.
• On the radio (#159). The hourly newscasts of NPR (NPR News Now) and ABC News (ABC News Update), which we already transcribe, are now split into their stories in the order they aired. The bigger local model (Qwen3 30B-A3B) reads each transcript with the list of the 40 stories most outlets had on their front pages at the nearest run, writes a short headline for each newscast story, and says which front-page story it is, if any (the same event, not the same topic). Splitting at reporters' sign-offs alone missed anchor-read items. ABC's ad wrapper (a promo before the anchor's "ABC News, I'm ...", a sponsor read after "This is ABC News") is cut off by that pattern first: the model's own ad calls were unsteady. A story is kept only if its first words occur in the transcript after the previous story's (once the model copied the front-page list as a newscast, reusing one real opening for all of them), and a story told twice is kept where it first aired. The page sets each hour's newscasts beside the front pages' biggest stories and counts how often a newscast led with their biggest one. Newscasts from before Oct 4, 13:05 UTC have no front-page snapshot to compare with and are left out.
• Rumors stated the way their tellers state them. The one-line summary of a fact-checked claim sometimes stated the fact-checker's correction instead of the rumor ("the video shows Muslim women queuing for a clothing sale, not free childcare"); 10 of the 80 in the motif index read that way, so motifs filed from them described the debunking, not the story. The summarizer is now told to state what the rumor's spreaders say is true, never the correction, and those claims were rewritten ("a video shows Muslim women queuing for free childcare"). Motif names follow the same rule: the story as told ("Caught on camera"), not the verdict ("Historical footage misattribution"); analysts' words like "narrative" and "controversy" were dropped from names.
• Hand labels. A headline checked by hand on our label-check page now carries the person's mood, loaded-wording and feelings labels in every measure (the model's where the person agreed), marked as hand labels so no rescore replaces them. Only a few dozen headlines, chosen evenly across the model's labels; they also measure the model: of the first 19, mood matched 11 (16 within one step, 1 opposite sign), loaded wording 14 (18 within one), feelings 12. Of the 7 checked since the Oct 4 rescore: mood 5, loaded wording 6, feelings 7.
• Motifs reused, and checked by hand. Named one claim at a time without seeing the index, nearly every motif was a one-off (223 of 259 held a single claim), and the same name was coined as separate entries ("Blame shifting" four times; now merged). Now the model sees the 15 motifs closest to a claim before naming it, files it under one that fits, and coins a new name only for a shape none of them covers; a new name already in the index, word for word, is that motif. Re-filing 20 claims on a copy of the index, 17 found an existing motif (the old way, 1 of 12), though some fits are loose while the index's names are young. A person checks filings one by one ("is this claim an instance of this motif?"); a no takes the claim out of that motif for good.
• Motifs for political news, and cleaner names. Folklore groups that are news reports or shared reactions got no motif, but news is powerful when it confirms a story people already tell (a Russian lab worker's plague death, retold as a bioweapon): political news is now filed in the motif index too. Non-political news, sports picks and shared topics aren't (a WNBA semifinal "prophecy" had become "Valkyries to defeat Aces"). Motif names over seven words are dropped: the model sometimes gave the claim's own sentence, or echoed its instructions, as a name. 25 such entries were removed from the index (10 from sports and moods, 15 sentence names), and their claims refiled where they had no other motif.
• Names merged on the "mentioned:" filters. The model names one subject several ways (Trump / Donald Trump, GOP / Republican Party, Supreme Court / Supreme Court of the United States); names now pass through aliases to one name, seeded for the common cases and curated by hand.
• A saga taken apart by hand, and a saga tracker. The 8B judge had joined an Energy Department employee's arrest for aiding the Houthis with the Yemeni offensive against them ("Yemen-Houthi military conflict"); the 30B keeps them apart, and saved sagas never come apart on their own, so it was taken apart by hand (as the Iran plot saga was on Oct 4). A new Sagas tracker lists every saga kept, with each part's time on the front pages.
• Sagas: a looser name rule and a bigger judge. Two stories could become parts of one saga only if a name rare on the day was in 30% of both stories' headlines; the RAF Fairford bail story and the bombers-pulled-from-Fairford story (0.67 alike) never qualified, since "Fairford" was in 42% of one and 10% of the other. Now 30% of one and 10% of the other is enough to ask the judge (6 more pairs over the week's 91 stories). The judge moves from the 8B, which called the Fairford pair "related cases but different outcomes", to the 30B. Two fixes were needed before it could be trusted. Cut off mid-reason, or answering a bare true/false, it said no after reasoning its way to "the same developing story" (Cornell); it now finishes its reason and answers in words. And the second wording asked about a story "followed over time", which two groups on the same development aren't (it called FlyDubai "redundant"); it now counts the same development and later developments alike. On 14 hand-labeled pairs, 13 agree; the 14th (an Energy Department arrest and the Yemen campaign, which the 8B had joined) it keeps apart, which we think is right. Sagas already saved don't come apart. First new sagas: the Fairford plot, and Alito's remarks (retirement, intimidation of the Court).
• Reuters and AP titles lose Google's source tag. We read both through Google News, which tags each title with the outlet; we removed the tag as a name (" - AP News") but Google has lately tagged by domain as well (" - apnews.com", " - reuters.com"), switching back and forth for the same article: 306 of the 364 Reuters and AP articles stored with more than one title in the last three days differed only by the tag (all but 8 were already ignored as minor rewrites). Both forms are now removed as headlines come in. The tag had made "AP News" a name in AP's stories on the "mentioned:" filters; names are now checked against headlines without it.
• Quotes inside one another count once. Outlets often cut the same quote differently ("Bad Things", "world should accept some bad things", "We Should Accept Some Bad Things Happening For The Benefits"); as separate chips they filled a story card with one quote and pushed other quotes past the four-chip limit (7 of 28 cards with quotes this morning). A quoted phrase that sits inside another, whole words, now joins it: one chip counting every outlet that used any of the wordings, shown as the wording most of them used (the longer on a tie), the others in its tooltip.
• Stories by who and what they name. The local model names the people, countries, places, organizations and groups each current story is about (up to six, by their common short names, remembered per story). A name is kept only if the story's headlines actually contain it (asked about an OPEC+ story, the model added Russia, Iran and Iraq), and the United States is left out (nearly every story is about it). Names in two or more current stories become "about:" filters over the story cards (renamed "mentioned:" the same day: often a story only mentions the name). Not sagas: these are unrelated stories that name the same subject.
• Our own motif index replaces Thompson's chapters. Thompson's Motif-Index is built from folktales and fits modern political rumor poorly, so the site no longer shows it. Instead we build our own: each folklore narrative and each fact-checked claim gets one to three motif names from the model (reusable framings, no names or dates, two-sided when contested), and each is filed under the closest existing entry the model judges the same shape (candidates at 0.7+ similarity, forced choice with "new"), or starts a new entry (M001, M002…). The naming and matching go to the bigger local model (qwen3:30b-a3b); a person merges, renames and deletes entries in an organizer on the label-check page, which suggests merges by name similarity. First fills, with the 8B and one motif per claim: at 0.6 vague early entries snowballed (one took 16 unrelated claims); at 0.76 almost nothing merged (111 motifs from 115 claims), and the 8B still made absurd matches (a tax-vote claim filed with a harassment story). Folklore, Rumors and the Motif index now share their own "folklore" menu.
• Gap: no runs from 3:04 to 7:10 AM Eastern. The desktop's idle timer suspended the machine during the 3 AM run's GPU transcription (which had grown to its full 10 minutes with the new call-in and podcast feeds); it never resumed, so the 4, 5 and 6 AM runs and the nightly folklore report were lost until a cold boot at 7:10. Headlines live on front pages only during those hours are missing. Fix: each run holds the desktop's sleep lock until it's done, and a missed nightly report is made up by the next run after 4 AM.

2026-10-04

• Mood rescore done (8:22 PM; 20,378 stored headlines plus 1,441 that arrived during the run). Over 15,425 news headlines, 79% kept the same mood and 1.1% flipped sign; all leans read grimmer, right-leaning outlets most (average mood -0.60 to -0.76; left -0.50 to -0.62; center -0.29 to -0.39; RedState -0.78 to -1.06), as expected once a headline's own dismay counts. Spin: 77% the same; right-leaning outlets' average rose (0.65 to 0.72) from counting outrage bait, the others held or dipped. Old scores: data/archive/scores-before-rescore.csv.
• Rumors: are they circulating? Each fact-checked claim is looked for nightly in the last 72 hours of our Bluesky and Mastodon sample: posts at least 0.72 alike by meaning are candidates, and the model reads up to 12 one at a time, saying whether each is telling the claim, arguing against it (sharing the fact-check counts), only about the news around it, or something else. The Rumors page shows counts of people, never posts. Stance matters here (unlike for a story's shape): the closest posts are often people sharing the fact-check or the news report itself. First look (half a day of sample, 66,722 posts): 15 of 91 claims seen. Weak spots: some fact-checker items are explainers of real news, so "telling it" is people sharing news; single posts can be misjudged.
• Beyond the front pages. The front page's "Investigations" and "On the shows" link lists are gone (we're not an aggregator): each piece now shows on the story it covers (🕵️ investigated by, 🎙️ on the shows), and a new page lists what the shows, newsletters, streams and investigative desks discuss, tagged by subject by the local model (one to three from a fixed list of 35), plus a directory of all 121 sources beyond the front pages and what we do with each. Subject tags pad: asked for one to three, the model gives two or three even for teaser titles; asking for one unless others are plainly discussed only partly helps.
• Conspiracy scope held back. Even asked first whether a secret plot is claimed, the 8B model called 43 of 94 fact-checked claims conspiracies (e.g. "Jon Husted pushed for data centers and tax incentives"). The scope shows only in local previews until a confidence cascade (the model's own answer probabilities, with unsure answers sent to a bigger local model) is calibrated against hand checks. Rumor class and subject stay on the site.
• Rumor shapes: three open label sets, and a Rumors page. Thompson's index fits folktales, not modern political rumor: in a blind test (151 claims: folklore narratives and fact-checked claims; up to 16 candidate entries each by meaning), the local 8B model picked a motif for 30 of its first 53 claims and the user accepted 1 of 6 checked; a strong model (Claude, in-session) picked 5 of 151, and the user accepted both checked. Specific motifs stay off the site. Instead every narrative and fact-checked claim gets three short labels from the study of rumor, concepts rather than anyone's text: rumor class by what drives it (wish, dread, wedge; Knapp 1944, DiFonzo and Bordia 2007), conspiracy scope (event, systemic, superconspiracy; Barkun 2003), and a subject family (our own list, after the chapters of urban-legend collections). The new Rumors tracker lists the fact-checkers' recent work with these labels, filterable; verdicts are always the fact-checkers' own, behind their links. Survey of other schemes (SemEval 2025 Task 10 narratives and roles, UK election narratives, EIP, Brunvand, ATU, Berezkin, EUvsDisinfo, NewsGuard and more, with their licenses) on #145.
• Narrative labels from one point of view, and only when most posts are about it. The labeler now names the claim most posts share (one or two posts' claim isn't the narrative), says for each part (villain, victim, hero) whether the story has one before naming it (it used to fill every part: "hero: the Bills"), and casts parts and side from the tellers' point of view. Each narrative's sampled posts are then asked one by one whether they're about its claim (telling it or arguing over it alike); under half, it's not counted. On the side: the model's guess is unchecked (it called trans people's own posts "right"), so the site shows a side only from evidence, the AllSides lean of outlets whose articles the posts share (two or more rated shares); the guess shows only in local previews. News reactions get no cast or motif. Re-run on the flagged groups: the Jets penalty and the trans group's misreading are fixed; perspective still sometimes mixes ("Canadians in Florida").
• How a claim's motif is judged. A motif is the story's shape as its tellers tell it: not whether it's true, and not whether a poster tells it or argues against it. (Calling a claim a "false accusation" puts a verdict into the classification; that's the fact-checkers' part, shown separately.) The site presents shapes only in the tellers' voice ("the story casts…"). The model's prompt and the hand check ask the same question.
• A folklore card withheld after a misreading. "Trans women are predators and their identities are inherently tied to dark kinks", told by 12 people and labeled right-wing, was a group of mostly trans people talking about their own gender; one post's line, from one trans woman arguing with others, became the group's "shared claim". It's withheld by hand (data/narratives/withheld.json). Cause: the finder groups posts by topic and the labeler writes a claim even when most posts don't concern it. Fix (next report): a narrative counts only if most of its sampled posts are about its claim, telling it or arguing over it alike.
• Fact-checks (#153). Seven fact-checkers' feeds are archived (PolitiFact, FactCheck.org, Snopes, Lead Stories, Full Fact, NewsGuard Reality Check, Science Feedback; Check Your Fact was dropped: its feed is mostly the Daily Caller's ordinary news). Each check from the last week is tied to a current story, and each from the last 30 days to a folklore narrative, the same way as satire: the closest three at 0.6+ similarity, then the model picks one or "none". Shown with its title (which usually carries the verdict) on story and folklore cards.
  • Checked by hand on Oct 4. Below 0.58 the closest pairs were mostly another claim on the same subject (FactCheck.org on Trump's taxpayer-funded ads matched Ken Paxton's ads at 0.56). The first prompt ("the same event or the very claim") made the model answer "none" even when its own reason said the check was about the story; asking for "N claim" / "N event" / "none" made it worse (no story ties at all). Asking only for the number, and saying a check of a fake video or rumor about an event is about that event: 7 ties, 6 right (FlyDubai, Cornell ×4, Christa Pike), 1 on the same subject rather than the same event (an AI-safety check tied to the "Super Intelligence Force" narrative). The satire ties (now the same code) were unchanged, 3 of 3.
• Call-in shows archived and transcribed (research only, #153). Seven shows whose free feeds carry full hours with callers: Jesse Kelly, Clay Travis & Buck Sexton, Michael Berry, Mark Levin (right); Brian Lehrer (WNYC), KQED Forum, 1A (public radio). No free feed on the left carries callers (Thom Hartmann's and the Majority Report's caller hours are YouTube-only or for paying members), so public radio stands in there; the sample leans right. C-SPAN's Washington Journal, which sorts callers by party line, has no feed and its archive runs on a lag, so it's out. Never shown; callers are private people, so only patterns will be published.
• Satire (#139): which stories become jokes. Ten satire sites are read every two hours (the Babylon Bee; the Onion, Borowitz, New Yorker humor, McSweeney's, Reductress, the Hard Times, NewsThump; Duffel Blog, ClickHole; leans by reputation). Each joke from the last three days is offered the closest current stories with mxbai-embed-large similarity 0.5 or more (up to three) and the model picks one or "none". Shown only as 🃏 "joked about by" on the story card, never as coverage or in any measure.
  • Checked by hand on Oct 4. With no similarity floor, the model tied 6 jokes to stories and 2 were wrong, both on a shared word ("Woman survives two minutes in Primark" -> a botched execution; a paste-eating Marine -> a Marine arrested in Japan). Asking it to name the event the joke mocks first made it worse (9 ties, 4 wrong). Every wrong tie was under 0.48 similarity and three of four right ones over 0.54, so with the 0.5 floor: 3 ties, all right. Missed: the Bee's oblique Cornell joke (0.44). With only the fallback embeddings, satire is skipped rather than matched on another scale.
• Mood redefined: how the headline presents the news (rubric b31d654f). Mood was how good or bad the event is for the people it affects, whoever reports it. For political news that depends on whose side you're on, so the model picked a side itself and a reader couldn't check it: RedState's "Oh, Hell No: Judge Rules Federal Noncitizen Voting Ban Unconstitutional" scored ☀️ upbeat+ (good for noncitizens). Mood is now whether the headline presents its news as good or bad, from the outlet's own words and tone; plainly reported harm is still grim. Spin now also counts outrage bait in plain words ("Oh, come on:").
  • Tested on 21 headlines (12 hand-checked in the label check, nine outlets' headlines on that ruling): mood exact 13 vs 10 under the old rubric, within one step 19 vs 16. On the hand-checked 12 alone, exact 6 vs 6 and within one 10 vs 9; the one sign flip (the Daily Wire's "Leftism's appetite for the destruction of order" read as upbeat+) is fixed. The ruling set's expected moods are the assistant's reading, not a hand check.
  • Every stored headline is rescored with the new rubric (--rescore-outdated, between hourly runs). The old scores are kept in data/archive/scores-before-rescore.csv with the rubric that made them. Mood before and after Oct 4 means different things, so any mood trend starts here.
• Shortened links scrubbed (privacy fix). Bluesky writes links in a post's text shortened and without https:// ("youtu.be/…", "twitch.tv/someone"), so the scrubber missed them: 4,998 stored posts kept a full link, some naming an account. They now become their site like every other link, and the stored posts were rewritten. Posts need six real words to be grouped: a run of emoji and a video link is promotion, and a ring of such posts had reached the page as an 11-person "conspiracy theory".
• Folklore page floor: 10 people. In the first full report (25,095 accounts), half of the 60 retold narratives had five to nine tellers, and those were mostly a handful reacting to one news item or game. The page shows narratives told by at least 10 people, or 1 in 2,500 accounts sampled if that's more; the report keeps the rest.
• Shared article links kept (research only). A link in a post to an article on a site we scrape is kept whole (without its query string), since it ties the post to a story exactly; other links are still cut to their site.
• What people say, unprompted (research only, #153, #142). Each run samples five minutes of Bluesky's public stream (Jetstream) and three pages of mas.to's federated timeline, on a thread during the scrape. Kept: post text, reply/quote flags, a salted fingerprint of the author. Handles in text become "@someone", links just their site; images and video are never fetched. Skipped: non-English (by script and common words, since posts often don't say), posts their authors labeled sensitive, sex spam, Mastodon accounts that haven't opted in to being searchable or are marked as bots. Deleted posts are deleted here too; raw text expires after 30 days. Once a night the day's posts are grouped into paraphrase groups (mxbai embeddings, mutual nearest neighbours at 0.78, accounts posting more than 20 times in six hours and bare link shares left out), and each group of 5+ people is scored for wording variety (one minus mean word overlap): varied wording from many people is a told narrative, the same words copypasta. The local model labels the biggest groups: the shared claim, whether it's retold or just a shared topic, genre (rumor, contemporary legend, conspiracy theory, folk belief, joke formula, testimony, news reaction...), Thompson Motif-Index chapter and a motif, Propp-style roles (villain, victim, hero), politics and side. Each narrative is matched to current stories: the model chooses among the three closest stories ("same event", "same issue" or, most often, "none"). A yes-or-no check, one candidate at a time, agreed to nearly anything (a Drake Maye fumble "matched" voters calling J.D. Vance "a goober"); the forced choice linked 12 of 60 narratives on Oct 4, 10 of them rightly. Matching narratives to The Focus Group transcripts (12 episodes, Ep50-61, transcribed Oct 4) was tried and dropped from the page: the passages rarely overlap a day of social posts, and the yes-or-no matches were wrong almost every time. The first report (40 minutes of stream, 15,430 posts) found 13 retold narratives, among them a folk belief that Trump is being kept alive by doctors because "they need him alive" and a contemporary legend of "Julie bots". Reports stay in data/narratives; nothing is published.
• Stories use a bigger embedding model. Four models were compared on the day's 6,733 headlines at equal coverage (the threshold at which each put about 700, 900 and 1,100 headlines into stories of 6+ outlets), with precision from cross-outlet pairs sampled inside each model's stories and recall from a shared pool of 657 pairs sampled across every model's similarity range, all judged by the local language model under two wordings (agreement only; about 80 judged pairs per cell, so ±0.04). At ~900 headlines: potion-base-8M (until now) precision 0.77, recall 0.57; mxbai-embed-large 0.91, 0.57. At ~1,100: 0.82 / 0.64 against 0.88 / 0.68. potion-base-32M and nomic-embed-text fell between. Stories now use mxbai-embed-large through the local Ollama at 0.80 (about 1,000 headlines in stories); its vectors are cached by headline, so each run embeds only new headlines. If Ollama can't embed, stories fall back to potion-base-8M at 0.70. Sagas, wire share and aggregator matching keep potion-base-8M, whose thresholds were tuned on its scale.
• Each side's wording. Story cards show two-word phrases that outlets on one side of the lean scale put in their headlines and no outlet on the other side did ("hero pilot" and "FlyDubai hijacker" from the right against "extremism concerns" and "terrorist act" from the left on the FlyDubai attack; "Trump doxes" against "Trump urges" on the senator's phone number). Our outlets lean left as a group, so a story needs 3 or more rated outlets on each side, and a phrase needs 2 or more outlets and a quarter of its side's outlets, with none on the other side. Function words and headline filler don't count; phrases the same outlets used back to back are joined ("small Texas town stopped Trump's"). No model involved.
• Every-headline file packed. headlines-table.json now lists each row's values in a fixed column order, with outlets (and their lean), topics, stories and feelings listed once and referred to by position, and URLs without "https://": about 12% smaller compressed (keys were a third of the raw file, but compression already removed most of that; URLs and titles are the bulk). The page unpacks it, and the download button saves the rows with named fields as before.
• Cloud paper by spiciness. Each word's scrap of paper is colored by how loaded the wording of its headlines is, white (plain) through cream to peach (spicy), instead of a random tint.
• Supreme Court page (#144). The term's cases come from the Court's Granted & Noted list (a public-domain government record); what each is about comes from its question presented on Wikipedia's list of pending cases (CC BY-SA), which the local language model puts in a few plain words ("Cook County AR-15 gun rights case"). Both are fetched at most weekly. Headlines naming the Court or a justice are read by the model with the outlet and its country (India's Supreme Court had been slipping in): whether each is about the US Supreme Court, which of the glossed cases it's about, a few topic words, and the stage it reports. A small model reading a list of thirty cases matched every term preview to five or six of them, and two headlines about Justice Alito and the Dobbs leak to the Arizona voter case, so a suggested case is kept only if the headline shares a word specific to that case (from its gloss, parties, question presented or the state court it came from, counting only words in at most two cases' texts) and the model, asked about that case alone, agrees it clearly refers to it. A distinctive party name ("Suncor", never "Congress") ties a headline to its case as well. Each case in the news gets a card like the front page's story cards: the lean of the outlets covering it against all Court coverage, mood, loaded wording, feelings, quoted phrases, and each side's most loaded headline side by side. Court news about no case on the docket is grouped by topic; docket cases no headline covered are listed. Window: 14 days; each outlet counts once per card.
• Story coherence check tried, not adopted. The model was shown each story's three most central headlines and asked about each loosely tied member (best similarity to that core under 0.75); rejected ones were taken out and clustered again among themselves. On Oct 4's headlines (labels as in the threshold test above, 90 pairs per setup): at 0.70 it raised precision from 0.87 to 0.91 but cut recall from 0.79 to 0.73; at 0.67 and 0.65 it didn't help, and the biggest stories kept their size (they were real stories: the FlyDubai attack ran to 121 headlines). All within the noise of the sample, so stories stay as they were.
• Story threshold re-tested; it stays at 0.70. Stories are still the connected groups of headlines whose embeddings are 0.70 alike or more. On one day's headlines (6,888) the local model judged about 1,000 headline pairs, under two wordings, keeping the pairs both wordings agreed on. One set was spread across the similarity range; it measures how many same-story pairs end up in one story (recall). The other was drawn from inside the stories each threshold would make, grouped by the highest threshold that still joins each pair; it measures how often a story's headlines really belong together (precision). Pairs that only join below 0.70 are the same story 73% of the time between 0.675 and 0.70, 51% between 0.65 and 0.675, and 16% between 0.625 and 0.65, because low thresholds chain unrelated stories (the largest group grows from 127 headlines at 0.70 to 451 at 0.60). Scored with precision counting double (F0.5), 0.69 and 0.70 were best and within noise of each other. Lowering to 0.65 catches more same-story pairs (recall 0.79 to 0.88) but drops precision from 0.85 to 0.75. Splitting the chains instead with Louvain communities or mutual nearest neighbors scored lower. A threshold has no gradient (scores move in steps as it crosses pairs), so it was swept every 0.005 from 0.55 to 0.80.
• Saga links checked by the language model. Before two stories first join a saga, the model reads a few headlines from each side by side and says whether they're one running story; the similarity and shared-name rules still apply first. A shared name alone had linked separate Supreme Court cases and separate plots by Iranian nationals. Verdicts are cached (saga_judgments.json); those two sagas were taken apart.
• Headline A/B tests (#119). Every wording Slate tests on a front-page card is recorded each run with when it was first and last up; when one is left, it's the winner. Shown on the headline changes page. Recording begins Oct 4. Checked on Oct 4 whether other outlets expose their variants the same way (each homepage once): none do. Outlets using Chartbeat's or Optimizely's headline testing run it in the reader's browser, so the variants never appear in the page a scraper reads.
• Headline download. The every-headline page links its rows as JSON (the file the table already loads).
• Quote selection (#149). Story cards list the phrases outlets put in quotation marks in their headlines (one headline per outlet, up to 4 phrases, most-quoted first), each tinted by the lean of the outlets that quoted it. Phrases of 2 to 80 characters count (an apostrophe inside a word doesn't open a quote); matching ignores case and punctuation, and each outlet counts once per phrase.
• Daily editions begin (#125). At each run the front page is saved as that news day's edition; past days are published as fixed pages under /YYYY/MM/DD/ with an archive index. Editions start Oct 4; earlier days have none.
• Outlet pages (#126). Each outlet gets its own page: live headlines and the stories they're in, recent rewrites, and the mix of feelings its headlines stir, over the outlets page's 30-day window.
• RSS feed (#123). feed.xml lists the current stories; each item keeps its id across runs while the story lives.
• Aggregators aren't outlets (#151). Google News, Drudge Report, RealClearPolitics and Political Wire no longer count toward stories, lean, the news-day sticker or the cloud (they mostly link to other outlets' stories); their live headlines are matched to stories (embedding 0.7) and shown as "picked up by" on the cards.
• Wire share (#152). Outlet cards show the share of an outlet's news headlines over 7 days that are AP or Reuters headlines near word for word (embedding 0.90 and 70% of words shared; checked by hand). Outlets mostly rewrite wire copy: on Oct 4 the highest was about 11%.
• Story snapshots and trend arrows (#133). Every run saves each story's lean, mood and outlet count (story_snapshot). Cards show arrows when lean or mood moved over the last 6 hours (needs 2+ hours of history): lean 0.15 / 0.35 / 0.6, mood 0.1 / 0.25 / 0.45 for one, two or three arrows.
• Bright side removed. Its picks (hope/joy plus a "widely good news" model check) read as feel-good filler rather than a measure.
• Blindspots move onto story cards. Same rule; the separate box is gone. A blindspot's card shows a lean meter (left/center/right counts) with a tick where an average story splits, and a 🙈 filter shows only blindspots.
• Transcription: temporary failures retry. Out-of-memory and network errors are no longer recorded as failures (they stopped every item for good on the first prod run); only a missing or unreadable file is. Transcription waits for the language model to unload first and skips the round if it can't. The test suite no longer calls the real language model (it was loading it onto the shared GPU mid-transcription).
• More shows. News outlets' own daily podcasts shown (Reuters World News, BBC Global News, WSJ What's News, Bloomberg News Now, Politico Playbook, The Economist's Intelligence, CNN 5 Things, the Guardian's Today in Focus, Morning Wire); more political YouTube channels shown (Pakman, Secular Talk, Brian Tyler Cohen, MeidasTouch, TYT, Adam Mockler, Timcast IRL, Benny Johnson, Ben Shapiro, Dan Bongino, Matt Walsh). Archived only: ABC News Update (hourly), NBC Top Story, Al Jazeera's The Take, Steven Crowder, Candace Owens and long-form talk (Rogan, Theo Von, Flagrant, Jimmy Dore, PBD, Pod Save America's channel). 72 sources, 37 shown.
• Headline changes: junk prefixes. A rank from a "most read" list ("4 , ") or a view counter ("114.7k views : ") before a headline is stripped when scraping and ignored when comparing versions, so a story moving from #4 to #2 is no longer a "change".
• Headline changes: how a rewrite changed. The language model reads both versions side by side and says, in a few words, what the rewrite changed ("added the death toll"), cached once per change. "Plainer wording" / "reads as better news" notes show only when that side-by-side reading and the two versions' own scores agree on the direction; either alone was noise (identical headlines came out "plainer").
• Backups every run. Each run first takes an online copy of the database (newest 5 kept), besides the nightly copies (30 days).
• Trending ranks by what's on front pages now. "Trending in the news" ranks stories by the share of outlets carrying them on their front page right now, weighted by age the same way as the news-day sticker (full weight for 12 hours, then halving every 24), instead of every outlet that covered them in the last day. A story outlets have dropped falls down the list. Items show both counts: on front pages now, and in all.
• Sagas are kept across days. Sagas are saved (saga table, story.saga_id). Each hour, today's stories are merged with the last 7 days of saved stories, starting from the saved sagas, so a saga grows or merges but never loses a part. Linking now needs a shared distinctive name (a word outlets capitalize mid-sentence) rather than any distinctive word, and hyphenated words split ("UK-Iranian" carries "Iranian"). Checked on Oct 3-4: Cornell keeps all 4 parts; a false link ("Trump rally in Alabama" with "Ohio layoffs before Trump rally", via "rally") is gone.
• Podcast transcription. Archived podcast audio is transcribed on our own GPU with Whisper (faster-whisper large-v3-turbo, int8), after each hourly build, hourly newscasts first, for items up to 3 days old. Sponsor reads at the start and end are trimmed. Transcripts are kept with the model's name, for analysis only.

2026-10-03 (evening)

• Shows archive begins. 40 newsletters, podcasts and political video channels are collected into side_item (append-only), each feed read at most every 2 hours with conditional requests. Sources are sorted by how they follow the news (news of the day vs. thinking, #135), not format; only news-of-the-day ones are shown on the site.
• Stories: different money stays apart. Headlines that both name dollar amounts more than 3x apart can't link ($90 Medicare checks vs. a $5,000 promise); $90 vs. "nearly $100" still can. Headlines without amounts unchanged.
• Stories: a split story gets two titles. When one saved story's coverage splits into two clusters in an hour, the bigger part keeps the story and its title and the smaller becomes a new story (before, both showed the same title).
• Headline changes: live-blog filter. Only an all-caps LIVE tag or phrases like "live updates" mark a live blog; the plain word "live" no longer does, so ordinary rewrites containing it now count.
• Scraping. Bare photo credits with no text left after cleaning are dropped; hrefs with stray spaces are kept.

2026-10-03

• Score provenance. Each scored headline now records which model and rubric version scored it, when, and the model's own note on who the event affects. The rubric id is a hash of the prompt and schema, so it changes on any edit to either.
• Trend sightings. Every appearance of a trend in a source's list is now logged with its rank (before this, only first and latest sightings were kept).
• Outlet ratings. Lean switched from MBFC to AllSides (5 levels, CC BY-NC 4.0); reliability from Wikipedia's perennial sources list (CC BY-SA 4.0). Outlets AllSides doesn't rate are left out of every lean average.
• Our lean estimate. Wording-only model (tf-idf + ridge) trained on rated outlets over 30 days; published only when leave-one-out rank agreement is 0.45+ and the outlet is 0.55+ from center. Never counted in lean averages.
• Emotions. Ranked top three per headline, averaged ranked-choice style (3:2:1). All headlines rescored with this rubric (14,404).
• Stories and sagas. Stories: headlines linked at embedding similarity 0.7+, six or more outlets, 24-hour lookback. Sagas: stories whose centers are 0.58+ similar and share a distinctive word.
• News-day sticker. Live headlines only; a story's weight is full for 12 hours, then halves every 24 hours.
• Duplicates merged. 2,470 duplicate articles and 1,468 duplicate headlines merged (backup kept).
• New outlets. 11 added from their feeds (TPM, Puck, The Washington Sun, The Lever, Mediaite, Straight Arrow News, UnHerd, Just the News, The Daily Signal, The American Conservative, Washington Reporter). Vox and National Review switched to their feeds.

2026-10-02

• Archive begins. The current database starts this day; anything earlier isn't in it.

History

• 2024-01 to 2024-08: built by hand as Maudlin: per-outlet scrapers, the database, the word cloud, word-list sentiment (VADER, AFINN). About 680 commits.
• 2026-10: rebuilt with AI assistance (Anthropic's Claude as a coding assistant) for the 2026 midterms. The direction, design and decisions are the maintainer's; the assistant wrote and tested much of the new code. Separately, the site's measures come from a small open model run locally (see above for its provenance).