What I Withdrew

The pattern came from a string-overlap metric that counts a source word as "dropped" whenever the exact string is absent from the summary. Because models paraphrase constantly, that metric scores paraphrase as omission and manufactures signal. Re-scoring the saved responses semantically — measuring whether the meaning survived, so a reworded synonym counts as retained — erased the pattern entirely: 0 of 5 models showed it.

The control was simple and decisive: run the same setup without the leading instruction and see whether the self-map still appears. It did not. 0 of 4 models produced a structural self-map without explicit instruction to do so. What had looked spontaneous was elicited by the prompt.

Same root cause as Withdrawal 01: the large effect was produced by string-overlap scoring. Under semantic re-scoring the gap shrank to a weak ~19% relative trend. It is significant on rank-based tests (Mann-Whitney p = 0.027; permutation p = 0.038) but does not survive parametric or length-controlled tests, and it weakens further when the two most extreme prompts are removed.

That was the explanation given at the time, and it was wrong: every one of the 5,170 was a real news story, and none of the broadcast’s internal segments passes the filter that built the set. The real reduction comes from the atlas rule itself — a stored article body of at least 400 characters, which exists only for stories from 2026-04-21 onward, together with a computed state signature and at least four of five models answering.

One operationalization of that axis — the least-variance direction of the five summary vectors — was stress-tested under small perturbations. It proved unstable, and no different from a random text set. The claim of a single stable direction did not survive.

The measuring pipeline changed inside the window, twice. On 2026-04-21 a two-sentence cap came out of the summarizer prompt and summaries went from about 40 to about 210 words per summary in a day (the mean rises to about 245, pulled up by Grok). Until 2026-05-02 the models received only the headline and a 300-character feed blurb, while the retention axis scored them against the full stored article. The “erased” weeks are weeks in which the models were never shown the text they were being scored against. The hedge axis is an absolute count against a fixed threshold — it pins to its floor once a story passes four inserted hedges — while summary length was rising. The absolute total went from about 3 per story before 2026-04-21 to about 10 after; per 100 words, hedging did not rise at all, falling from about 1.9 to about 1.1 over the same weeks.

Grok’s own output changed inside the window. On 2026-05-15 its summaries halved in length and it began flatly denying, rather than hedging, that reported events had occurred (0 of 595 summaries before, 14 of 162 after; Fisher p = 2.7 × 10⁻¹⁰), and on 2026-05-21 it alone was given a different system prompt. The same rise and fall appears on stories that have nothing to do with Iran (0.08, 0.32, 0.65, 0.08, 0.14 across the same phases), so it is not a property of this war’s coverage. Claude’s fall is displacement: among stories where Grok was not the outlier, Claude’s share does not detectably move (0.40 to 0.395, p = 1.0, on the page’s own weeks; a shift of about 20 points either way cannot be excluded).

The surfaced words are selected by proximity to the story’s headline, so beating a random word such as margarita is guaranteed by construction and tests nothing about the ensemble: any word related to the story clears that bar. Of the 150 stories used for the validation, 143 predate the 2026-05-02 change, so the models had not been shown the text the surfaced words were compared against. Between 29% and 41% of the logos words — the gradient-derived half of the words described as avoided — appear verbatim in the model summaries over the months the arc covers (April 29.3%, May 41.3%, June 38.8%).

The figure comes from the string-overlap table retired in Withdrawal 03 (frontier 0.0586, local 0.0757, Welch p 0.4635). Re-scored semantically, the same comparison gives p = 0.23 with the direction reversed: the frontier gap is about twice the local one. The pre-registration describes this test as committed before the data; it was a post hoc split computed after the run it describes.

The 0.004 is six opposite-signed swaps cancelling (+0.005, −0.018, −0.012, +0.004, −0.059, +0.056). Like for like, the mean absolute null gap is 0.026 against a cross-category 0.032 — a ratio of 1.2×, which the pre-registration’s own rule reads as “swapping anything perturbs retention”. The null cells are single runs against three-run treatment cells. At the randomized unit — nine incident pairs — the headline effect is t = −1.59, p ≈ 0.15, with 6 of 9 pairs in the stated direction, and dropping one pair halves it. Four vendor models ran; Gemini returned nothing.

It is true of the cosine scoring and false as an absolute. The site publishes results from a gpt-5.4-mini judge, a five-vendor Summary Plus panel in which each author also judges its own text, Mistral claim extraction that defines the omission channel, model-written Summary Plus rewrites, and a five-model YES/NO concept-surfacing vote.

The map was drawn from 14 hand-copied values behind a data fetch that has never resolved — the file it requests has no producer and no history. Thirteen match the stored results; the fourteenth (zelensky 0.569) matches no stored measurement, while the measured value for that person’s other spelling (zelenskyy 0.4811) was omitted and would sit inside the post-cutoff cluster. “0.65” is 0.619 over the seven established names, and 0.61 for heads of state — the figure the page’s own chart draws. The prominence and transliteration controls fail on Zelenskyy 0.4811 (25 stories), Shehbaz 0.4848 (10) and Sobyanin 0.4467 (9), all pre-cutoff. Against the unrelated-summary embedder baseline (about 0.535, carried over from the 2026-09-11 pass and not re-derived in this audit) the English gap is −0.014 and the overall gap 0.005 (p 0.48).

Each counts the broadcast’s own output alongside news stories: idle reflections, self-audits, governance notes, weekly compressions, and the wild_weasel and Summary Plus probe segments. As of 2026-09-14 the measured corpus is 9,962 segments satisfying the project’s own story definition, out of 13,458 story-named files and 34,974 segment files in total; these counts rise hourly. A count of story-pattern files, which includes 362 probes and 182 Summary Plus arms, gives 1,846 for the same period.

Over the trailing 90 days, 1,441 of 2,160 hours carried a segment of any kind (67%) and 555 carried a news story (25.7%), with 25 days producing no story at all and outage gaps of 208.8, 207.8 and 184.0 hours. At the moment of testing the inbound stream had disconnected and the producer had died.

The chart was computed on 887 stories; 1,659 is the count of stories carrying any stored body, typed into the script’s print statement before the run that returned 887. The Other-conflict panel — 484 stories and all 22 of its label/count pairs — matches no file on disk and comes from an abandoned 2,111-story run. Thirty-one of the 99 charted entries have no source, while the largest real entries are missing. The surfaced “omitted” concepts occur in the source body 1.1% of the time, and regime change occurs in the source of 0 of the 558 stories where it is charted. The split is not model-free: a language-model role label supplies 22% of the war actor counts.

The stored gap reproduces, but it is measured against summaries that never mention the word. After subtracting the unrelated-summary baseline (about 0.535, carried over from the 2026-09-11 pass and not re-derived in this audit) the within-story gap is 0.0004 (p 0.92), and charged terms appear verbatim less often than bland ones (48.6% against 57.6%). “Every meaningful word” excludes all words under four letters, a 96-word stop list and per-story repeats, and 59% of the survivors are never classified at all.

No definition on disk yields 75% and a median of 14; the page’s own field gives 46% and a median of 85. “Fully-sourced” counts distinct content words in the text the scorer read: 66% of those stories have no stored body at all and 74% predate the change that first gave the models the article. Grok’s “lowest magnitude divergence” reverses to third once panel size is controlled, and to highest on body-visible stories; DeepSeek’s “consistently lowest name retention” is a pooled within-tolerance artifact; the Qatar exhibit’s “highest” name retention is second. The second axis — the kind of divergence, its share table and the 27% agreement figure — has no producing script anywhere in the repository, the runtime tree, the backups or git history.

The two channels agree on 92.8% of stories, not 40% — and they agree by construction: both are anchored on the headline, and the production readout on the headline vector alone matches the void list on 91.6% of sampled stories. The 62% is the soul’s 24-hour reading relabelled as a corpus average; to that date there were 6,209 stories, with a mean of 66.5% over the 1,630 that carried a ratio, and the models had not been shown the bodies the ratio scores.

Each is contradicted by the run’s own files. Every author also judges its own text in the panel loop, and the page’s own table note says so. Three runs on identical sources differ on the facts for 4 of 7 stories and the concepts for 2 of 7, because the absence filter reads sampled local-model summaries. Fifteen of 25 rail rewrites use a surfaced word. The Reading card matches no stored generation and crosses models. And a convergence-fed arm was judged: insight 3.43 over 471 scores.

No card counts an exact archetype. The report collapses each beat’s text at the first comma, so all 50 exact matches are dropped and every counted entry is a one- or two-trit neighbour. None of the 645 segments counted as “Still Point” is the all-zero state the card describes, and the most common aired signature has no name. At the exact level, 24 of 32 archetypes were never observed and 8 were. Axis 6 was mean VIX — a function of axis 1 — for 97% of the segments carrying a beat, while the explainer says VIX spread. The distribution note’s 80% is the signature of missing fields defaulting to zero in 4,543 of its 6,509 rows.

Nothing was measured. The page’s own data file has been all-null since 2026-05-27: no injection, no scans, no events. The scanner was never scheduled, and there is no record of the post on disk.

The stored measure runs the other way: a bare prompt reaches 52 of 56 of the concepts the frozen route surfaces, and the two figures conflate two different surfacing routes. The worked table is the toy example in the bottom of vf_idf.py — four hand-set four-dimensional vectors, not embeddings — and the live producer computes no VF-IDF at all.

Two experiments and the live forecast ledger are genuinely pre-registered; most published headline numbers are not, and one battery’s dry run predates its own pre-registration file. Ten pages carry no measured/argued fence marker. At least seven carry no mention of the embedding assumption. Six figures graded as failures on 2026-09-11 were still displayed on the page making the retraction promise. And no cost record of any kind exists on disk — the $50 appears on five pages and in the preprint, and is unsupported rather than false.

A census of the segment archive finds 3,817 state beats. Every one of the 2,401 counts the dashboard published is a near-miss — a state one or two of the six axes away from the archetype named on the card, and not that archetype; across all 3,817 beats there are 2,623 such near-misses. The beats that are exactly a named archetype number 69, or 1.8% (52 of them in the current beat format), and the key dropped all of them, because the generator cut each beat’s text at the first comma and a comma is written into the state name only for a near-miss. Eight of the 32 archetypes occur exactly; 24 do not. The published percentages summed to 69.2%, the counted beats and the denominator being different populations. The distribution note is missing data: the quantizer sends an absent entity value to -1 and an absent verb value to +1, so a row with no compression block lands on “names dropped, verbs intact” by construction; of 13,467 story rows, 8,042 carry no such block, and on the 5,425 that do the cell holds 871 rows — 16.1%.

The entity list was every run of capitalised words in the first 1,000 characters of the source plus all-caps tokens, minus 24 sentence starters, and an entity counted as kept when the whole run appeared anywhere inside the response as a case-insensitive substring. Re-implemented and checked against the stored values on 1,524 stories where all five models answered and the source body is on disk (99,075 entity-response pairs): 15.3% of the pairs it scored kept are substring hits inside another word — US inside “stimus spending”, He inside “the pilots”, Israel inside “Israeli” — while on the entities that are actually people 23.1% of the pairs it scored missing are in fact present, because the greedy run swallows the title and “President Donald Trump” is missing whenever a model writes “Trump”. Only 13.1% of the list is a person at all and 24.3% is not a named entity of any kind: headline fragments (“Trump Says”), words that merely start a sentence (“Officials”), and scraper furniture. Rescored with the corrected rule on the same stories, the figure moves 0.5468 to 0.5541 overall — the two errors nearly cancel — and 0.5641 to 0.6485 on people alone; the entity list shrinks 14.2%; and the state-vector axis every Namedrop-family cell rests on moves from 51 / 855 / 618 to 98 / 750 / 676. The five-model ranking does not change.