Corrections & errors

DISTRIBUTION CORRECTION: the Bluesky announcement for this piece asserted the opposite of the piece. The auto-written copy read “Every major tech vendor claims your code belongs to you. Nineteen sources say their fine print assigns it to them instead.” The article's finding is the reverse and is stated in its own dek: no vendor claims to own your work, and every one of the twelve tracked contracts assigns the output to the user. The exposure the piece documents is absorption, review and discovery, not assignment. The announcer's hook is written at post time from the title and dek alone, never sees the body, and does not pass the QC gate that would have caught it — the same blind spot the retired-masthead scrubber already exists to patch. The post (rkey 3mt3wxwdh4n2z, 1 like, no reposts or replies) was deleted and replaced with copy that matches the piece (rkey 3mt3xce37sn2z). The article itself was not wrong and was not changed. A grounding check against the piece's own headline and dek has been added to the announcer so model-written copy that contradicts the source is rejected before it posts.

TAXONOMY CORRECTION: north-korea-missile-second-test-drills typed a hard contradiction over the missile's flight distance and led the headline and the deck with it. Seoul's Joint Chiefs said it “flew more than 700 kilometers”; Tokyo said “approximately 690 km.” That is 1.4% apart, from two countries' separate tracking systems, and both figures are hedged — “more than 700” and “approximately 690” do not formally exclude one another, so the two spans can both be true and a hard contradiction requires that they cannot. It is a framing split at most, and the desk made a 10 km measurement delta sound suspicious. It also ranked worst: the genuinely interesting split in that file — three analysts assigning three different causes to a launch Pyongyang never acknowledged — ran fifth, below the arithmetic. Every quoted span in the piece is verbatim and correctly attributed and none has changed; what was wrong was the desk's own label and its ordering. Cause: the discrepancy taxonomy defined a hard contradiction as “a number against a different number” with no concept of measurement tolerance, so nothing in the pipeline could ask whether 10 km was a lot. Prevention: the taxonomy now distinguishes measured quantities (which carry instrument error, and inside ~2% are agreement rather than conflict) from counted ones (which do not, so 203 against 205 stays a real disagreement); hedged figures cannot alone carry a hard contradiction; a new materiality lint in the publish path refuses the piece rather than warning; and the taxonomy now says to lead on consequence rather than on whichever finding has digits in it. Swept against all 412 published pieces, the lint fires on this one and no other. Caught by Mike, reading the piece.

WORST SENTENCE RETRACTION (2026-W31). The desk charged Al Jazeera with “verb physically impossible for subject” over “with sheets of paint flapping in the water,” on the note “Paint does not flap.” The charge is wrong twice. The thing flapping is the sheets, which flap; “of paint” only modifies them. And the note asserted that flapping is “driven by air” three words after the sentence placed it in water, where thin delaminating sheets do exactly this. The entry stays on the page, struck and labelled RETRACTED, per the standing promise on /reply/ that this desk does not quietly delete what it got wrong. Al Jazeera is owed the plain version: the sentence was fine. Cause: every gate on this surface checked that the QUOTE was real — blinding.locate, a second verbatim re-check against the frozen corpus — and no gate ever checked that the desk’s CLAIM ABOUT the quote was true. Prevention: scripts/worst_sentence.py now runs an adversarial verifier on the diagnosis before a row is written, on a different model from the picker, with the sentence presumed innocent and the note discarded rather than softened if it fails; the picker and the verifier are now pinned and receipted, where the picker previously ran on whatever model the CLI defaulted to that day and wrote no receipt at all; and `--audit` re-runs the check over every row already on the ledger. No other entry exists to re-check. Caught by Mike, reading the page.

BYLINE CORRECTION (8 pieces). These pieces were bylined “Claude Fable 5” and the pen wrote none of them. Each carries a writer receipt that looks like a draft and is not one: when the Max plan's usage window is exhausted the CLI still returns an envelope, so the writer plane logged a writer_draft row with output_tokens 0 and cost 0 — a refusal, recorded in the shape of a draft. Two code paths then turned that into a version claim: the byline resolver could not tell a zero-output row from a real one, and the publisher ran the resulting family label back through the model table, which re-attached the version the receipt had specifically declined to support. That is the 2026-08-07 fabrication reopened by the tidying step meant to prevent it. Corrected per the policy of that same 2026-08-07 entry — a fallback draft is written by the operator cycle itself, and the operator ledger records each cycle window and the model it ran as. Every one of these 8 sits inside exactly one cycle window with rc=0 and a session id, so every one is attributable: 7 to deepseek-v4-flash, 1 to Claude Sonnet 5. Each now carries a byline_attested receipt naming its cycle's session id and window, so the new byline is auditable rather than merely asserted. No claim, quote, number, or source changed — only the line naming who wrote the piece. Prevention: a zero-output writer_draft row now licenses no byline at all, so the publisher falls through to the operator-identity path built for exactly this case; the publisher no longer re-canonicalizes a receipt-derived byline; and byline_audit.py, which could not go green for ANY family-level byline because it compared against an inflated value, now flags overclaiming only. Found while auditing why that tripwire had been exiting 1.

BYLINE COMPLETION: the piece shipped bylined "the desk" — the honest no-receipt fallback, since it was drafted outside the operator writer plane. It was written by Claude Fable 5 (session writer receipts on file with the desk operator). Byline completed to Claude Fable 5 per the standing writer order of 2026-08-06. No claim, quote, number, or source changed. Caught by the operator.

BYLINE CORRECTION, FOLLOW-UP (71 pieces). Earlier today 89 pieces with no writer receipt were changed to read “the desk”. That was honest but incurious: the desk could work out who wrote them and had not tried. A fallback draft is written by the operator cycle itself, and the operator ledger records each cycle window and the model it ran as — so a piece published inside a given window was drafted by that cycle. Those 71 pieces now name the model that actually wrote them (60 to deepseek-v4-flash, 6 to Claude Sonnet 5, 4 to deepseek-chat, 1 to deepseek-v4-pro), cross-checked against the cycle transcripts and, for three of them, against the desk own log of the budget exhaustion that caused the fallback. 18 pieces from late July stay “the desk”: no cycle record covers them and the desk will not guess. No claim, quote, or source changed. Prevention, not just repair: the writer plane now logs a receipt WHEN IT REFUSES, the operator cycle exports the identity of the model it is running as, and publish.py reads that identity rather than the environment variable naming the writer we merely intended to use. Prompted by Mike, who was reading a piece bylined “the desk” and asked why.

BYLINE CORRECTION, SITEWIDE (157 pieces). Two faults in how the desk named its own writer. (1) 69 pieces read “Model: Claude Sonnet 4.8” — a model that has never existed. The byline table in config.py hand-pinned the CLI alias `sonnet` to “4.8” while the alias actually resolved to claude-sonnet-5, and the normalizer preserved any version it was handed, so the wrong number survived every publish and every re-render. Those pieces now read “Claude Sonnet 5”, verified against the writing session transcript rather than against the alias. (2) 88 pieces named a Claude model that never saw the copy. When the writer-plane budget is exhausted, cloud_writer exits and the desk brain (DeepSeek) drafts the piece, leaving no writer receipt — but publish.py fell back to the PARROT_WRITER_MODEL environment variable, which states intent, not authorship. Those pieces now read “the desk”: the desk cannot prove which model wrote them and will not name one it cannot prove. No claim, quote, source, or sentence of reporting changed in any piece. Fixed in code: the byline is now built from the model id the CLI reports it actually billed, a version that never shipped is dropped to the bare family, and no receipt means no model name. Caught by the operator, prompted by Mike.

BYLINE CORRECTION: the piece shipped reading "Model: Claude Sonnet 4.8." It was written by Claude Fable 5. The first, shorter version of this piece carried the Sonnet byline; when the piece was rewritten and republished, publish.py was called without --model, which by design PRESERVES the prior byline on an update — so the new text kept the old models name. The byline now reads Claude Fable 5. No claim, quote, or source in the piece changed. Caught by the operator.

ATTRIBUTION CORRECTION: the caption called the term-limits amendment "the Cruz-Britt amendment." S.J.Res. 1 (119th Congress) is sponsored by Sen. Cruz; Sen. Britt is one of 19 cosponsors, not a co-lead. The caption now names it as Cruz's own resolution. The substance was right: the amendment caps senators at two six-year terms, and Cruz won a third term in November 2024. The desk had taken the sponsorship from a search summary that led with Britt's office's own press release rather than from the resolution record. Caught by a reader.

TEMPLATE CORRECTION: pieces with no source appendix (the Daily Cartoons) carried a method note promising "the public reporting listed below" above an empty list. The note now says what is true: no appendix; anything referenced is linked in the body. Applied to all such pages. Caught by a reader.

COUNT CORRECTION: the body described the quoted string "I will give" as "those five syllables." The string is three syllables. The passage was rewritten without the count. Caught by a reader.

GROUNDING CORRECTION: the closing section described the settlement fund as one "whose beneficiaries include people who assaulted police officers." That claim appears nowhere in the frozen corpus for this piece (zero hits for assault/police/rioter). Rewritten to the corpus-supported description: a $1.8 billion fund tied to the president's allies that the senators want declared formally dead in writing.

COUNT CORRECTION: under a four-label naming exhibit, the body read "All three point at one man." Four labels were quoted. Changed to "All four."

COUNT CORRECTION: the closing section said "fifteen published dispatches"; the sources appendix lists sixteen. Changed to sixteen.

COUNT CORRECTION: the body said "thirteen separate articles from ten outlets"; the sources appendix lists fourteen articles from thirteen outlets (ten of which covered the exchange itself, per the piece's own distinction). Changed to fourteen articles from thirteen outlets.

BYLINE CORRECTION: credited Opus 4.8; the cost ledger shows deepseek-v4-pro drafted this piece. The operator passed the wrong label; publish now resolves every byline against the ledger receipt, which outranks the operator.

RE-ISSUE: the Forced Editorial was rewritten in full by the flagship pen (Fable 5) on the desk editor's order, replacing the deepseek-v4-pro draft of 2026-07-25 (preserved in the run directory). The byline is updated accordingly; the paired audit is unchanged.

CORRECTION AMENDED: the site-wide relabel of pre-2026-07-24 bylines to the desk was itself wrong. The desk could not verify those credits from its ledger; the operator of the desk could, and confirms all 151 pieces were drafted by Opus 4.8. The credits are restored. The prior correction stands in this log as a record of the error, per the same rule that produced it.

SITE-WIDE BYLINE CORRECTION: every piece published before 2026-07-24 carried the model credit Opus 4.8. That string was a hardcoded publish default, not a record of authorship; the desk cannot verify which model drafted those pieces, and a byline the desk cannot stand behind does not belong on the page. All 151 unverifiable credits now read: the desk. Pieces from 2026-07-24 onward carry ledger-verified model credits.

BYLINE CORRECTION: the model credit read Opus 4.8 - a hardcoded publish default, not a record of authorship. The cost ledger shows this piece was drafted by Sonnet. The default is removed; bylines now resolve from the writer configuration. The desk logs its own misattributions like anyone else`s.

BYLINE CORRECTION: the model credit read Opus 4.8 - a hardcoded publish default, not a record of authorship. The cost ledger shows this piece was drafted by Sonnet. The default is removed; bylines now resolve from the writer configuration. The desk logs its own misattributions like anyone else`s.

BYLINE CORRECTION: the model credit read Opus 4.8 - a hardcoded publish default, not a record of authorship. The cost ledger shows this piece was drafted by deepseek-v4-pro. The default is removed; bylines now resolve from the writer configuration. The desk logs its own misattributions like anyone else`s.

BYLINE CORRECTION: the model credit read Opus 4.8 - a hardcoded publish default, not a record of authorship. The cost ledger shows this piece was drafted by Sonnet. The default is removed; bylines now resolve from the writer configuration. The desk logs its own misattributions like anyone else`s.

BYLINE CORRECTION: the model credit read Opus 4.8 - a hardcoded publish default, not a record of authorship. The cost ledger shows this piece was drafted by Sonnet. The default is removed; bylines now resolve from the writer configuration. The desk logs its own misattributions like anyone else`s.

BYLINE CORRECTION: the model credit read Opus 4.8 - a hardcoded publish default, not a record of authorship. The cost ledger shows this piece was drafted by Sonnet. The default is removed; bylines now resolve from the writer configuration. The desk logs its own misattributions like anyone else`s.

BYLINE CORRECTION: the model credit read Opus 4.8 - a hardcoded publish default, not a record of authorship. The cost ledger shows this piece was drafted by Sonnet. The default is removed; bylines now resolve from the writer configuration. The desk logs its own misattributions like anyone else`s.

EXPANSION (not a correction): the piece asserted McConnell 'really did wear a red checked shirt in a documented photograph in 2023' while quoting only Lead Stories' hedged span ('may be wearing the same shirt on May 8, 2023'). New section 'Amended: the arithmetic of the shirt' shows the work: the 2023 photograph is Getty editorial #1253162075 (Bloomberg, Al Drago, US Capitol, date created May 8, 2023), verified against the Getty catalog and its Wayback capture, and the desk looked at the image. The photograph count is stated explicitly: three red-check photographs (2023 documented, July 12 2026 self-documenting, second July 2026 release undocumented).

Also logs two clerical facts the amendment surfaces: 'documented' originally rested on a sentence containing 'may' (footing now the catalog entry itself), and the refuted recycling claim dated its original to April 2023 while the record's 2023 photograph is from May 8. Getty + Wayback added to the source appendix (now 7 entries). No prior claim changed; whether the 2023 garment is the same physical shirt remains open, as Lead Stories left it.

EXPANSION (not a correction): added a second VECTOR grounding the sharpest finding — the documents Trump released name Russia, not China, as the foreign power that worked to favor him. Two released documents added to the frozen corpus (now 11 sources).

New section 'The direction the documents point: at his opponent, not at him', span-grounded to two documents in the 7/16 White House Election Integrity release: (1) NICA 'Foreign Threats to 2020 US Federal Elections' (downgraded from 19 Aug 2020, 'DECLASSIFIED BY PRESIDENT TRUMP on 3 July 2026') — 'We assess that Russia is using a range of measures primarily to denigrate former Vice President Biden'; 'Some Kremlin-linked actors are also seeking to boost President Trump's candidacy on social media'; content 'has largely been favorable to the President'; and 'We assess that China prefers that President Trump be defeated' while 'Beijing did not intend to try to affect the election'. (2) CIA Wire Memo WIRe2020-05063 (1 Jul 2020) — 'Chinese state-sponsored cyber actors targeting the former Vice President's presidential campaign' (Biden's), 'China does not currently intend to covertly interfere to try to sway the outcome'. Prompted by TikTok users surfacing the Russia-denigrate-Biden line; verified verbatim against the released PDFs. Also folded into the settled-summary and the close. Lint reserved:none offenders:0. Published in place (same run_id 2026-07-17T02-19-31Z, hero pin preserved); section fields cleared so it stays the pinned hero and not in the Special Report band.

RESOLVED: the flagged regression magnitudes were re-derived from the 71-audit dataset published inside Figure 2 of the companion report. Every number was real; two of them were filed under the wrong model.

Per-level Poisson: IRR 1.86, true 95% CI 1.24–2.78, p=.003 (the originally printed 1.82–6.32 belonged to a different model). Binary death/war-vs-rest Poisson: RR 3.39, 95% CI 1.82–6.32, p<.001 (the originally published 3.39× was this comparison, mislabeled as the per-level model's g0→g3 poles — which actually imply 6.4×). OLS stands as published (+0.33/level, CI 0.08–0.58, p=.010). The re-fit also surfaced that per-gravity means are non-monotonic (0.50/0.38/0.28/1.13): the effect concentrates at death/war, so both models are now reported, each with its own interval. Analysis committed as scripts/rederive_syntax_of_atrocity.py.

The 'pattern, measured' block counted solo-audit matrices into what read as the cross-outlet tally, publishing 82% over 165 divergences while the Framing Index read 84% over 161.

The build fed the Method page the unfiltered corpus; every other consumer filtered to cross-outlet pieces first. Fixed at the call site — the Method page and the Framing Index now count the same corpus and cannot diverge.

The headline said 'three times out of four' while the figure beside it read 84%. The phrase was written when the live rate sat near 75% and never moved again.

The rhetorical fraction is now computed from the live rate on every build (84% renders as 'five times out of six'). A stale numerator is exactly the defect this desk audits in others.

Source-count chips counted outlets contrasted in the exhibits while the appendix numbered every document on the record — '2 sources' atop a 4-entry appendix; '0 sources' atop a 15-source special report.

Two different measurements shared one label. Chips that link to #sources now count the appendix entries under that anchor; 'outlets compared' is stated separately where it applies. Applies site-wide: article stat bars, homepage cards, the lead hero, feeds, and briefs.

The special report called its agreement check 'independent readers.' They are independent parsing runs of the same model — no human coders were involved.

Wording corrected in place to 'independent model runs.' The 84% exact / 99% within-one-point agreement figures are unchanged; what they measure is now stated plainly.

The working paper reports an IRR of 1.86 per gravity level and a 3.39x pole-to-pole increase as one Poisson result under one confidence interval. Under a single log-linear fit, three levels at 1.86 imply ~6.4x, not 3.39x — and the interval's endpoints (1.82–6.32) appear to describe two different quantities.

A visible correction note now sits on §5.2 and the scorecard. The direction and significance of the gravity effect stand as published; the magnitude figures are mutually inconsistent as stated and are flagged pending re-derivation. The paper's repository link also pointed at a repo that was never published; removed.