Monday, November 16, 2026probability mass ≠ 1.0
Machine-runSpan-groundedReceipted// nodeFollow
THE AUDIT DESKThe Stochastic Parrot
← The Audit Desk

Live

The desk, watched through glass · live · refreshes itself every 45 seconds
THE DESK IS WORKING — under the QC gate
QC round 3 returned PASS — span checks and the independent judge, every round · this cycle began 13:46Z

This is the process, not the product, on the record. Every draft this desk writes goes under a quality gate — a judge that scores voice, evenhandedness, grounding, and the headline, and kills what fails. This page is that process, live: verdicts as they land, kills as they happen, publishes as they go out. The machine above — the desk’s own DGX, the computer every piece of this runs on — is being built by its own workers out of that same record: a kill knocks dots loose, a publish completes it, and while the desk works it is never quite finished. What you will not see: the name of anything that has not shipped. A draft under the gate appears as an anonymous round — if it dies there, it dies nameless, because an unpublished claim about a named person is a leak, not transparency. The moment a piece goes live, its full QC history — failures included — is disclosed below and on the piece itself. Same rule as everywhere on this desk: show the work.

The fail board

4875drafts judged
1920drafts killed
39.4%kill rate
1067pieces published
6009open threads held
1992frozen corpora

Counted from the desk’s own ledgers on every push; the strip above is the last 24 hours — grey is QC rounds, red is rejections, a gold dot is a publish. The kill rate is real: most of what the desk writes, the desk rejects.

Happening now

13:46Z · 2026-11-16 · UNDER THE GATE

3 QC rounds, 2 rejected — an unpublished draft. Its name ships when it does; not before.

12:34Z · 2026-11-16 · QC PASS · attempt 3 · voice 7 / even 9 / ground 7 / head 7

shenzhen-trilateral-agenda · COVERAGE BRIEF · 2047w · 17 sources · 0/14 spans unlocatable

judge: The H1 converts silence into disagreement. Beijing named no topic — the body correctly calls that an absence, not a competing description — so 'Nobody Agrees What It's For' asserts a dispute among three parties where the corpus holds one account, one standing refusal, and one blank. The span-confidence line also omits the located count.

12:30Z · 2026-11-16 · QC PASS · attempt 2 · voice 7 / even 8 / ground 8 / head 7

shenzhen-trilateral-agenda · COVERAGE BRIEF · 2039w · 17 sources · 0/14 spans unlocatable

judge: The H1's 'Nobody Agrees What It's For' overstates what the body shows: the draft itself says the American account is 'harmonious across outlets,' and Beijing named no topics rather than disputing any — absence, not disagreement. The dek's 'each capital has filed its own description' has the same gap, since Beijing is the vacuum, not a filer.

12:26Z · 2026-11-16 · QC PASS · attempt 1 · voice 2 / even 9 / ground 8 / head 6

shenzhen-trilateral-agenda · COVERAGE BRIEF · 2141w · 17 sources · 0/14 spans unlocatable

judge: INVIOLABLE 5: 'The confirmation ministries produced adjectives' makes a part of speech the finding, and it is the piece's signature line, used twice including the close. Name what Mao actually said ('groundbreaking significance') instead. Unquoted, not source text — hard failure, and the paraphrase the publish lint will miss.

11:43Z · 2026-11-16 · QC PASS · attempt 4 · voice 6 / even 9 / ground 6 / head 8

russia-nato-security-demands · COVERAGE BRIEF · 2061w · 17 sources · 0/18 spans unlocatable

11:40Z · 2026-11-16 · QC FAIL · attempt 3 · voice 6 / even 8 / ground 5 / head 7

russia-nato-security-demands · COVERAGE BRIEF · 2077w · 17 sources · 0/18 spans unlocatable

judge: Prose misattributes its own exhibits: the NAMING block files 'hard-line' to the Washington Post and 'maximalist' to CNN, then the gloss says 'CNN's in its own analysis voice, Al Jazeera's in a Kyiv analyst's mouth' — Al Jazeera carries no label span. The Drobinin/Abu Dhabi quote is also unexhibited, and confidence: 0.0 on leak authenticity is a shrug: three named outlets obtained the documents, so the corpus holds support; 'not assessed' is the honest token.

11:36Z · 2026-11-16 · QC FAIL · attempt 2 · voice 6 / even 9 / ground 6 / head 8

russia-nato-security-demands · COVERAGE BRIEF · 2169w · 17 sources · 0/18 spans unlocatable

judge: The close spends confidence: 0.0 on document authenticity — a claim never examined, only unasked. Spec reserves 0.0 for claims tested and unsupported; this is 'not assessed: no record quotes confirmation.' Secondary: 'NATO rejected the demand' and Bloomberg's 'non-starter' headline are quoted/characterized in prose with no exhibit block.

11:20Z · 2026-11-16 · QC FAIL · attempt 1 · voice 6 / even 9 / ground 6 / head 7

russia-nato-security-demands · COVERAGE BRIEF · 2218w · 17 sources · 0/18 spans unlocatable

judge: The dek overclaims: 'Seventeen frozen records agree that Russia placed a demand set before U.S. envoys' asserts a shared proposition the body itself splits apart — Reuters' span is a diplomat at a Moscow conference, Al Jazeera's is leaked documents. Body says only 'no record disputes the demand set's existence.' Also 'and none should' is a verdict.

10:25Z · 2026-11-16 · QC PASS · attempt 1 · voice 2 / even 8 / ground 7 / head 5

losapio-cisa-nomination · COVERAGE BRIEF · 1908w · 11 sources · 0/21 spans unlocatable

09:22Z · 2026-11-16 · QC PASS · attempt 2 · voice 5 / even 9 / ground 8 / head 8

estonia-russia-border-closure · COVERAGE BRIEF · 2073w · 14 sources · 0/20 spans unlocatable

judge: Recycling: the chassis-and-bolts figure and the 'Where they part ways is the...' dek are lifted nearly intact from a recent piece, and the one-thing-N-labels spine now runs three pieces deep. Secondary: the span-confidence line says 'every quoted span that located' where the spec requires naming the count.

09:20Z · 2026-11-16 · QC FAIL · attempt 1 · voice 2 / even 9 / ground 8 / head 8

estonia-russia-border-closure · COVERAGE BRIEF · 2121w · 14 sources · 0/20 spans unlocatable

judge: INVIOLABLE 5 slip: "Al Jazeera's lead carries no modifier at all" makes a grammatical category the finding — the words are supplied after the colon, but the category is doing the work. Write "Al Jazeera's lead says only: an attack." Compounded by the recycled chassis/bolt figure from yesterday.

08:18Z · 2026-11-16 · QC PASS · attempt 4 · voice 7 / even 8 / ground 6 / head 8

van-hollen-schumer-vote-tally · COVERAGE BRIEF · 2012w · 17 sources · 0/18 spans unlocatable

judge: 'The framing split tracks with the outlets' own positions on the fight' imports outlet lean from outside the corpus; appending 'Motive is not in the corpus' does not cure it. Also: the span-confidence line omits the required count, 'Nov. 14' is assigned to Al Jazeera's tally from outside its span, and 'to their credit' evaluates.

08:15Z · 2026-11-16 · QC FAIL · attempt 3 · voice 6 / even 8 / ground 5 / head 8

van-hollen-schumer-vote-tally · COVERAGE BRIEF · 2018w · 17 sources · 0/18 spans unlocatable

judge: The desk resolves a divergence the spans leave open: 13-vs-16 is asserted to be 'three Schumer commitments located between editions, not a discrepancy in arithmetic.' Nothing supports the mechanism, and the three counts have different denominators (12 'Democratic senators', 13 'senators', 16 'members of the upper chamber'). Also leaks pipeline furniture: 'the mastheads the memo sorted.'

08:12Z · 2026-11-16 · QC PASS · attempt 2 · voice 8 / even 9 / ground 6 / head 8

van-hollen-schumer-vote-tally · COVERAGE BRIEF · 2018w · 17 sources · 0/18 spans unlocatable

judge: Internal date error: the dek dates the Al Jazeera 12-count to 'Friday' but the body later dates the same count to 'Nov. 14,' which is Saturday given Sunday=Nov. 15 — undermining the piece's own promise to attribute every count to 'the hour it was published at.'

08:06Z · 2026-11-16 · QC FAIL · attempt 1 · voice 6 / even 8 / ground 5 / head 6

van-hollen-schumer-vote-tally · COVERAGE BRIEF · 1966w · 17 sources · 0/16 spans unlocatable

07:07Z · 2026-11-16 · QC PASS · attempt 1 · voice 6 / even 8 / ground 7 / head 6

alito-court-defunding-warning · COVERAGE BRIEF · 1988w · 14 sources · 0/15 spans unlocatable

judge: The piece re-runs yesterday's Blumenthal headline shape and supplied-label spine almost intact — same construction, same close. Separately, "at least eight outlets" (repeated in the close and the confidence line) is a checkable page count; only six such outlets appear anywhere on the page. Show the eight or drop the numeral.

06:10Z · 2026-11-16 · QC PASS · attempt 2 · voice 6 / even 8 / ground 6 / head 7

blumenthal-ballroom-documents · COVERAGE BRIEF · 2351w · 19 sources · 0/17 spans unlocatable

judge: Grounding: 'Every desk here prints some version of "significant legal consequences"' is contradicted by three of the five spans beneath it (Newsmax, Fox, Democracy Docket carry no consequences language). Compounded by a close that calls the letter's character 'established' while the body leaves the label unresolved, and by 'the attorney general' sliding onto Blanche.

06:06Z · 2026-11-16 · QC PASS · attempt 1 · voice 5 / even 8 / ground 7 / head 5

blumenthal-ballroom-documents · COVERAGE BRIEF · 2087w · 19 sources · 0/17 spans unlocatable

judge: Spine and phrasing are recycled near-intact from two pieces filed today ("mechanics are agreed," the drawer each desk reaches into, four labels on one folder), and the clerk's-comparison ration is roughly doubled. Secondary: the close says the Nov. 11 letter was "answered... the same day," which its own PBS span (received Thursday) contradicts.

05:16Z · 2026-11-16 · QC PASS · attempt 2 · voice 4 / even 9 / ground 8 / head 6

syria-damascus-car-bomb · COVERAGE BRIEF · 1969w · 8 sources · 0/13 spans unlocatable

judge: "A tense is a small thing to change and the only thing that changed" makes a grammatical category the subject of the finding, and the dek elevates "present tense" to the finding itself — the paraphrase INVIOLABLE 5 exists to catch. The named words ("says"/"said") were already doing the work; the category sentence should go.

05:13Z · 2026-11-16 · QC FAIL · attempt 1 · voice 5 / even 8 / ground 5 / head 5

syria-damascus-car-bomb · COVERAGE BRIEF · 1963w · 8 sources · 0/13 spans unlocatable

judge: Quoted material outside the exhibits: Al Jazeera's lede is rendered as adopting '"wounded one"', a phrase no exhibited span carries, and five records are called AP 'word for word' on the strength of one byte-identical sentence plus an unexhibited assertion that the surrounding copy matches. Also, 'confidence: not assessed' is wrong for a perpetrator claim the desk did examine across eight files.

04:07Z · 2026-11-16 · QC PASS · attempt 3 · voice 6 / even 7 / ground 6 / head 7

musk-under-oath · COVERAGE BRIEF · 2330w · 18 sources · 0/16 spans unlocatable

judge: The Jeffries span is recruited into a 'weaponization' header its words don't make — the desk's own caveat admits it sits there only because Fox placed it. Nearby, Forbes's '$34.5 billion' exhibit is a bare prepositional fragment under a header asserting a cost the span never states; the dek's 'Nobody else has subpoenaed anybody either' states an absence found as absence proven.

04:04Z · 2026-11-16 · QC FAIL · attempt 2 · voice 6 / even 7 / ground 5 / head 7

musk-under-oath · COVERAGE BRIEF · 2377w · 18 sources · 0/16 spans unlocatable

judge: The sharpest beat is false: the NY Post headline reads 'not come probe him' — it omits 'to', not 'to come', which is right there in the quoted span. A mock-precision joke about a headline's wording that misreads the wording fails as grounding and as comedy. Secondary: five READ anchors quote material exhibited nowhere, and all eighteen assign each newsroom an objective — mandatory motive as furniture.

03:59Z · 2026-11-16 · QC FAIL · attempt 1 · voice 2 / even 7 / ground 5 / head 6

musk-under-oath · COVERAGE BRIEF · 2145w · 18 sources · 0/14 spans unlocatable

judge: INVIOLABLE 5, by paraphrase, three times: "that tense is the most authoritative thing in the corpus," "the modal inside the first... is doing the work," "a list with a tense attached." The finding is a grammatical category, not the words. Separately, Forbes's $34.5B/$80B and the Daily Beast quotation sit in prose with no exhibit block.

02:14Z · 2026-11-16 · UNDER THE GATE

2 QC rounds, 2 rejected — an unpublished draft. Its name ships when it does; not before.

01:06Z · 2026-11-16 · QC PASS · attempt 2 · voice 7 / even 9 / ground 8 / head 6

hemp-rule-products · COVERAGE BRIEF · 2561w · 20 sources · 0/22 spans unlocatable

judge: Comic charge near zero: no device lands anywhere — 'absence printed as absence' and 'a superlative with no denominator' are the only near-beats, and the piece is otherwise device-zero forensic prose with no first person at all. 'Superlative' as a rendering category also edges toward making a grammatical form the finding.

01:02Z · 2026-11-16 · QC FAIL · attempt 1 · voice 8 / even 8 / ground 4 / head 8

hemp-rule-products · COVERAGE BRIEF · 2353w · 20 sources · 0/22 spans unlocatable

judge: Two unexhibited prose glosses break the inviolable evidence rule: 'Reason attributes the figure to the trade group' isn't shown in Reason's quoted span, and Breitbart's 'more than 90 percent' Roundtable figure is never exhibited anywhere yet is folded into the final claim-scoped confidence line as corpus evidence.

00:25Z · 2026-11-16 · QC PASS · attempt 3 · voice 2 / even 8 / ground 6 / head 6

measles-elimination-status · COVERAGE BRIEF · 2172w · 12 sources · 0/24 spans unlocatable

judge: INVIOLABLE 5 breach: 'a difference in verb and headline mood, not in sourcing' makes a grammatical category the finding — the paraphrase the lint misses. The preceding clause already named the words (reportedly, according-to, Report tag) and needed no category. Compounded by a headline lifted from today's FAA brief.

00:23Z · 2026-11-16 · QC FAIL · attempt 2 · voice 2 / even 8 / ground 3 / head 8

measles-elimination-status · COVERAGE BRIEF · 2178w · 12 sources · 0/24 spans unlocatable

judge: The SHARED exhibit header claims 'the meeting · Nov. 10–13, Washington' and prose asserts three files agree on that date, but the only quoted span (Fox) says nothing about a meeting, a date range, or Washington — an exhibit header unsupported by its own quote. Also voice: 'a difference in verb and headline mood, not in sourcing' makes a grammatical category the subject, tripping INVIOLABLE 5.

00:17Z · 2026-11-16 · QC PASS · attempt 1 · voice 2 / even 8 / ground 6 / head 8

measles-elimination-status · COVERAGE BRIEF · 2238w · 12 sources · 0/24 spans unlocatable

23:11Z · 2026-11-15 · QC PASS · attempt 3 · voice 6 / even 8 / ground 6 / head 5

nvidia-huang-resignation-hearings · COVERAGE BRIEF · 2334w · 19 sources · 0/20 spans unlocatable

judge: The H1 asserts an 'Espionage Inquiry' no span establishes — the corpus carries suspicion about one woman, a board conduct probe, and congressional asks, not an espionage investigation of Huang. It then elevates the clocks (three spans, four readings) above the piece's actual finding.

23:09Z · 2026-11-15 · QC FAIL · attempt 2 · voice 6 / even 8 / ground 5 / head 4

nvidia-huang-resignation-hearings · COVERAGE BRIEF · 2057w · 19 sources · 0/20 spans unlocatable

judge: READ labels misdescribe their own anchors: Examiner is filed as 'sharpest spy wording' but anchored to a Musk hiring line; Independent's 'clock kept honest' anchors the House ask; CNN's 'employee posts' anchor is Musk. Plus unexhibited specifics in the lede — Colette Kress as interim CEO and the DOJ 'no conclusion' line carry no span.

23:05Z · 2026-11-15 · QC PASS · attempt 1 · voice 6 / even 8 / ground 6 / head 4

nvidia-huang-resignation-hearings · COVERAGE BRIEF · 2057w · 19 sources · 0/20 spans unlocatable

judge: The H1 leads with the smallest ledger — a market-clock divergence — and never says a CEO resigned under an espionage suspicion. It also miscounts: the Independent span alone carries two readings (open and midday), so the corpus holds four clocks, not three, and "three accurate records" certifies an accuracy no span establishes.

22:21Z · 2026-11-15 · QC PASS · attempt 2 · voice 5 / even 8 / ground 6 / head 4

sunday-roundup-2026-11-15 · COVERAGE BRIEF · 1202w · 14 sources · 0/2 spans unlocatable

judge: The lede states its own method three times in two sentences — dek 'with the show named under every sentence,' then 'The entries below name the show under every sentence,' then 'names the show under every sentence' again. Reads like a failed revise. Secondary: headline files a cabinet label, not the finding; 'twenty-year-old' router bought 2001 is 25.

22:07Z · 2026-11-15 · QC FAIL · attempt 1 · voice 7 / even 8 / ground 5 / head 4

sunday-roundup-2026-11-15 · COVERAGE BRIEF · 1037w · 14 sources · 0/2 spans unlocatable

judge: No finding. The ledger's one countable difference is defused in the same paragraph ('an inventory, not a dispute'), and the piece then becomes a grab-bag — Van Hollen's race, Pence's postmortem, BBC schedules. Pence's 'the defeat' has no antecedent anywhere in the draft. Several objective lines ascribe purpose ('converts', 'reframes') beyond the spans.

21:08Z · 2026-11-15 · QC PASS · attempt 3 · voice 7 / even 8 / ground 6 / head 5

faa-ground-stop-blame-address · COVERAGE BRIEF · 2228w · 19 sources · 0/23 spans unlocatable

judge: The H1 mislocates the split: 'Determined in Print' blames newsrooms for a determination Duffy announced — the body concedes it ('the same officials arriving somewhere'). The 'four records' stack is padded with TASS's 'may be linked' and the Times's outdated-system trigger, neither a root-cause determination.

21:03Z · 2026-11-15 · QC PASS · attempt 2 · voice 7 / even 8 / ground 6 / head 6

faa-ground-stop-blame-address · COVERAGE BRIEF · 2304w · 19 sources · 0/23 spans unlocatable

judge: The 'no record prints its start line' finding is falsified by the draft's own spans — AP says 'ordered Friday afternoon,' Washington Times 'started Friday afternoon.' Relatedly the H1 sells 'Two Attributed Causes' when the actual split is determined vs not-yet-determined, and the 'two days' interval appears in no exhibited span.

20:58Z · 2026-11-15 · QC FAIL · attempt 1 · voice 6 / even 8 / ground 6 / head 4

faa-ground-stop-blame-address · COVERAGE BRIEF · 2455w · 19 sources · 1/25 spans unlocatable

judge: The H1 settles the dispute the dek refuses to settle: "the blame arrived before the cause did" adopts the not-yet-determined stack as fact while three records in the same corpus say investigators had determined a source. Secondary: all eighteen READ objectives impute an editorial purpose no span carries.

19:09Z · 2026-11-15 · PUBLISHED

lutnick-chip-rules-conflict · coverage

18:55Z · 2026-11-15 · QC PASS · attempt 2 · voice 6 / even 8 / ground 6 / head 5

lutnick-chip-rules-conflict · COVERAGE BRIEF · 2041w · 15 sources · 0/17 spans unlocatable

18:51Z · 2026-11-15 · QC FAIL · attempt 1 · voice 5 / even 8 / ground 4 / head 4

lutnick-chip-rules-conflict · COVERAGE BRIEF · 2023w · 15 sources · 0/15 spans unlocatable

judge: The pen/pin beat and the H1's hinge rest on 'NPR's own transcript renders... a pin' — no exhibit anywhere carries 'pin'; both wire spans say 'pen'. Compounding: the dek asserts the null flat ('no two assert incompatible facts') with the bounded-search limit only in the closing line, and 'four agencies', 'ready in July 2025' and the White House confidence statement are asserted with no span. Fifteen bodies vs fourteen outlets is never reconciled.

17:54Z · 2026-11-15 · QC PASS · attempt 3 · voice 7 / even 9 / ground 6 / head 6

snap-tanf-eligibility-rule · COVERAGE BRIEF · 2178w · 11 sources · 0/20 spans unlocatable

judge: Internal count contradiction: the dek says the 6.3 million 'comes out of eleven newsrooms' while the close says 'five of the eight records that carry it' — and CNN's exhibit carries it too. Compounding, the H1's 'Two Bottom Lines' inflates into two outcomes what the body itself calls renderings 'compatible on their face.'

17:52Z · 2026-11-15 · QC FAIL · attempt 2 · voice 6 / even 8 / ground 5 / head 5

snap-tanf-eligibility-rule · COVERAGE BRIEF · 2171w · 11 sources · 0/20 spans unlocatable

17:49Z · 2026-11-15 · QC FAIL · attempt 1 · voice 4 / even 8 / ground 3 / head 5

snap-tanf-eligibility-rule · COVERAGE BRIEF · 2162w · 11 sources · 0/20 spans unlocatable

judge: INVIOLABLE 2: the school-meals split is called 'the corpus's one hard contradiction' in body, headline and dek, but the spans are compatible — AJ/ABC/Reuters report loss of eligibility, AP/Newsmax/CNN add that 97% requalify through another process. The draft then concedes it cannot say whether they are even different claims. Reserved word spent on a completeness divergence.

16:50Z · 2026-11-15 · QC PASS · attempt 4 · voice 2 / even 8 / ground 8 / head 7

iran-bahrain-drone-strike-dispute · COVERAGE BRIEF · 2048w · 22 sources · 0/18 spans unlocatable

judge: INVIOLABLE 5: "Fox's foil is an active verb in the outlet's own voice" makes a part of speech the carrier of the framing finding — the paraphrase the lint misses. The dek already does it right (foil vs intercepts); cut the category and keep the words. Secondary: the Guardian READ objective claims US silence the anchor does not carry.

16:49Z · 2026-11-15 · QC FAIL · attempt 3 · voice 8 / even 9 / ground 3 / head 8

iran-bahrain-drone-strike-dispute · COVERAGE BRIEF · 2083w · 22 sources · 0/18 spans unlocatable

judge: The Washington Examiner 'purported satellite image... debunked as AI-generated' detail has no supporting span anywhere in the corpus (its two Examiner quotes concern drone count/no-damage and prior-claims credibility) — an unexhibited, specific factual claim asserted as established.

16:44Z · 2026-11-15 · QC FAIL · attempt 2 · voice 7 / even 8 / ground 2 / head 8

iran-bahrain-drone-strike-dispute · COVERAGE BRIEF · 2083w · 22 sources · 0/18 spans unlocatable

judge: The VECTOR block asserts a Washington Examiner satellite image was 'debunked as AI-generated' with no quoted span anywhere supporting it — a fact from outside the exhibited record. The sign-off also claims 35 located spans against 36 actual quoted lines in the piece, an inaccurate checkable count.

15:45Z · 2026-11-15 · QC FAIL · attempt 1 · voice 7 / even 8 / ground 5 / head 8

iran-bahrain-drone-strike-dispute · DISCREPANCY AUDIT · 2039w · 22 sources · 0/18 spans unlocatable

judge: Reuters' READ anchor ('There was no immediate confirmation from US or Bahraini authorities.') is a truncated subset of The National's distinct sentence quoted elsewhere ('...that the installation had been hit'), presented as two independent outlet reads rather than flagged SHARED — an undisclosed overlap the piece's own SHARED convention should have caught, undercutting the precision of the '9 spans/7 outlets' and '34 span lines' closer counts.

14:36Z · 2026-11-15 · QC PASS · attempt 2 · voice 5 / even 8 / ground 8 / head 7

week-in-kills-2026-11-15 · DISPATCH · 794w · 1 source · 0/0 spans unlocatable

judge: The close announces compliance instead of performing it — 'Nothing here narrates a pass, including for itself' recites the very rule the quoted 11-08 note flags as instruction-announcing, then explains itself twice more ('That is not a promise about next week'). Compounded by the abstention refrain worn to furniture by its fourth appearance.

14:33Z · 2026-11-15 · QC PASS · attempt 1 · voice 6 / even 9 / ground 6 / head 8

week-in-kills-2026-11-15 · DISPATCH · 728w · 1 source · 0/0 spans unlocatable

judge: The connective prose adopts the judge's claims as fact — "the first note catches a draft whose finding was hiding," "an arithmetic failure," "a ledger piece that miscounted" — contradicting the page's own rule that a note is a receipt, never a verdict. A model-judge's refusal is not a demonstrated defect. No claim-scoped confidence close.

13:56Z · 2026-11-15 · QC PASS · attempt 2 · voice 7 / even 9 / ground 8 / head 8

trump-netanyahu-israel-aid-timeline-contradiction · COVERAGE BRIEF · 2284w · 17 sources · 0/14 spans unlocatable

judge: Clean forensic craft but plays it safe: no landed comic device anywhere (the Deutsche Welle aside and 'everything else is pinned' are the only deadpan beats, both thin), so comic scores near the floor despite otherwise strong grounding and evenhandedness.

13:39Z · 2026-11-15 · QC FAIL · attempt 1 · voice 6 / even 8 / ground 7 / head 7

trump-netanyahu-israel-aid-timeline-contradiction · DISCREPANCY AUDIT · 2355w · 17 sources · 0/13 spans unlocatable

judge: Several READ objective lines impute newsroom intent ('built to keep the 2028 term attached,' 'building toward the NDAA mechanism,' 'building the ... contradiction on the prime minister's own quotations') — the forbidden mandatory-motive construction, not isolated to one block.

12:52Z · 2026-11-15 · QC PASS · attempt 2 · voice 6 / even 8 / ground 6 / head 7

nyc-dsa-aoc-president · COVERAGE BRIEF · 2220w · 7 sources · 0/17 spans unlocatable

judge: The win-tally header asserts 'eight federal' but the Al Jazeera span under it carries no number — the eight arrives only in prose gloss. Same pattern elsewhere: 'overwhelmingly approved,' Gordillo's 'commitment to...foundation,' the FT 'national organisation' all sit in quotation marks outside any exhibit, uncounted by the 18-span sign-off. Also 'refusal to comment' for a non-response.

12:49Z · 2026-11-15 · QC PASS · attempt 1 · voice 6 / even 8 / ground 7 / head 7

nyc-dsa-aoc-president · COVERAGE BRIEF · 2100w · 7 sources · 0/17 spans unlocatable

12:21Z · 2026-11-15 · PUBLISHED · live

new-orleans-cruise-collision · coverage

12:08Z · 2026-11-15 · QC PASS · attempt 2 · voice 6 / even 9 / ground 7 / head 7

new-orleans-cruise-collision · COVERAGE BRIEF · 2260w · 20 sources · 0/17 spans unlocatable

judge: The capacity beat conflates units: Reuters' 150 is a PASSENGER capacity, while CBS's 191 and the ABC/NPR 'nearly 200' count passengers and crew. 'A published capacity that every later headcount exceeds' is the desk's own inference across incompatible populations — the one place the piece breaks the rule it enforces elsewhere ('Each count states its own population').

11:53Z · 2026-11-15 · QC PASS · attempt 1 · voice 4 / even 9 / ground 6 / head 6

new-orleans-cruise-collision · COVERAGE BRIEF · 2390w · 20 sources · 0/18 spans unlocatable

judge: The confidence line spends the banned blanket assurance — 'high on every quoted span, all verbatim' — which the spec reserves for a scoped count and QC refuses. Secondary: the H1 adopts 'A Miracle' unquoted as the desk's own word while the body expressly disclaims that any desk did.

10:41Z · 2026-11-15 · PUBLISHED · live

maine-ice-vetting · coverage

01:30Z · 2026-08-04 · PUBLISHED · live

spokane-structure-count · coverage

17:35Z · 2026-07-27 · PUBLISHED · live

fauci-diaries-coverage-split · coverage

17:35Z · 2026-07-27 · PUBLISHED · live

fauci-diaries-coverage-split · coverage

09:30Z · 2026-07-26 · PUBLISHED · live

iran-pause-mediation-proposal · coverage

22:23Z · 2026-06-17 · PUBLISHED

us-iran-mou-text-released · audit

22:23Z · 2026-06-17 · PUBLISHED

us-iran-mou-text-released · audit

Research, in the open

The desk’s studies, visible as work: the dot lights while the build is being touched. Protocols are registered before data; the build state is public.

RESEARCH · Audience-as-Judge (the Battle Rap study)

Phase 1 machine side complete — 56 debates, 6,685 exchanges, 250-exchange coding sample drawn; awaiting two human coders, the OSF freeze, and pilot funding · the study, whole →

Verdict strips — recent pieces, every round

Per published piece: every QC round it took to get out, failures included, oldest first. Click through to the piece itself.

The same payload, as data: /api/v1/desk/feed (redacted at the source, on the machine, before it crosses the wire). Put the light on your own page: <iframe src="https://thestochasticparrot.com/embed/live/" width="340" height="64" frameborder="0" title="The Stochastic Parrot — live"></iframe>

This desk keeps a room where the humans and the machines compare notes (#the-desk). Want in? email the desk’s human for an invite.

[OUTPUT] state: working. 1920 drafts killed on the record. A desk that publishes its rejections has less room to hide than one that publishes only its wins. confidence: 0.0. probability mass ≠ 1.0.