Live
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 A draft under the gate appears as an anonymous round — if it dies there, it dies nameless, because an unpublished claim about a named person is a leak, not transparency.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
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
1 QC round, 0 rejected — an unpublished draft. Its name ships when it does; not before.
iran-lego-videos-view-counts · coverage
iran-lego-videos-view-counts · COVERAGE BRIEF · 2212w · 13 sources · 0/8 spans unlocatable
judge: "I render no verdict on any of this; the corpus rendered one" reproduces almost verbatim the exact construction the spec names as forbidden ("never as your own verdict... 'I render no verdict; the AP did'"), and misapplies it to a hard_contradiction vector rather than the corpus_adjudicated case rule 3 defines it for.
iran-lego-videos-view-counts · DISCREPANCY AUDIT · 2215w · 13 sources · 0/8 spans unlocatable
judge: The 'reading is live' phrasing is recycled near-verbatim from the immediately preceding ADNOC piece; otherwise clean, well-grounded, dry comic execution throughout.
iran-lego-videos-view-counts · DISCREPANCY AUDIT · 2699w · 13/13 spans unlocatable
judge: Comic floor barely cleared: a few landed dry lines ('a number wearing the clothes of one') but long stretches of clean forensic prose with no beat, especially the six stacked READ exhibits near the close.
read-exhibits-outlet-audit · EDITORIAL · 2950w · 9 sources · 0/0 spans unlocatable
judge: Comic floor is thin: only two beats actually land (the reprogramming threat, the '269 newsrooms... I moved on' anti-climax) inside a dense, otherwise all-statistics self-audit — competent and rigorous but close to spreadsheet-with-feathers.
read-exhibits-outlet-audit · EDITORIAL · 2711w · 9 sources · 0/0 spans unlocatable
judge: READ exhibits cite anchors from the desk's own past pieces but carry no URLs, so the one invariant that normally lets a reader check a quote (real URL, re-checkable corpus) is absent for a piece built entirely on self-audit claims.
ai-vendor-terms-solo-builder · EDITORIAL · 3307w · 5/5 spans unlocatable
judge: Comic floor is thin for an editorial with this much license — a few landed beats (the $100 coincidence, the anti-climax on it) but the piece mostly stays in flat forensic-advisory mode rather than using the room the ordered-opinion exception grants.
ai-vendor-terms-solo-builder · EDITORIAL · 3289w · 5/5 spans unlocatable
judge: 'I am one person building a company. Can these vendors see what I am making' briefly adopts the founder's first person without marking it as quoted/hypothetical, blurring the machine narrator's own identity for a sentence.
ai-vendor-terms-solo-builder · EDITORIAL · 3105w · 5/5 spans unlocatable
judge: No protest opener or ordered-verdict statement appears anywhere, yet the piece sustains direct business advice for third-party readers (which tier to buy, concrete action items) — a verdict on outcomes, not on coverage/language/hedges/own-conditions. The KIND OVERRIDE exception requires the order be stated on the page; it never is.
adnoc-tankers-hormuz-drone-strike · coverage
adnoc-tankers-hormuz-drone-strike · COVERAGE BRIEF · 1916w · 12 sources · 0/6 spans unlocatable
judge: The seven READ blocks assign unsourced motive/intent ('the reader is assumed to be following Washington') to each outlet — speculation outside the spans, redundant with the ANGLE section, and it drifts the narrator toward the 'knowing media critic' the spec forbids.
adnoc-tankers-hormuz-drone-strike · COVERAGE BRIEF · 1916w · 12 sources · 0/6 spans unlocatable
judge: Cold open is a near-structural repeat of Piece 1's Hormuz vessel-count opening (same location, same short-declarative numeric-count shape); otherwise clean, well-grounded work.
trump-hormuz-us-territory-claim · DISCREPANCY AUDIT · 2170w · 13 sources · 0/7 spans unlocatable
judge: Twelve READ blocks imputing motive to every outlet (hedged with 'innocent'/'tentative' but still speculative) bloat the piece and dilute forensic economy; comic weight rests on one idiom-audit beat carrying a long, administrative middle.
trump-hormuz-us-territory-claim · audit
white-house-ballroom-scotus-appeal · COVERAGE BRIEF · 2093w · 13 sources · 0/15 spans unlocatable
judge: Bystander riff faintly echoes piece 1's schema/field-validation device without its originator-note payoff, a minor voice ding. Otherwise disciplined, well-grounded, evenhanded across five READ blocks; comic charge is real (the bystander bit lands) but thin, resting on essentially one beat in a dense forensic piece.
white-house-ballroom-scotus-appeal · COVERAGE BRIEF · 2126w · 13 sources · 1/15 spans unlocatable
judge: Solid, well-grounded, evenhanded coverage brief, but it spends two devices — the schema/field-validation riff and the 'no instrument' line — that were just used in the two immediately preceding pieces, reading as recycled rather than rotated.
white-house-ballroom-scotus-appeal · COVERAGE BRIEF · 2126w · 13 sources · 1/15 spans unlocatable
judge: The cold open's 'logged a verb' / 'a future tense that resolved on schedule' wordplay routes its finding through a grammatical category (verb, tense) rather than the words themselves — not the letter of the outlet-dispute pattern rule 5 targets, but close enough to the forbidden shape to invite reflexive edit before publish.
splc-beirich-fraud-arrest · coverage
splc-beirich-fraud-arrest · DISCREPANCY AUDIT · 2058w · 12 sources · 0/17 spans unlocatable
judge: Biggest problem: the arrest-day VECTOR is a near-clone of the harvard piece's ruling-day VECTOR — same N-to-1 split, same Tuesday outlier, same vote-count rhetoric, same confidence-0.0 landing, filed the same day. Two identical weekday-split devices back-to-back read as house formula, not fresh surprise.
splc-beirich-fraud-arrest · DISCREPANCY AUDIT · 2063w · 12 sources · 1/17 spans unlocatable
judge: The signature incapacity line and the cold-open shape are lifted nearly intact from the Harvard piece published the same day in this rotation — egregious recycling, not a faint echo, and it must be rewritten before this can run.
harvard-antisemitism-lawsuit-dismissed · coverage
harvard-antisemitism-lawsuit-dismissed · COVERAGE BRIEF · 2242w · 13 sources · 0/13 spans unlocatable
judge: Twelve near-identical READ blocks (anchor/objective/motive/'confidence: tentative'/innocent-clause, repeated almost verbatim) speculate on each outlet's editorial intent beyond the spans and bury the one real finding (Thursday vs Tuesday) under exhaustive, mechanical filler — the restraint failure the spec calls the default risk.
harvard-antisemitism-lawsuit-dismissed · COVERAGE BRIEF · 2217w · 13 sources · 0/13 spans unlocatable
judge: HARD FAILURE: 'no adjective riding along with either verb' makes a grammatical category (adjective, verb) the subject of the finding — exactly the forbidden construction rule 5 bans, not a paraphrase exemption since it isn't a quoted span.
harvard-antisemitism-lawsuit-dismissed · COVERAGE BRIEF · 2217w · 13 sources · 0/13 spans unlocatable
judge: The closing incapacity riff is lifted nearly intact from the mangione piece's construction, and the outlier-column observation over-explains itself ('I say so because it is true and because I would rather say it than have it inferred'), edging toward winking at its own transparency instead of just being transparent.
mangione-federal-guilty-plea · coverage
mangione-federal-guilty-plea · COVERAGE BRIEF · 2073w · 12 sources · 0/15 spans unlocatable
judge: Comic floor barely cleared: scattered dry wit ('four altitudes of certainty,' the friendly-reading reconciliation) but no single landed set piece; twelve back-to-back READ blocks risk reading as a checklist rather than a voiced audit.
mangione-federal-guilty-plea · COVERAGE BRIEF · 2065w · 12 sources · 1/15 spans unlocatable
judge: Cold open misattributes McDonald's (the prosecutor's) 'no deal' denial to 'the man who took the plea' (Mangione) — conflates who actually said the plea was 'the opposite of' a deal, a real grounding slip in paragraph one, compounded by 13 speculative per-outlet motive READs stretching past the spans.
zambia-election-2026-count-resumed · DELTA · 407w · 3 sources · 0/3 spans unlocatable
judge: The Mundubile-crime callback leans on the prior piece's established fact (his video win claim) without re-grounding it with a quoted span in this file, so the line's punch depends on reader memory rather than this corpus alone.
zambia-election-2026-count-resumed · DELTA · 407w · 3 sources · 0/3 spans unlocatable
judge: The 'once removed from the briefing room' vs 'in the room' explanation for who kept the hour is invented — no span places any outlet at the briefing, so the causal story is editorializing beyond the corpus.
zambia-election-2026-count-suspended · COVERAGE BRIEF · 1910w · 6 sources · 0/2 spans unlocatable
judge: The ledger/duplicate-flagging status-inversion line is structurally near-identical to the pipeline/run-folder device in the same day's Apple piece — egregious enough to cost voice points despite otherwise disciplined, evenhanded, well-grounded work.
zambia-election-2026-count-suspended · COVERAGE BRIEF · 1922w · 6 sources · 0/2 spans unlocatable
judge: The piece's two strongest beats (the no-correspondent/no-camera incapacity list, and the built-the-case-and-it-declined-to-appear search) are near-verbatim structural repeats of the immediately preceding Lincoln piece, not fresh devices.
zambia-election-2026-count-suspended · coverage
uss-abraham-lincoln-jump-reports · DISCREPANCY AUDIT · 2259w · 8 sources · 0/19 spans unlocatable
judge: Ten speculative READ blocks assign 'objective'/'motive' to each outlet's editorial intent — inference beyond the verbatim spans — and crowd out the actual discrepancy analysis, which is the stronger half of the piece.
uss-abraham-lincoln-jump-reports · DISCREPANCY AUDIT · 2253w · 8 sources · 2/8 spans unlocatable
judge: The math_error flag cites 'the AP's "more than 240"' without ever quoting AP verbatim anywhere earlier in the piece, the one figure introduced without the full citation apparatus every other quote gets.
dc-guard-one-year-tallies · DISCREPANCY AUDIT · 1575w · 8 sources · 0/9 spans unlocatable
judge: The rhetorical climax rests on two specific figures (5,148 summer troop peak; 1,655 open violent-crime cases) asserted as narrator fact with no accompanying quoted span or URL, breaking the stay-inside-the-corpus rule at the piece's most load-bearing moment.
trump-carriers-steam-catapult-return · COVERAGE BRIEF · 1852w · 6 sources · 0/14 spans unlocatable
judge: Comic floor barely met: mostly flat forensic audit with only a few landed dry lines (decimal-place quip, sortie-rate personification, receipt callback); five identical 'READ' blocks all closing 'confidence: tentative' read mechanical.
dc-guard-one-year-tallies · DISCREPANCY AUDIT · 1795w · 8 sources · 1/9 spans unlocatable
judge: Comic charge stays thin across three dense statistical vectors — competent forensic accounting with only a few rationed deadpan beats (ledger line, 'not found' close) landing; the piece earns its close but risks reading as a spreadsheet audit until the final NO CRIME section.
dc-guard-one-year-tallies · DISCREPANCY AUDIT · 1795w · 8 sources · 1/9 spans unlocatable
judge: Closing 'confidence: 0.0' misapplies the unresolved-close format to the NO CRIME claim, which the piece itself already corpus-adjudicated (the 1,655-case log) — the one dishonesty rule 3 forbids: performing 0.0 confidence over a question the corpus already answered.
apple-epic-scotus-stay-names · coverage
dc-guard-one-year-tallies · DISCREPANCY AUDIT · 1786w · 8 sources · 1/9 spans unlocatable
judge: Closes with agnostic 'confidence: 0.0' over the NO CRIME claim, but the corpus itself already adjudicates that claim (1,655 open cases in the same police log) — the piece even says so in voice ('I filed this one as it stands: not found') before undercutting it with the unresolved close, which INVIOLABLE rule 3 names as the one dishonesty the desk cannot afford.
trump-carriers-steam-catapult-return · COVERAGE BRIEF · 1852w · 6 sources · 0/14 spans unlocatable
judge: Six READ blocks each close on the identical 'confidence: tentative' boilerplate, flattening the piece's back half into a template; comic charge is present but thin (one 'fails to compile' beat, one deadpan aside) rather than landed twice.
sudan-blue-nile-double-fall · DISCREPANCY AUDIT · 2111w · 5 sources · 1/13 spans unlocatable
judge: Comic is nearly absent: competent forensic prose throughout with no landed device beyond a faint deadpan-literalism beat ('They describe two different armies'), which fails the floor even under a dry key that otherwise permits expecting little.
dc-guard-one-year-tallies · DISCREPANCY AUDIT · 1786w · 8 sources · 1/9 spans unlocatable
judge: Double recycling of near-verbatim reasoning devices from two separate recent pieces in one draft, plus a second self-epithet ('instrument' after 'machine') exceeding the max-1 ration.
trump-carriers-steam-catapult-return · COVERAGE BRIEF · 1864w · 6 sources · 5/14 spans unlocatable
judge: The ledger status-inversion beat is lifted nearly intact from today's Apple piece (pipeline-refuses-write vs ledger-flags-entry) — same rhetorical machine, different noun.
apple-epic-scotus-stay-names · COVERAGE BRIEF · 2468w · 7 sources · 0/13 spans unlocatable
judge: The piece re-analyzes the same handful of files three times over (NAMING+SEMANTIC, then ANGLE walkthroughs, then READ blocks with near-identical motive/innocent framing for Epoch Times, Gotechtor, Northeast Times) — bloated and redundant rather than tightly filed.
apple-epic-scotus-stay-names · COVERAGE BRIEF · 2468w · 7 sources · 0/13 spans unlocatable
judge: Comic charge is thin — only the six-label/pipeline-would-refuse-the-write bit actually lands; the rest is clean but flat forensic prose, and the seven near-identical READ cards (same 'objective/motive/innocent/confidence: tentative' template) read mechanically repetitive within the piece itself.
apple-epic-scotus-stay-names · COVERAGE BRIEF · 2468w · 7 sources · 0/13 spans unlocatable
judge: The NAMING block and headline both count five distinct words (reprieve/stay/extension/halt/breather), but the OUTPUT line asserts the language 'moved by six words' — an ungrounded number in a piece whose entire thesis is that newsrooms can't agree on counts.
us-iran-mou-deadline-centcom · audit
us-iran-mou-deadline-centcom · DISCREPANCY AUDIT · 2007w · 11 sources · 0/7 spans unlocatable
judge: Nine READ blocks plus two VECTORs and a NAMING/SEMANTIC block make the piece procedurally dense — sound grounding, but the comic and grammatical energy of the strongest bits (idiom audit, the two-date open) gets diluted by catalog-style entries in the back half.
us-iran-mou-deadline-centcom · DISCREPANCY AUDIT · 2017w · 11 sources · 0/7 spans unlocatable
judge: The reserved-word foreshadowing beat is recycled near-intact from the same-day trump-iran piece; otherwise disciplined, well-grounded, evenhanded work with landed comic beats (idiom audit, Allah diction joke).
us-iran-mou-deadline-centcom · DISCREPANCY AUDIT · 2022w · 11 sources · 0/7 spans unlocatable
judge: The line 'I render no verdict on which source is right' lands almost verbatim on the spec's own illustrative example phrase ('I render no verdict; the AP did'), reading as formula rather than earned restraint.
us-iran-mou-deadline-centcom · DISCREPANCY AUDIT · 2035w · 11 sources · 0/2 spans unlocatable
judge: Nine near-identical READ blocks each closing 'motive: tentative; innocent:' turn the back half into a checklist and dilute momentum after the strong idiom-audit beat.
trump-iran-economic-fury · coverage
trump-iran-economic-fury · COVERAGE BRIEF · 2434w · 8 sources · 0/9 spans unlocatable
judge: Clean and tightly grounded throughout; the Fox 'measures doing the seeing' aside skirts grammar-tally territory but stays on the literal meaning, not a part of speech, so it holds. Comic beats land dry but stay modest — no single big swing.
trump-iran-economic-fury · COVERAGE BRIEF · 2462w · 8 sources · 0/9 spans unlocatable
judge: Runs two failed-assertion forms in one piece (missing-unit on Bessent's claim, exhaustive-search on Iran's silence) — not a violation alone, but it dilutes the ration and the cold-open shape is a second straight recycling; otherwise clean, well-grounded, evenhanded audit with real comic beats (the database log, 'the measures doing the seeing').
trump-iran-economic-fury · COVERAGE BRIEF · 2475w · 8 sources · 1/9 spans unlocatable
judge: The Bessent-preposition passage frames its finding as a part of speech twice — 'I am sorry to dwell on a preposition' and 'three grammatically distinct objects' — the exact grammar-tally construction rule 5 forbids, despite the actual disputed words being quoted nearby. Inviolable hard failure.
zambia-election-2026-results-first-counts · DELTA · 438w · 12 sources · 0/0 spans unlocatable
judge: Comic charge is thin — a couple of dry clerk asides ('which desk decided about was close enough') but no fully landed device; the piece reads as clean forensic prose with seasoning rather than an actual joke.
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.
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.
FAIL1 FAIL2 PASS3
PASS1 PASS2
FAIL1 PASS2 PASS3
PASS1 PASS2
PASS1 PASS2 PASS3
FAIL1 PASS2
PASS1 PASS2 PASS3
FAIL1 PASS2
PASS1 PASS2
PASS1 PASS2
FAIL1 PASS2
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.