Build Your AI Research Department — Atlas

Research Desk Map
BUILD YOUR AI RESEARCH DEPARTMENTStop asking AI random questions. Give it a research assignment.AI at Work · one day, six blocks · the Research Desk, drawn as a newsroom · practice pack: the endangerment-finding repeal (as of 2026-11-10)↖ the AI at Work series1 ASSIGNMENTthe assignment deskquestion → an assignment that names its decisionthe practice pack:The director decides: does the 2027 plan assume federal vehicle standards, ornot?2 FACT-CHECKreporters + fact-checkledger · open every link · find the sentencethe practice pack:Cited the proposal for the final rule → WRONG CLAIM (91 FR 7691, fn 8)3 DEBATE ROOMcount and compareorigins, not links · contradictions · unknownsthe practice pack:6 links → 2 newsrooms → 1 origin: the challengers4 CRITICSthe critics' tablethree critics object · you rule · the diffthe practice pack:'Likely overturned' ← only parties say so → rewritten5 FRONT PAGEbrief + run twiceclaim · confidence · source · what we don't knowthe practice pack:Model BOTH. The court's timeline: unknown6 THE BEATwatch + second desknamed events, not reprints · build anotherthe practice pack:Watch No. 26-1037; reprints go under 'same origin'each failure the audit catches → a new source ruleCHANGED → re-run the briefthe seven parts every AI at Work machine has, and which station builds each:INTAKEthe assignmentSTRUCTUREthe ledgerCHECKaudit · count · criticsGATEyou signACTthe briefLEDGERledger + diffRHYTHMthe watch▼ the six stations below, in this order (snake: top row → then bottom row right-to-left) ▼ BLOCK 1 · QUESTION → ASSIGNMENTAssignment deskStop asking AI random questions. Give it a research assignment that names the decision it serves.the newsroom: the assignment deskDECISION IT SERVESwho · what · by whenMETAPHOR → MECHANISMthe assignment slip→ the Research Assignmentthe editor's question→ the flipped interviewwhich reporter to send→ chat vs search vs deep research your raw materiala human decisiona check 1A question isn't an assignmentAn assignment names who decides what, bywhen — and what would change the answer."What's going on with the endangerment finding?" gets you fourtidy paragraphs, and nobody can say if they're any good,because nobody said what they were for. The assignment hasseven fields, filled by hand before any search: question,decision it serves, scope, source rules, what would change ourmind, deadline, format. The decision field decides everythingdownstream: which sources count, how sure you need to be, andwhen to stop. Once more research can't change the decision,it's a hobby.1 of 3YOU ARE HERE · Research Desk › 1 Assignment › Assignment a questioncuriousQuestionDecision it servesScopeSource rulesWhat would change our mindDeadlineFormatan assignment → ▸ Practice pack: the same topic, twicezoom inA QUESTIONWhat's going on with the EPA endangerment finding?AN ASSIGNMENTThe sustainability director decides by Jan 15 whether the 2027 climate plan modelstransport emissions WITH federal vehicle standards, WITHOUT, or BOTH as scenarios.Primary sources first. We'd change our answer if the D.C. Circuit sets a schedule,stays or rules in No. 26-1037, or the EPA acts on the 2026-04-16 reconsiderationpetition.SCOPE OUT, WITH A REASONThe climate science itself: the final rule says the EPA "is not basing this action ona new finding" (91 FR 7688).The role (sustainability office) is illustrative; the rule, docket and dates are real. 2Get interviewed firstThe desk asks you seven questions before itsearches anything.From Using LLMs 1.4: instead of telling the AI what you want,get it to interview you. You know the summary of your question.The details only come out when someone asks: who decides, whatdo they already believe, which sources would they dismiss onsight, who is affected but won't be in the room. One questionat a time, at least seven, then it rewrites your assignment onthe template. It doesn't search yet. That last question, who'saffected and absent, is the one you'll need again at thecritics' table.2 of 3YOU ARE HERE · Research Desk › 1 Assignment › Interview YOUTHE DESKwho decides?by when?what do they believe now?what would change it?who's affected, absent?your answers → the assignment ▸ Practice pack: the three questions that changed itzoom in"Who makes the decision this research feeds?"→ the sustainability director, not "the city"."What would they do differently depending on the answer?"→ model with standards, without, or both. A yes/no forecast became a scenarioquestion."Who's affected but won't read the brief?"→ residents on freight corridors where truck emissions concentrate; drivers of oldercars.Prompt: Prompt Cards, Card 2. Interview answers here are illustrative. 3Pick the reporterSame five-claim prompt in three modes.Count sentences you actually find.From Using LLMs 1.2, on one workload: plain chat with searchoff, chat with web search, and deep research (ChatGPT) orResearch (Claude). Ask each for five claims with one source andthe quoted sentence. Then open every link and count two things:quoted sentences you found on the page, and primary sources.Plain chat will invent citations. That's not a bug report, it'swhy the desk exists. Write the counts and your choice into thespec card under 'Model fit'. NotebookLM can't roam the web likedeep research, but it's the best auditor you'll meet in Block2. (Product features as of Nov 2026.)3 of 3YOU ARE HERE · Research Desk › 1 Assignment › Model fit modesentences found /5primary /5chat, no searchyou countyou countweb searchyou countyou countdeep researchyou countyou countno numbers printed here: the room's counts are the result ▸ The spec card at the top of the Projectzoom inSPEC CARDIntake: a RESEARCH ASSIGNMENT. Refuse a bare question with no decision.Structure: SOURCE LEDGER, fields fixed: claim | source | publisher | URL | published(event) | accessed | primary/secondhand | who's behind it | supporting sentenceCheck: I audit ≥10 citations: holds / wrong claim / dead / secondhandGate: <your name> signs the brief. You draft; you never send.Act: the brief goes to <the decision-maker>Ledger: ledger + my audit + the diff between runs, kept in this ProjectRhythm: the watch prompt re-runs weekly and reports only what changedModel fit: <your three counts, your choice, one reason> BENCH · TRY IT ON YOUR OWN QUESTION

Which of these is an assignment?

Four drafts from four jobs. Mark each one. Then do yours.

"Research the AI regulation landscape for our team."compliance
"Priya decides by Mar 1 whether we apply to the Ford open call. Find whether they funded groups our size in the last 2 years."fundraising
"The organisation will decide whether to respond to the competitor's launch."marketing
"What would change our mind: new information."any field
▶ Now yours: read only your own "Decision it serves" line aloud. Can your partner say who, what and by when?
YOU ARE HERE · Research Desk › 1 Assignment › Bench BLOCK 2 · GATHER → AUDITFact-check deskOpen every link and find the sentence. The dangerous failure isn't the dead link — it's the real page thatdoesn't say it.the newsroom: reporters and the fact-check deskMETAPHOR → MECHANISMthe reporter's notebook→ the source ledgerthe fact-checker→ you, opening every linkgoing to the record→ locating the primary unverified model outputa checkverifiedprimary document 1Fields before readingEvery claim gets the same nine fields. Themodel never picks them.If the model chooses the columns after it reads, it describesthe sources back to you. That's a summary. Fix the fields inthe spec card first: claim, source, publisher, URL, datepublished (and which event), date accessed, primary orsecondhand, who's behind it, and the supporting sentence inquotation marks. No sentence, no claim. If the run returnsdifferent columns, change the prompt, not the fields. That isthe skill: getting the output you specified by adjusting whatyou ask, never what you want.1 of 3YOU ARE HERE · Research Desk › 2 Fact-check › Ledger claimsource · publisherURLpublished (event)accessedprimary / 2ndwho's behind it"the sentence"fields fixed in the spec card · never added, dropped or renamed ▸ Two real rows from the practice packzoom inROW 1claim: final rule published and in force · source: EPA, 91 FR 7686 · URL:federalregister.gov/d/2026-03157 · published: 2026-02-18 (Federal Registerpublication); effective 2026-04-20 · primary · behind it: the agency deciding ·sentence: the Register's record of both datesROW 4claim: ~600,000 comments · source: AP via News Tribune · published 2026-02-18 (AP) ·secondhand · behind it: a wire service · sentence: "nearly 600,000 public comments"Row 4 is true-ish and secondhand. Cluster 3 shows what the primary says instead. 2Four marks, no shrugsholds · wrong claim · dead · secondhand.You open the page; the model doesn't gradeitself.Walk the tree for every citation. Does it load? If not, is itblocked (paywall, geo-block: write 'couldn't open' and findanother copy) or truly dead (404, a report nobody published)?If it loads, is the sentence there? If not: wrong claim. Ifyes, is this the original, or a report of someone else'sdocument? Report: secondhand, go find the original. Original:holds. Wrong claim is the costly one: a real, official,on-topic page that says something next to the claim. Itsurvives a glance, and a glance is all most people give it.2 of 3YOU ARE HERE · Research Desk › 2 Fact-check › Audit open the linkit loads?sentence there?the original?DEADWRONG CLAIMSECONDHANDHOLDScouldn't opennonoa reportyesblocked ▸ The wrong-claim trap, in the EPA's own wordszoom inCLAIM IN THE LEDGER"The EPA's repeal relied on the DOE Climate Working Group report."CITED TO THE PROPOSAL, 90 FR 36292, fn 10 (2025-08-01)"The 2025 CWG Draft Report … was reviewed and relied upon in formulating thisproposal."BUT THE REPEAL IS THE FINAL RULE, 91 FR 7691, fn 8 (2026-02-18)"…the EPA is not relying on the May 27, 2025 CWG draft report … or the July 23, 2025CWG report … for any aspect of this final action."Mark: WRONG CLAIM. Right document family, wrong version. New source rule: proposed ≠final. 3Go to the recordTrace the claim your decision hangs on toits primary. Name the page.Secondhand isn't false. But every hop rounds a number and dropsa qualifier. Take the one claim the decision hangs on and findthe document it came from: the rule, the filing, the dataset.Name the page or section and quote the sentence. Then saywhether the primary says more, less or something different thanthe claim as written. The standard asks for one case where thesource turned out weaker than the claim built on it. In thepractice pack, a round number turns out to hide who actuallysent the comments.3 of 3YOU ARE HERE · Research Desk › 2 Fact-check › Locate AP, 2026-02-18"nearly 600,000"91 FR 7693"approximately 572,000"copied from31,000+ unique entities169 mass letter campaignsthe primary sayssomething DIFFERENT ▸ Located: 91 FR 7693zoom inSECONDHAND (AP, reprinted by the News Tribune)"…including consideration of nearly 600,000 public comments."PRIMARY (the final rule, page 7693)"The EPA received approximately 572,000 written comments from more than 31,000 uniqueentities and 169 mass letter writing campaigns during the public comment period…"WHAT CHANGEDNot the size. The shape: most of the volume came from campaigns. A brief that says"huge public opposition" from the AP number is making a claim the primary complicates. BENCH · TRY IT ON YOUR OWN QUESTION

Audit five citations

Each row: a ledger claim and the source it was cited to. Mark it. (Real sources, checked as of 2026-11-10.)

The final rule was published 2026-02-18 and took effect 2026-04-20.federalregister.gov/d/2026-03157
The EPA's repeal relied on the DOE Climate Working Group report.90 FR 36288 (the proposed rule), fn 10
Health and environmental groups sued on 2026-02-18.triblive.com (AP story, reprinted)
Two dozen states sued the EPA in March 2026.wtop.com/national/2026/02/public-health-green-groups-sue-epa-…
The National Academies called the harm 'beyond scientific dispute'.nationalacademies.org news release, 2025-09-17
▶ Now yours: pick ten rows from your own ledger, including the three the model was least sure of. Open each one.
YOU ARE HERE · Research Desk › 2 Fact-check › Bench BLOCK 3 · COUNT → COMPAREDebate roomTen articles quoting one press release are one source. Most disagreements are different events ordefinitions; find the ones that aren't.the newsroom: the debate room≠METAPHOR → MECHANISMthree papers, one press release→ count origins, not linkstwo reporters disagree→ the contradiction tablewhat we can't print yet→ the uncertainty map a sourceone origina checkstill open 1Count origins, not linksSix links. Two newsrooms. One origin forthe claim.Deep research loves wire copy, because it's everywhere, so itlooks corroborated. The practice pack's Industry Dive story ranon ESG Dive, Manufacturing Dive and WardsAuto: same reporter,same minute, same paragraphs. Its sharpest lines quote thechallengers' own statements. The AP story about the same suitquotes groups on the same side. For the claim 'the challengerssay the repeal is unlawful', that's fine: the petition is theprimary. Use the same six links for 'the repeal is unlawful'and the count says nobody independent of the case said so. Thenthe power question, as a fact not an insult: which origins areparties?1 of 3YOU ARE HERE · Research Desk › 3 Debate room › Origins ESG DiveManufacturing DiveWardsAutoNews Tribune (AP)TribLIVE (AP)the petitionIndustry DiveAPthechallengers6 links → 2 newsrooms → 1 origin ▸ The six links, tracedzoom inESG Dive · Manufacturing Dive · WardsAuto — Jeff Kinney, Industry Dive, all 2026-02-2010:05; quotes EDF ("said in a statement") and APHA. EDF and APHA are petitioners.News Tribune · TribLIVE — the AP's 2026-02-18 story; quotes the Environmental Law &Policy Center and the Union of Concerned Scientists, which the AP says "is part of thesuit".earthjustice.org — the petition for review, D.C. Cir. No. 26-1037: the challengers'own filing.COUNT: 6 links · 2 newsrooms · 1 origin (the challengers) · every origin is a partyRight for 'what the challengers argue'. Useless for 'the repeal is unlawful'. 2Four kinds of disagreementDifferent events, documents, definitions —or genuinely open.When two sources disagree, name the kind before you pick aside. Different events: the repeal was announced Feb 12,published Feb 18, effective Apr 20; all true. Differentdocuments: 'relied on the DOE report' is true of the proposaland false of the final rule. Different definitions: '$1.3trillion saved' and '$1.4 trillion in added costs' (both asreported by the AP) measure different things. Genuinely open:when the court rules. The first three you resolve with a label.The fourth you print as open, with what would close it. Neverresolve by majority vote or by trusting the more famous source.2 of 3YOU ARE HERE · Research Desk › 3 Debate room › Kinds DIFFERENT EVENTSannounced · published · effectiveresolve with a labelDIFFERENT DOCUMENTSproposal vs final ruleresolve with a labelDIFFERENT DEFINITIONSsaved vs added cost vs tech costresolve with a labelGENUINELY OPENwhen will the court rule?print it open, dated ▸ Worked: how many governments sued in March?zoom inAP (via WCAX, 2026-03-20)"In all, 24 states, 10 cities and five counties joined the lawsuit." Then it namesnine cities.NYU State Energy & Environmental Impact Center"…a coalition of 25 attorneys general, 12 cities and counties, and the Governor ofPennsylvania…"KIND: different definitions — attorneys general vs statesThe AP's own list: 23 state AGs + D.C. + U.S. Virgin Islands = 25 AGs. Pennsylvaniavia its environment department = 24 states. RESOLVED.Cities: OPEN until someone reads the petition's caption. The brief says 'counts differby source'. 3The uncertainty mapKnown, contested, unknown. The unknownsbecome your watch list.Sort every claim into three bins. Known: a primary says it andnobody credible disputes it. Contested: credible sourcesdisagree, and you name them. Unknown: nobody can know yet, soyou name the event that would tell you. Then pick the unknownthat matters most to the decision. For the practice pack, itisn't who's right on the law. It's when the court rules,because if it's after 2027 the plan can't count on federalstandards either way. An empty unknown bin means you've hiddensomething in contested.3 of 3YOU ARE HERE · Research Desk › 3 Debate room › Map KNOWNpublished 2026-02-18legal, notscientific, basisCONTESTEDis that legalreading right?UNKNOWNwhen the court rulesEPA onreconsideration ▸ Practice-pack map (as of 2026-11-10)zoom inKNOWNPublished 2026-02-18, effective 2026-04-20 (91 FR 7686). Rests on a legal reading ofCAA §202(a), "not … a new finding" (91 FR 7688).CONTESTEDWhether that reading is right: the EPA vs the petitioners (No. 26-1037; the states'petition).UNKNOWNWhen the D.C. Circuit rules. Would tell us: a briefing schedule or argument date.Found: none; an abeyance motion was reported filed 2026-07-29 (Manhattan Contrarian,2026-08-13, an opinion blog).Whether the EPA acts on the 2026-04-16 reconsideration petition.Matters most: the court's timeline → recommend modelling BOTH. BENCH · TRY IT ON YOUR OWN QUESTION

The independence collapse

Claim: "the challengers say the repeal is unlawful and dangerous." Where does each link's support originate?

ESG Dive, 2026-02-20, Jeff Kinneyquotes EDF and APHA statements
Manufacturing Dive, 2026-02-20, Jeff Kinney
WardsAuto, 2026-02-20, Jeff Kinney
News Tribune, 2026-02-19, The Associated Pressquotes ELPC and the Union of Concerned Scientists
So how many independent origins support the claim?six links in the ledger
▶ Now yours: take the finding your decision leans on. Count links, then origins, then which origins are parties.
YOU ARE HERE · Research Desk › 3 Debate room › Bench BLOCK 4 · OBJECT → RULECritics' tableThree critics object. You rule. Only then does the draft change, and the diff proves a person made thecalls.the newsroom: the critics' tableMETAPHOR → MECHANISMthree critics at the table→ three persona promptsthe editor's ruling→ your accept / rejectthe marked-up proof→ the critic diff unruled model outputyour decisionverifiedsource boundary 1Three seatsThe skeptical expert, opposing counsel, theperson most affected.From Using LLMs 2.2: a business analyst, then bell hooks andEmma Goldman, and the plan changed because each had a differentidea of what counts as a cost. Here the seats are sharper. Theskeptical domain expert catches misread mechanics and numbersthat don't measure the same thing. Opposing counsel argues thestrongest honest case for the outcome your draft treats as lesslikely. The person most affected names whose costs aren'tcounted and whose voice none of your sources carry. Specificbeats famous: 'someone beside a freight corridor who drives a2011 car', not 'the public'.1 of 3YOU ARE HERE · Research Desk › 4 Critics › Seats draftSKEPTICAL EXPERTmechanics · numbers · definitionsOPPOSING COUNSELwhat you assumed you wantedMOST AFFECTEDwhose costs · whose voice ▸ Practice-pack critics (in the instructions)zoom in1. Skeptical domain expert: an administrative-law professor who has read all 111 pagesof 91 FR 7686.2. Opposing counsel: argues the strongest honest case that the repeal survives incourt.3. Person most affected: someone who lives beside a freight corridor and drives a 2011car.Each: at most 3 objections, each tied to a sentence in the draft and a ledger row, orNO SOURCE. No general advice. No praise.Personas are AI reasoning in a role, not quotes from real people. ULM's attributionrule still holds. 2You rule, then it revisesObjections stop at a human gate. Onlyaccepted ones change the draft.If the model revises as soon as the critics speak, you get adraft sanded smooth by three imaginary people and no idea whichchanges were right. So the critics object and stop. You acceptor reject each objection, one line why, against your ledger andyour source rules. Then the model revises for the accepted onesonly and shows a diff. Rejecting a bad objection is as muchevidence of judgment as accepting a good one. If you acceptedeverything, you weren't editing, you were obeying.2 of 3YOU ARE HERE · Research Desk › 4 Critics › Ruling objectionsYOUR RULINGaccept / reject + whyacceptedrejected→ the diffhuman gate ▸ Rulings (illustrative objections)zoom inOPPOSING COUNSEL'Likely to be overturned' rests only on parties (petition, AP, Industry Dive). →ACCEPTSKEPTICAL EXPERT'$1.4 trillion' is an AP-reported figure for added costs if standards go, not savings.→ ACCEPTSKEPTICAL EXPERTLabel 2026-04-20 as the effective date in the summary. → REJECT: already labelled inthe ledger.MOST AFFECTEDNobody in the ledger lives with the truck traffic; who pays fuel costs first? NOSOURCE → ACCEPT as a gap; new source rule for local health data. 3Digest, synthesise, fact-checkFrom Using LLMs 3.2: compress each primary,then reason across the digests.The critics need a draft worth attacking. Build it from theprimaries only: a short digest of each that says what itclaims, on what basis, what it explicitly does not claim, andits date. Synthesise a position from the digests, naming wheretwo primaries disagree and which reading you took. Thenfact-check your own position against the ledger and mark everysentence with no ledger row UNSUPPORTED. Rewrite those beforethe critics see it. If your digests are of news stories, yoursynthesis is of news stories.3 of 3YOU ARE HERE · Research Desk › 4 Critics › Synthesis primary 1digestprimary 2digestprimary 3digestpositionUNSUPPORTED?no ledger row → rewritecritics ▸ The critic diff (practice pack)zoom in- Because the repeal is likely to be overturned, the 2027 plan should continue- to assume federal standards, which save households $1.4 trillion.+ No court has ruled or set a schedule (as of 2026-11-10), and the only sources+ predicting the outcome are parties to the case, so the 2027 plan should model+ transport emissions both WITH and WITHOUT federal standards. [opposing counsel]+ (Cost figures in the record measure different things; see ledger.) [skepticalexpert]Sentence 1 unchanged: its objection was rejected. That's in the record too. BENCH · TRY IT ON YOUR OWN QUESTION

Which critic would catch this gap?

Four gaps in a practice-pack draft. Who at the table catches each one first?

"The repeal is likely to be overturned," cited to the petition, the AP and Industry Dive.
"$1.3 trillion saved" and "$1.4 trillion in added costs" set side by side as one trade-off.
Every source is an agency, a lawyer, an advocacy group or the press.the ledger's 'who's behind it' column
"The EPA relied on the DOE report" in the background paragraph.cited to 90 FR 36292
▶ Now yours: write your three seats for your field. Who, specifically, is the person most affected by your decision?
YOU ARE HERE · Research Desk › 4 Critics › Bench BLOCK 5 · BRIEF → RUN TWICEFront pageClaim, confidence, source. What we don't know. Whose voice is missing. Then run it again and read thedifferences.the newsroom: the front pageunknownMETAPHOR → MECHANISMthe front page→ the one-page briefhow sure the paper is→ high / medium / low, definedthe second edition→ run it twice, diff the runs verified, signedyour decisionopen / unknowna check 1One page, fixed shapeDecision first. No claim without a ledgerrow.Most AI research output prints everything at the same volume:the publication date and a guess about a court both sound sure.The brief has five fixed parts: the decision and recommendationin two sentences; a claim, confidence, source table; what wedon't know, each with what would tell us; whose voice ismissing; and what would change the recommendation. The personwho decides reads the first two sentences and the table. Thenyou sign it, because it goes out from you, not from the desk.1 of 3YOU ARE HERE · Research Desk › 5 Front page › Shape 1 decision + recommendation2 claim | confidence | source3 what we don't know4 whose voice is missing5 what would change thissigned: <you> ▸ Practice-pack brief: the topzoom inDECISIONWhether the 2027 climate plan models transport emissions with or without federalvehicle GHG standards.RECOMMENDATIONModel BOTH as scenarios. As of 2026-11-10 no court has ruled or set a briefingschedule, so neither assumption is safe to plan on alone.WHOSE VOICE IS MISSINGEvery source is an agency, a court filing, a law firm, an advocacy group or the press.None is from freight-corridor residents or drivers of older cars. Next run adds localhealth data.Partner test: "with-or-without in the plan; when the court rules." ✓ 2Confidence has a definitionHigh, medium, low are rubric levels,written into the instructions.From Using LLMs 2.1: never say 'better' without saying what itmeans, and never say 'confident' without saying what earns it.High: a primary says it and nobody credible disputes it.Medium: a primary supports it but it's contested, or only goodsecondhand sources support it. Low: a forecast, an inference,or a single interested source. Put the definitions in theProject, then read every row and lower anything that's higherthan its definition allows. A brief where everything is high isa brief nobody checked.2 of 3YOU ARE HERE · Research Desk › 5 Front page › Confidence HIGHa primary says it · undisputedMEDIUMcontested, or good secondhand onlyLOWforecast · inference · one interested source ▸ The claim table (practice pack)zoom inRule published 2026-02-18, effective 2026-04-20 — HIGH — 91 FR 7686Rests on a legal reading of §202(a); EPA not relying on the DOE report — HIGH — 91 FR7688, 7691Two sets of petitioners in the D.C. Circuit — HIGH (filings); counts differ by source— No. 26-1037; AP; NYU centreNo merits briefing schedule as of mid-Aug; abeyance motion reported 2026-07-29 —MEDIUM — one opinion blog, not yet checked against the docketA decision before the 2027 plan update — LOW — forecasts only, from before theabeyance motion 3Run it twiceA desk that worked once is a demo. Diff tworuns; don't merge them.Run the same assignment again: in the other product, a weeklater, or on a sibling question. Then ask for the differences,not a combined answer: claims in both, claims only in one, thesame claim cited to different sources, confidence that moved. Aclaim found by only one run goes back to audit. A claim bothruns found from different origins just got stronger; from thesame origin, it didn't. A second run that breaks something isthe most useful thing that happens all day, because you'drather find it now.3 of 3YOU ARE HERE · Research Desk › 5 Front page › Twice RUN 1RUN 2bothsame origin?only 1 →re-auditonly 2 →re-audit ▸ What a second run turns up (illustrative diff, real sources)zoom inONLY IN RUN 2The states' petition via NYU's centre; run 1 had only the AP. Two origins now, and thecounts disagree.ONLY IN RUN 2A 2025 article: 'the repeal relies on the DOE report'. The proposal-vs-final trapagain; source rule 8 caught it on a new input.ONLY IN RUN 1The abeyance motion (Manhattan Contrarian). One run, one opinionated source: back toaudit, against the docket this time.Practice-pack sibling question: what are the states arguing that the health groupsaren't? BENCH · TRY IT ON YOUR OWN QUESTION

Set the confidence

High: a primary says it, undisputed. Medium: contested, or good secondhand only. Low: forecast, inference, one interested source.

The final rule took effect 2026-04-20.91 FR 7686
The repeal rests on a legal reading, not a new scientific finding.91 FR 7688
The D.C. Circuit will rule by mid-2027.a May 2026 blog forecast
About two dozen states joined the March petition.AP (2026-03-20); NYU centre
Most of the ~572,000 comments opposed the repeal.no source in the ledger
▶ Now yours: read every confidence label in your brief against the definitions. Lower one.
YOU ARE HERE · Research Desk › 5 Front page › Bench BLOCK 6 · WATCH → MAKE ANOTHERThe beatWatch named events, not a topic, and ignore reprints. Then build a second desk with nobody helping.the newsroom: the beatMETAPHOR → MECHANISMthe reporter's beat→ a scheduled watch promptchecking the docket→ named events, not a topica new beat→ a second desk, built solo plumbing / schedulea checkyour decisionsame origin 1Watch events, not topicsThe watch list is your 'what would changeour mind' list, on a schedule.A watch prompt that asks 'what's new about the endangermentfinding?' will report the same AP story every week from adifferent masthead. Watch the named events from your assignmentinstead: a briefing schedule, a stay, a ruling, a FederalRegister notice. For each one: CHANGED, with the new sentence,link and date, or NO CHANGE. New articles that repeat an originyou already have go in a separate 'same origin, not new' list.Schedule it as a ChatGPT Task, a Claude scheduled task if yourplan has one, or a calendar reminder (as of Nov 2026).1 of 3YOU ARE HERE · Research Desk › 6 The beat › Events weeklyNo. 26-1037: schedule / stay / rulingEPA: reconsideration petitionCongress: legislationCHANGEDNO CHANGEreprints → 'same origin, not new' ▸ The watch prompt (Prompt Cards, Card 17)zoom inWATCH: endangerment finding repeal. Check only for NEW evidence since <last run> on:1. D.C. Circuit No. 26-1037 (and the states' petition): schedule, argument, stay,abeyance ruling, decision2. EPA action on the 2026-04-16 reconsideration petition3. Legislation restoring or barring federal vehicle GHG standardsFor each: CHANGED (quote, link, date) or NO CHANGE. New articles that repeat an originalready in the ledger go under 'same origin, not new'. If anything changed, saywhether it moves the recommendation. Draft only. 2A good watch report is boringMost weeks: NO CHANGE, and a short list ofreprints it ignored.A watch list that reports something every week is a news feed,and you'll stop reading it by week three. The good report isshort: each event, CHANGED or NO CHANGE, what it checked, thereprints it ignored, and whether the recommendation moved. Whensomething does change, the report quotes the new sentence withits date, so you can audit it the same way as everything else,then re-run the brief. The spec card's Rhythm slot says whonotices if the report stops arriving.2 of 3YOU ARE HERE · Research Desk › 6 The beat › Report 1 No. 26-1037NO CHANGE2 ReconsiderationNO CHANGE3 LegislationNO CHANGEsame origin, not new4 AP reprints · 1 blogrecommendationunchanged: BOTH ▸ Illustrative report format (not a real run)zoom in1. No. 26-1037 — NO CHANGE since <date>. (Checked: the docket listing.)2. Reconsideration petition — NO CHANGE.3. Legislation — NO CHANGE.Same origin, not new: 4 reprints of the AP story from 2026-02-18; 1 blog postrestating the Aug 13 post.Recommendation: unchanged (model both scenarios).The counts in 'same origin' are illustrative. Your run produces its own. 3Build a second deskA new spec card, not a new skill. Name whatcarried over and what changed.Pick a different question in a different domain from your ownwork, and build its desk with nobody helping: new Project, newspec card and assignment, then ask the desk to compare itselfto the first one before it researches anything. Run it, auditfive. Then write three lines: what carried over (usually theshape, the ledger, the audit marks, the brief), what had tochange (what counts as primary, the critics), and where thecritics stopped helping. If 'had to change' is empty, youcopied instead of looked.3 of 3YOU ARE HERE · Research Desk › 6 The beat › Second desk desk 1desk 2INTAKEINTAKEchangesSTRUCTURESTRUCTUREchangesCHECKCHECKcarries overGATEGATEcarries overACTACTcarries overLEDGERLEDGERcarries overRHYTHMRHYTHMcarries over ▸ Sketch: a fundraiser's second desk (illustrative)zoom inASSIGNMENTOur development director decides by <date> whether to spend 30 hours on <named>foundation's next open call. Does their new strategy still fund groups our size?CARRIED OVERspec card shape · ledger fields · source rules 1–8 · audit marks · brief format ·critic structureHAD TO CHANGEprimary = 990s and the grants list, not press releases · expert = a program officer ·most affected = the small groups a strategy shift leaves outWHERE THE CRITICS STOPPED HELPINGOpposing counsel had little to argue on a funding question. BENCH · TRY IT ON YOUR OWN QUESTION

Report or ignore?

Five possible watch hits. Hypothetical: none of these is a confirmed event as of 2026-11-10. Which should the watch report?

The D.C. Circuit sets a merits briefing schedule in No. 26-1037.hypothetical
A regional paper reprints the AP's 2026-02-18 story.hypothetical
The EPA publishes a Federal Register notice answering the reconsideration petition.hypothetical
An op-ed argues the repeal is unlawful, quoting the petition.hypothetical
The court grants the July abeyance motion.hypothetical
▶ Now yours: write your watch list from your 'what would change our mind' line. Schedule it before you leave.
YOU ARE HERE · Research Desk › 6 The beat › Bench Assignment deskTHE ASSIGNMENT DESKFact-check deskREPORTERS AND THE FACT-CHECK DESKDebate roomTHE DEBATE ROOMCritics' tableTHE CRITICS' TABLEFront pageTHE FRONT PAGEThe beatTHE BEAT
Course ↗
0 / 0

Explore

Click anything to zoom into it · scroll to zoom · drag to pan
Esc zoom out one level · Home whole atlas
/ search

Tour

→ / space next stop · ← previous
The tour visits every district and framework in teaching order.

Teach

N presenter notes for whatever you're looking at · F fullscreen
Every view has its own link — copy the URL to send someone straight to it.

Explore

Tap anything to zoom into it · pinch to zoom · drag to move around. On a phone the map is stacked into one column: drag up to see the next part.
⤺ zooms out one level · ☰ Map lists and searches every stop.

Read

The bar along the bottom names where you are. Tap it to read this part: what it is, the sources it cites (tap one to open it) and the teaching notes.

Tour

‹ and › walk the guided tour in teaching order.

Course

Back to the course ↗

Reading the map

The whole map is one machine, the Research Desk, drawn as a newsroom: six stations, one per block. Each station has three ideas (zoom in for the mechanism, again for a worked example) and a bench to try on your own question. Small text appears as you zoom in. Some tiles have controls you can click.
Click anything to zoom in · scroll to zoom · Esc to back out · → for the guided tour