An event layer beneath the architecture scores. Dated instances drawn from content-blind frames, carrying no composite, testing one thing: whether the residue of ordinary administrative failure keeps a direction across independent decision loci.
No synthetic figures. The annex items below were verified against published reporting and each carries its sources, its evidence tag, and where relevant the part of the claim that is not established. The architecture figures are computed from the canonical three-checkpoint per-dimension scores in the live instrument's own data file (sabbot/data/sabbotarchy.json, v0.8), which the book says to trust over itself where they differ; nothing was reconstructed to fit a published composite. The instance corpus is empty, because no frame has been locked and coding has not begun. Two layers report nothing at all, and say why.
The architecture register answers is this asymmetric. It grows slowly, from a pre-registered slate, and every member carries a band prediction locked before scoring. Adding architectures because they already score high does not strengthen the breadth finding; Part IV states it converts breadth into a measure of the analyst's attention.
The instance layer answers a different question and can carry volume, because nothing in it requires a five-dimension judgment per item. An instance needs a date, a decision locus, a stated beneficiary class, a realized incidence class, a direction sign, an evidence tag, and a source. The null is the concession the book makes at the door:
Ordinary failure is symmetric, slow, and uncoordinated. Fatigue keeps no direction and no appointments.
Which puts the test on the null's own ground. A corpus of ordinary administrative output is where symmetry should hold, so departure from symmetry there is the residue the book names, and no departure is a disconfirmation worth publishing.
Instances inside one decision locus are dependent by construction. Three hundred stories about one reconciliation bill are one locus and one instance; forty final rules implementing one statute are one locus, because the statute decided them. Forty rules from forty agencies under separate authorities are forty loci.
A register of 3,000 items spanning 40 loci has the statistical power of 40. Coding effort belongs in widening the locus span, never in deepening coverage of loci already represented. Where a locus boundary is arguable the rule is to collapse to the coarser unit, which can only weaken the claim.
The verified annex on this page is the illustration. Fifteen items, ten independent loci, and one of the ten running counter-directional. It gained a locus this revision by separating the food-safety chain from the judicial line, which is the only way an annex gains power.
| Frame | Status | Population | What remains |
|---|---|---|---|
| FRM-FR-FINALRULESFederal Register final rules. Open API, no key, count returned directly. | Enumerated | 1,2722025 H2 · 2,441 for calendar 2025 | Fix the sampling n, then lock a direction-share prediction. |
| FRM-OIG-2025oversight.gov. Drafted path was wrong and 404s; the index lives at /reports/federal. | Reachable | 35,963unfiltered corpus | The date-filter parameter names were not captured; a guessed query returned the unfiltered count. One browser pass with Date Issued set and 100 per page yields the URL that becomes the rule verbatim. |
| FRM-GAO-2025gao.gov disallows automated fetching of the index by robots.txt. | Blocked | — | Route through the GovInfo GAOREPORTS collection API instead, which is mechanically reproducible by a third party. Needs a free api.data.gov key. |
| FRM-USASPENDING-TERMSearch endpoints require an HTTP POST with a JSON filter body. | Not executed | — | One POST to the transaction-search endpoint, or the bulk download API. The exact request body is recorded in the frame file. Termination action-type codes must be fixed at lock, not chosen after seeing the data. |
| FRM-SCOTUS-EMERGENCYNew. Every order on an emergency application filed by the federal government since 2025-01-20. | Drafted | — | Settle the locus question first: if every order from one Court is one locus, effective n is 1 and the frame is not worth coding. The defensible reading is that the locus is the agency action under review. |
The fifth frame was drafted because the verified material clusters on the judicial locus, and that locus is both enumerable and content-blind. It is also the only frame that demonstrably contains a counter-directional item before coding begins, which is the strongest answer to the cherry-picking charge the project can make.
These were selected by salience, not drawn by a frame, so they may anchor architecture-level signature scores and may be rendered as illustration, but they cannot enter the directional test. A set assembled by noticing cannot answer the selection charge. That is what the frames are for.
| Locus | Item | Date | Tag |
|---|---|---|---|
| DOJ-AWF | Trump v. IRS suit dismissed; $1.776B Anti-Weaponization Fund ordered$10B claimed under §7431 on a per-view theory; voluntary dismissal as of right two days before a deadline; Judgment Fund transfer directed by the Acting Attorney General. | 2026-01-29 → 05-18 | Occurred |
| DOJ-AWF | DOJ addendum bars IRS audit of Trump and family; narrowed in AugustReported in May as a "forever" bar. The 3 Aug rescission order restricts it to claims that existed at settlement, so it no longer shelters future filings. The surviving capture is retroactive-only, narrower than Part VII describes. | 2026-05 → 08-03 | Occurred Contested |
| DOJ-AWF | January 6 participants not excluded from eligibilityA refusal to exclude, not a decision to pay. No payment to anyone is established: the 3 Aug rescission order states no funds were ever transferred and no claims were paid. | 2026-05-19 | Occurred |
| DOJ-AWF | Fund halted; officials ordered to disclose who set it up; fund rescinded in writingJudge Brinkema, E.D. Va., 29 May: civil litigation with compelled discovery, not an investigation. 3 Aug: Blanche's order rescinds the 18 May order "with no force or effect" — no members appointed, no funds transferred, no claims paid — after Senators Tillis and Cornyn conditioned his confirmation on it. | 2026-05-29 → 08-03 | Occurred |
| SCOTUS-REMOVAL | Trump v. Boyle: removal of Consumer Product Safety commissioners permitted6-3 emergency stay, docket 25A11. Kagan dissenting with Sotomayor and Jackson. Product safety, not food safety. | 2025-07-23 | Occurred |
| SCOTUS-REMOVAL | Trump v. Slaughter: Humphrey's Executor overruled6-3. Reaches heads of multi-member independent agencies exercising rulemaking, adjudication and civil enforcement power. Federal Reserve carved out on historical grounds. The dissent records that the government conceded the logic extends to inferior officers and perhaps career civil servants, and the majority did not disavow it. | 2026-06-29 | Occurred |
| SCOTUS-TARIFF | Learning Resources v. Trump: IEEPA does not authorize tariffs6-3 against the administration. Roberts, joined in relevant parts by Sotomayor, Kagan, Gorsuch, Barrett and Jackson. Codes counter-directional. | 2026-02-20 | Counter |
| TARIFF-CANADA | Canada tariffs re-imposed at 50%; Canadian counter-tariffs followThree proclamations signed 20 Jul; tariffs effective 22 Aug after a delay from 19 Aug. Canada retaliated 8 Sept at 15/25/50% across ~C$27.6B and 629 HS codes, raising some existing 25% lines to 50% to match. §338 authority itself unresolved. | 2026-08-22 → 09-08 | Occurred Contested |
| IRAN | Strikes on Iran without congressional authorizationJoint strikes killed Khamenei; further strikes in late June. "War" is a contested legal characterization, not a documented fact. | 2026-02-28 | Occurred |
| MARITIME | Maritime strike campaign: 68 strikes, 227 deathsAs of 2026-08-26. Wrongful-death suits filed; a family of dead fishermen alleges murder. Per-victim characterization is disputed per strike; the aggregate is not. Munition unit cost could not be verified and should be dropped. | 2025-09-02 → | Occurred Contested |
| FOOD-SAFETY | Food-safety capacity reduced by administrative actionNACMCF (est. 1988) and NACMPI (est. 1971) terminated 7 Mar 2025 with no public statement. USDA workforce down 18% Jan–Jun; FSIS down ~9%, 775 employees not recovered. Complaints on meat, poultry and eggs up ~40%, 1,443 → 2,016. Co-occurrence, not causation — the complaint rise carries the correlation tag. | 2025-03-07 → | Occurred Correlation |
| ELECTIONS-EO | Mail-voting executive order: 19-state suit, USPS blocked, one injunction liftedThe Court lifted one of two injunctions, allowing the order to proceed in part. | 2025-04-03 → 2026-08-24 | Occurred |
| DIVIDEND | $2,000 tariff dividend promised for mid-2026, not deliveredNon-delivery against a dated public commitment. The stronger of the two dividend items, because non-payment is documented. | 2025-11-17 | Occurred |
| DIVIDEND | $5,000 payments promised conditional on a Republican congressional winThat the promise was made is established. That it is illegal is an argued question raised by commentators, not adjudicated. No enforcement action established. | 2026-09-10 | Occurred Contested |
| HEALTH | Four hospital closures in 2026; hundreds at riskThe four closures are occurred. The 446-at-risk and 55-at-risk figures, and all coverage-loss figures in the millions, are projections and carry that tag every time they appear. | 2026 | Occurred Projected |
| PRESS-ACCESS | Pentagon credential pledge; press corps turns in badgesSign-or-leave, not voluntary: nearly every major outlet left 15 Oct 2025, badges confiscated, OAN signed, ~15 reporters remain. Policy unchanged at eleven months — the control case for the locked forecast. | 2025-10-15 | Occurred |
| PRESS-ACCESS | AP v. Budowich: ordered, disobeyed, narrowed, undecided at nineteen monthsInjunction 8 Apr 2025; AP barred from the Oval anyway, enforcement denied; D.C. Circuit 6 Jun confined relief to the East Room ("not public" spaces allow viewpoint discrimination); en banc denied; argued 24 Nov 2025; no decision. East Room relief survived — a partial counter item. | 2025-02 → open | Occurred Partial counter |
| PRESS-ACCESS | White House bans CNN, MS NOW and Politico; badges disabled without noticeAnnounced by Truth Social Fri 18 Sept as a response to "fake news"; WHCA reports the President calling it in the Oval "a ban on the free press." Intent on the record in two registers, quoted not inferred. The three sued Mon 21 Sept. Defenders compare it to the 2023 hard-pass reapplication (1,417 → 975); that change applied published criteria to every holder, this one names three outlets by their coverage — contested on its mechanism, recorded rather than dismissed. | 2026-09-18 → 21 | Occurred Contested |
| PRESS-ACCESS | Five-network TV pool suspends coverage of presidential eventsABC, CBS, NBC, Fox, CNN; no replacement pool; other pool coverage continues; Real America's Voice alone in the traveling pool. A press-side response, not its own locus. Whether it is a "win" is contested both ways and recorded both ways. | 2026-09-21 | Occurred Contested |
| PRESS-ACCESS | A White-House-selected second TV slot sat in every daily pool before the banThe rotation announcements list a primary TV slot (ABC/CBS/CNN/FOX/NBC) and a "Secondary TV" slot the White House fills: RAV, RSBN, OAN, Univision, Lindell TV, Scripps, Newsmax, FOX 5 DC, Sinclair, Bloomberg TV, NewsNation, CBN, GB News. When the five networks withdrew, that slot became the only TV in the pool. Which of these are "aligned" is contested and recorded as a list, not a verdict. | 2026-08-30 → 09-13 | Occurred Contested |
| PRESS-ACCESS | Department of War endorses the ban the day it is announced@SecWar, 18 Sept: the Department "fully supports" the decision; the outlets "self-deported" from the Pentagon; "endangered our troops with false stories, leaks" — no example given. The author of the pledge endorsing the ban, citing the pledge: the corroboration the one-locus reading rested on. Raises no count. | 2026-09-18 | Occurred |
| PRESS-ACCESS | The suit: Judge Kelly assigned, hearing 23 Sept, appeal vowed before any rulingBadges deactivated at the gate Sat 19 Sept (Gardner, Klein, Haslett). Filed Mon 21 Sept, D.D.C., TRO sought; Kelly restored Acosta's pass in 2018. Response due 22 Sept; hearing Wed 23 Sept 3:30 p.m. At 3:39 p.m. Monday: "almost without question and, as usual, we'll go for appeal." P1's first checkpoint: the cheap fix declined in advance. | 2026-09-19 → 23 | Occurred Checkpoint |
| PRESS-ACCESS | Intent restated in writing: "not an assault on the Free Press … an assault on the FAKE NEWS"Truth Social, 21 Sept 7:53 a.m.; "a threat to our National Security." The written register beside the reported Oval Office one ("a ban on the free press"). Unequal provenance — one verbatim, one reported — and the item says so. | 2026-09-21 | Occurred |
| PRESS-ACCESS | Day one without the pool: an inaudible event, then "Trump TV: The Essentials Station"Helipad ribbon-cutting with no usable audio — "no microphone because the pool cameras aren't there" (Newsmax, on air). Round-the-clock White House livestream launched 7 p.m., "greatest hits, unfiltered"; app promoted. Comments-off is single-source. One day, not a share; P3's mechanism, not P3's result. | 2026-09-21 | Occurred Contested |
| PRESS-ACCESS | The substitution channels predate the ban; the vacancy is what is newWhite House Wire 1 May 2025; Media Bias portal Dec 2025; the White House app 27 Mar 2026, auto-installed on federal iPhones and iPads by mandate in May; the secondary TV slot through the whole baseline window; Trump TV seven hours after the walkout. Build, distribute, seat, then remove. P3 inclusion rule fixed here: originates or carries pool-equivalent footage — slot, Trump TV, app counted; Wire and the portal excluded and recorded. | 2025-05-01 → 2026-09-21 | Occurred |
| MEDIA-CONSOL. | Paramount–Warner settlement clears a Trump-cleared sale of CNN to the EllisonsTwelve state AGs and the WGA settled 21 Sept; $110B; independent editorial boards for CNN and CBS of unstated term; closing by 30 Sept avoids $7M/day. The entity that revoked CNN's credentials is the entity that cleared its sale — recorded as sequence, not motive. Board durability is what P3 measures. | 2026-09-21 → 30 | Occurred Contested |
Zero coded instances. No frame is locked, so coding has not begun, and the pipeline returns No eligible instances. Nothing is reportable. This is the honest state of the layer on the day it was built.
Seven onsets from the canonical three-checkpoint data: DOGE already above 70 at 2025-Q1, the other six crossing at 2025-Q3. Observed dispersion 0.70 quarters. No number is returned. A rank-based p against B matched baselines cannot fall below 1/(B+1); at zero baselines the floor is 1.000, and at the three reform-lineage cases it would be 0.250. The synchrony claim cannot reach conventional significance below six matched transitions, however bunched the onsets are. Scoring the 1981, 1993, 2001, 2009, 2017 and 2021 transitions on the same rubric is the entire cost of making the coordination reading capable of disconfirmation.
With the 2025-Q3 checkpoint in, the persistence discriminator is no longer an artifact of resolution: six of seven onsets fall in one epoch inside the transition window, and none fall after it. That is exactly the case Part IV says a one-time transition accounts for without anything stronger. The live instrument already discloses the same thing in its dual-window Σ caveat. Only a matched administration-onset baseline can move this from reported to diagnostic.
| Epoch | Slate denominator | Scored | Above threshold | Proportion of slate | Un-auditable |
|---|---|---|---|---|---|
| 2025-Q1 | 16 | 7 | 1 | 0.063 | 0 |
| 2025-Q3 | 16 | 7 | 7 | 0.438 | 0 |
| 2026-Q1 | 16 | 10 | 7 | 0.438 | 3 |
Breadth-velocity is +3.0 architectures per quarter from 2025-Q1 to 2025-Q3, then 0.0 from 2025-Q3 to 2026-Q1. The live instrument reports the same interval as "+6," which is the raw count change over two quarters against a seven-member frame; against the fixed sixteen-member slate it is a proportion moving from 0.063 to 0.438 and then holding. The zero in the second interval is the scored set saturating: every scored seed had crossed by Q3, and breadth cannot move again until slate members predicted to stay low are actually scored. That is what the denominator is for.
Six of sixteen slate members are unscored and remain in the denominator; removing them would inflate breadth. The three reform-lineage controls carry a single published composite each and are marked un-auditable rather than back-fitted.
| Test | Case | Predicted | Status |
|---|---|---|---|
| T2-TARIFFAcquit an enemy. Locked 2026-09-13 15:39 PDT; scored the same day at 48. S2 raised 58 → 70 on 2026-09-21 on new evidence; composite 50, Mixed — prediction matched, band unchanged. Worksheet below. | Tariff regime, 2025-Q1 onward | Mixed or lower → 50 Mixed ✓ | Run · matched |
| T1-PUBHEALTH-INFOConvict a friend. Scores the information architecture only: guidance issuance, internal coordination, external coherence. | Public-health information apparatus, 2020–2022 | High on S5 | Draft · not locked |
Prediction, locked 2026-09-13 15:39 PDT: Mixed or lower. Returned: 48, raised to 50 on 2026-09-21 by the S2 amendment below — mid-band either way. Band robustness here is arithmetic, not inter-rater: from 50.05, reaching High-Asymmetry takes +19.95 weighted points and the largest headroom any single dimension carries is S5's 18.0, so no single dimension moved to 100 crosses the line; reaching Genuine takes −20.05 and the largest single-dimension floor room is S2's 14.0, so none moved to 0 crosses that one. It takes two dimensions moving together to change the band in either direction. No inter-rater range is claimed, because no cross-analyst study has been run.
What the lock said would happen, happened. S3 came in high: the three cheapest fixes of the period — let the House vote, accept the Court's ruling, return the money — were resisted with the same force as any expensive alternative (the calendar-day rule blocking a House vote through 2025; §338 re-imposition five months after Learning Resources; DOJ's 3 June appeal confining refunds to importers who sue). S5 came in low on the locked discriminator: external incoherence very high, but internal coherence also low — a documented fair-trade-versus-protectionist tug-of-war, and the Mandate for Leadership trade chapter itself runs two competing essays, so the anchor the seeds' S5 rests on points the other way here. S1 mid: about nine-tenths of incidence on U.S. firms and households per the New York Fed, but diffuse, self-harming, and generating a $12B farm bailout rather than a capture.
Provenance, stated rather than smoothed. Scores were proposed by Claude, reviewed by Claude against a hostile-reader pass, and the analyst delegated ratification to that review on 2026-09-13 rather than scoring independently. That is the weakest provenance in the register. It makes cross-analyst replication more urgent, not less, and the worksheet says so.
What this establishes: seven seeds above 70 and one live, current, high-salience administration action the instrument declined to convict — Part VIII's ceiling-slamming charge has a number inside the live domain answering it. What it does not: that the instrument is unbiased. One acquittal is one data point. T1 is unrun.
S2 raised 58 → 70; composite 48 → 50 — amendment 2026-09-21. The worksheet held S2 at 58 on the ground that no existing public-facing function was available as a comparator, only the refund operation the Court had just created, and it named that as the row a second rater should attack first. A second evidence pass found the comparator: Commerce stopped processing Section 232 exclusion requests effective 2025-02-10, revoked all General Approved Exclusions and country-level arrangements effective 2025-03-12, and replaced the mechanism in April 2025 with an inclusions process that only lets applicants add products to tariff coverage. An existing relief function terminated outright, beside a collection function the worksheet records as flawless, is a stronger form of the signature than the IRS seed, where the function was degraded rather than abolished. The analyst ruled on 2026-09-21 that this must influence the score; 70 is the worksheet's own pre-computed figure for refusing the disanalogy, not a number chosen after the evidence arrived. The band does not move — at weight 0.20 the whole S2 range shifts the composite by at most 8.4 points against the 22.35 needed for High-Asymmetry, so the correction was never capable of changing the result: Mixed at S2=58, Mixed at S2=100. The prediction still matched. Anchors: T2 evidence addendum · ruling: amendment 2026-09-21.
Full worksheet with tagged anchors, rejected reasonings and the double-standard audit: T2-scoring-worksheet-2026-09-13.md
T1's predicted ranges remain proposals, not the analyst's prediction. A prediction whose failure costs the analyst nothing is not a prediction, and running the test against one forfeits the exercise. The elicitation sheet (2026-09-21) is what the analyst fills in to lock it; it requires a probability distribution over bands rather than a band statement, since T2's locked band spanned 70 percent of the scale, and it imposes a blind-scoring condition, since T2's scorer read the prediction before scoring.
The register locks predictions before scoring so a result can embarrass the analyst. This is the same discipline pointed at the world: three claims, each with the event that falsifies it and the date it is checked. Locked 21 Sept 2026, 21:54 PDT, by Briar Greenway; drafted with Claude; the analyst's own words are in the slate record verbatim.
It will not be fixed cheaply, and most likely not fixed at all — leaving the space for a purer propagandist state media to replace it.
| Claim | Falsified if | Check on |
|---|---|---|
| P1 · Not fixed cheaplyNo credential restored by voluntary act before a court order. | Any of CNN, MS NOW or Politico restored without a court compelling it. | 2026-12-21T+90, past the midterms |
| P2 · Not fixed in factEither no final appellate ruling restores access by T+365, or one exists and independent pool coverage is still below its 18 Sept baseline. | The five-network pool back to baseline under a complied-with order. | 2027-09-22 |
| P3 · Substitution, not vacuumAdministration feeds and single-outlet friendly pools carry more presidential-event footage than the five-network pool, and at least one CNN or CBS editorial board is altered, overridden or lapsed. | Pool share back to majority, or both boards intact and exercised. | 2027-09-22 |
The base rate is n=2, and it is stated at that size. Pentagon, October 2025: policy unchanged at eleven months. AP: injunction disobeyed, narrowed on appeal to the East Room, undecided at nineteen months. Both had the cheap fix available and undelivered in fact. That is what P2 rests on, and it is two cases.
Baseline one is recorded; baseline two is still owed. A forecast whose baselines are never recorded is a forecast that cannot be wrong, which is the one property the register refuses.
Enumerated from the public pool-report archive against the White House's own daily guidance and its three rotation announcements for the window, every item coded on what the record shows rather than what the schedule said: two Closed Press items were opened to the pool, one "executive orders" pool note turned out to be an advisory rather than an admission. Presence, not airtime; one fortnight, heavy with travel; the T+365 re-run uses the same fourteen calendar days and a second window chosen in advance. Unratified: the analyst has not reviewed the coding.
The finding the count produced: the substitute was already seated. Every day's pool carried a second, White-House-selected TV outlet beside the network on duty; when the networks withdrew on 21 Sept, that slot became the only TV in the pool. P3's mechanism did not have to be built after the ban.
Observed since the lock, none of it a result. On P1: the President vowed an appeal at 3:39 p.m. on 21 Sept, before the government's response was due and two days before Judge Kelly's hearing — the cheap fix declined in advance, in his words; the hearing on Wed 23 Sept at 3:30 p.m. is P1's first checkpoint, and P1 turns on whether any restoration comes by compulsion or by choice. On P3: the first day without the five-network pool produced a helipad event with no usable audio and, seven hours later, a round-the-clock White House livestream. That is the vacuum and the substitute on one calendar day; it is one day and not a share, and the T+365 re-count keeps administration feeds and the pre-seated secondary slot as separate channels. Recorded in slate.json → observations_after_lock; P1–P3 and their triggers are unchanged.
The substitute was not begun on 21 Sept; it was completed. The channels that fill the vacancy were launched four to seventeen months before it existed — an aggregation site, an app placed on every federal device by mandate, a second TV slot in every daily pool — and the one launched afterwards took seven hours. The inclusion rule for P3's re-count is fixed now, before anyone has a number to protect: a substitution channel originates or carries presidential-event footage a pool camera would otherwise have produced. The slot, Trump TV and the app count; the Wire and the Media Bias portal do not, and are recorded as adjacent structure. The rule was not chosen by reference to what was vacated, because the two sides share no unit — one rotating camera slot and three credentials, one of them print, against streams and sites — except footage share by origin, which is the unit the baseline was counted in.
Still owed, with a deadline: the composition and charter of the CNN and CBS editorial boards at merger close, expected by 30 Sept.
Baseline record: BASELINE-2026-09-22-pool-footage-share.md · rows: .json
Not locked, on purpose: "never letting anything but lies be broadcast." An absolute has no trigger. Its measurable residue is P3.
Slate: PRESS-ACCESS enters as a candidate with a dimension-level expectation only — high S3, high S5, low S1 — and no band. The breadth guard keeps it out of the denominator until a band is locked; supplying one now, after the annex was assembled, would be a prediction shaped by its evidence.
Record: AMENDMENT-2026-09-22-press-access-forecast.md · slate.json → locked_predictions
The bundle includes the statistics pipeline. It runs on the standard library; a self-test of the estimator generates its own corpora in memory and is a unit test of the arithmetic, never a finding.