What this site cannot see

What transparency cannot see

1. A registered share divides a state's count by a Census estimate

The registered share puts two different kinds of figure beside each other. The numerator is a count of registrations as the state's own election office files it (voter_registration, the state's newest period in the cycle): its total of every status where it files one, else the count it reads as its registered total, and the row says which and what population that count is of (active voters, all statuses). The denominator is a survey estimate of people, the Census Bureau's Citizen Voting Age Population tabulation from the ACS 5-year estimates (census_cvap, the newest release on file, its span and margin of error on the row), with no state's own eligibility rules applied. A state's rolls can hold more registrations than the estimate counts citizens 18 and over (voters who moved or died and were not yet removed, pre-registered 16 and 17 year olds, the estimate's own margin), so a share over 100 percent is served as computed and flagged, never clipped. The federal survey's counts are never the numerator, and a state that registers no voters has no share.

rendered on: registered_share

2. A state race is the certified returns as two public sets compiled them

State race pages list the candidates in the states' certified returns as the MIT Election Data and Science Lab (2016 to 2024, every state office) and Carl Klarner (1967 to 2016, state legislatures) compiled them. A winner on an MIT Election Lab row is derived by ranking the summed votes against the seats; on a Klarner row it is the filed outcome. Primaries are in these sets only for 2020 and for the types Klarner files in some states, statewide races start in 2016, and a year MIT has not published is not here. The state's own results page is the record.

3. Money beside a state race comes from a matched state filing

The money beside a candidate on a state race page is what a state register files under a filing matched to that ballot line, by the name and the seat or by an unambiguous name match. It covers every year of that filing on record, not only the race's cycle. A match is a match, never an identity the state stated. A candidate with no matched filing shows no money, which does not mean none was raised.

4. Who stands behind an entity is read from the registers that attest it

A committee's or an organization's principals are the sponsors, parents, officers, controlling candidates and funders a public register names, in three tiers. Filed: the entity's own filing names them. Register: another register attests the tie, such as the FEC's linkage file or the IRS group exemption. Inferred: a reading of filed gifts under a stated rule, shown with its share and its formula, never as a fact. Every row names its register, its row and the rule version, and every name is shown as filed. A tie no register attests is not drawn. An LLC's owners no register names, a social welfare group's donors, a person matched by a name that looks the same, and coordination are never inferred.

5. A spender's interests are read from who funded it, as the registers attest

A committee that spends on a race and files no sponsor is read by where its money came from: each gift it reported is followed to the giver's own filings, hop by hop through other committees, and each giver's face is what that organization tells Congress it lobbies on, or the industry a public register lists it under. The chain is shown with every hop and the register that resolved it. Where a hop ends at a group that files no donors, at a company no register knows, at a committee the trail cannot follow, or past the depth the rule follows, the page says so and that money carries no face. A face says what a funder lobbies on, never that the spender acted for it.

rendered on: cand_outside_face_shares · outside_industry_shares · spender_face_shares · subject_outside_totals

6. True dark money is legally invisible

501(c)(4)/(c)(6) organizations are not required to disclose donors. When they give to super PACs the transfer is visible but the trail ends at the org's name, by law, not by our limitation. PACrat shows the wall and labels it a wall; it cannot see through it, and neither can anyone else.

rendered on: cand_outside_face_shares · committee_money_share · outside_industry_shares · spender_face_shares · subject_outside_totals

7. Sub-threshold donors are invisible by design

Un-itemized small donations (≤$200/donor/cycle in federal races) appear only as lump sums, a privacy feature of the law we endorse. Any figure computed over itemized rows is therefore a floor or a mix, and is labeled as one. (Conduit-itemized earmarks can see below the floor; that exception is labeled where it applies.)

rendered on: bill_face_money · cand_face_shares · cand_industry_shares · cand_maxout_givers · cand_org_money · cmte_industry_shares · donor_band_shares · industry_cycle_totals · industry_party_shares · industry_receipts · industry_receipts_by_industry · money_above_district_median_share · money_above_district_median_share_candidate · money_above_state_median_share · money_above_state_median_share_candidate · money_zip_college_share · money_zip_college_share_candidate · money_zip_income_quintile_shares · money_zip_income_quintile_shares_candidate · money_zip_income_vs_district · money_zip_income_vs_state · money_zip_median_age · money_zip_median_age_candidate · money_zip_median_income · money_zip_median_income_candidate · occupation_unmapped_share · out_of_state_share · race_dark_share · race_face_shares · retired_share · small_donor_floor_share · subject_cycle_totals · top_occupation_groups

8. LLCs and shell entities can launder identity

A contribution from "Sunrise Holdings LLC" is disclosed as exactly that. PACrat can show shared addresses and treasurers; it cannot pierce veils, and refuses to pretend otherwise.

9. Self-reported fields lie casually

Employer, occupation, and spending purpose are unverified free text. The verbatim anchoring contract preserves the mess honestly; aggregate views over these fields are as-filed strings, not resolved entities, until a confidence-tiered resolution layer exists. Even then, both layers stay visible.

rendered on: bill_face_money · cand_face_shares · cand_industry_shares · cand_maxout_givers · cand_org_money · cand_outside_face_shares · cand_pac_industry_shares · client_year_lobbying · cmte_industry_shares · committee_party_lean · donor_band_shares · industry_cycle_totals · industry_party_shares · industry_receipts · industry_receipts_by_industry · money_above_district_median_share · money_above_district_median_share_candidate · money_above_state_median_share · money_above_state_median_share_candidate · money_zip_college_share · money_zip_college_share_candidate · money_zip_income_quintile_shares · money_zip_income_quintile_shares_candidate · money_zip_income_vs_district · money_zip_income_vs_state · money_zip_median_age · money_zip_median_age_candidate · money_zip_median_income · money_zip_median_income_candidate · occupation_unmapped_share · outside_industry_shares · payee_concentration · program_spend_share · race_face_shares · registrant_year_ld203 · registrant_year_lobbying · retired_share · spender_face_shares · subject_cycle_totals · subject_outside_totals · top_employer_share · top_occupation_groups

10. Payee concentration is arithmetic, not a verdict

Payee names on Schedule B are the committee's own free-text strings, grouped only by a versioned character fold and never resolved to vendors, so any concentration figure is a floor: one real vendor split across spellings reads as several. High concentration is the normal shape for some committee types (a super PAC buying advertising sends most of its money to a few media firms). The number states where the money went, not why, and is never by itself evidence of misconduct. Schedule B also ends at the first payee: sub-vendor arrangements are invisible, and where credit-card or payroll sub-itemizations make the ultimate payee ambiguous (memo dollars above the published gate) the measure refuses rather than guesses. Every surface serves it beside money-to-candidates and the committee's peer percentile.

rendered on: payee_concentration

11. Timing gaps are real

Quarterly filers mean money can move months before the public sees it; 48-hour notices cover only some windows. "Current" is never fully current, which is why freshness labels exist.

rendered on: cand_burn_rate · cand_burn_rate_cycle · cand_disb · cand_receipts · cand_runway_days · cand_spent · cand_velocity · disclosure_lag · race_contest_disb · race_contest_receipts · race_contest_spent · race_dark_share · race_disb · race_outside_floor · race_receipts · race_spent · small_donor_floor_share

12. Correlation is not purchase

Money follows ideology as much as it buys outcomes. Every money-beside- votes view carries this sentence.

rendered on: pre_act_move

13. A lobby group here is a pattern in the gifts, not an organization

The detector groups organizations whose people and PACs gave to the same members in the same three months, with money that faces the same two or three legislative subjects. That is a pattern in filed gifts. It is not a registration, a membership, an agreement or coordination, and a person who gives is not acting for their employer. An unnamed pattern is shown only as totals, with no giver named, until a reviewer's verdict names it; a vetoed one is not shown. The faces read only what lobbying reports file, and committee seats are the current roster, not the seats a member held on the day of a gift.

14. Coordination is invisible

Legal "non-coordination" between campaigns and super PACs is a legal fiction the data cannot test. PACrat shows parallel spending; it does not infer coordination.

rendered on: cand_spent · ie_oppose_general · ie_oppose_received · ie_support_received · ie_weight · race_contest_spent · race_dark_share · race_outside_floor · race_spent

15. Market moves carry no identities and no causes

A prediction-market price series shows THAT the odds moved. It never shows who traded (the regulated exchange discloses no counterparties, and on-chain wallets are pseudonymous and never attributed here), whether traders were informed or merely lucky, whether a move was information or manipulation, or what happened on offshore and over-the-counter books. Laying a move beside an official act is arithmetic on dates, never a causal claim. No measure of public expectation exists, so "surprising" is itself unmeasurable.

rendered on: pre_act_move

16. This market history is months deep, not decades

US-legal markets on congressional election outcomes have existed only since October 2024 (Kalshi v. CFTC: district ruling Sept 6, 2024; stay denied Oct 2, 2024), and the exchange's public API serves no market history from before roughly 2024. Every percentile and base rate here stands on months of history and a small number of market-by-act joins. That is enough to rank and not enough to conclude, and this page will say so until the history says otherwise.

rendered on: pre_act_move

17. Market prices are context, not records

Stock price series come from a public quote surface (an unofficial market-data endpoint), not a government register, and are labeled as such. A weekly walk stores one closing price per ticker per trading day for the tickers filed on trade reports; what is served is those stored closes, read from PACrat's own copy, never a live or intraday feed. Each close is the source's adjusted close (splits and dividends folded in) where the source gives one, and its plain close where it does not. Price movement around a trade is CONTEXT for the reader: it is never converted into a member's gain or loss, because the STOCK Act discloses brackets, not amounts; any dollar P&L would be an invented number.

18. A request is not an appropriation

Earmark request tables are member-submitted and committee-published: asking is not getting (FY2025 requests met an almost earmark-free enacted year), and an enacted line is not an outlay. Recipients carry no standard identifier, so joins to donors or lobbying clients are exact-text adjacency on self-reported strings: a name variant is a silent miss and a match is proximity, never causation.

19. Jurisdiction is authorization, not causation

The agency-to-committee map is curated AUTHORIZING-committee jurisdiction (reviewable data; unmapped is an honest miss). It is not hearings, not investigations, not appropriations. Contractor rankings are USAspending's top-N by obligation: complete at that grain only, small recipients invisible, pass-through and subawards not unwound. A member sitting on a committee whose agency pays a contractor who lobbies is a LOOP OF ADJACENCIES, never evidence that any leg caused another.

20. A vote is a recorded position, not proof of motive

Roll-call rows are the chambers' official records served verbatim (House Clerk EVS, Senate LIS); tallies are the chambers' own published totals, never recomputed. Bill links parse conservatively from the vote's own document reference; nominations, quorum calls, and unrecognized references stay unlinked, with the reference kept verbatim. Voice votes and unanimous-consent actions never appear here at all: most of what a chamber decides has no roll call. A member's vote shown next to money, trades, or lobbying is adjacency, never evidence one caused the other.

21. IE totals come from one deduplicated register

Schedule E rows come from e-filed F3X, F24, and F5 documents, verbatim. The same expenditure legitimately appears more than once in the raw rows (a 24/48-hour F24 notice re-reports on the quarterly F3X, and an amendment refiles under a new file_num alongside the original), so summing raw rows double-counts by construction. Every outside-money total served here is summed from one deduplicated register instead: a periodic report supersedes the notices for the window it covers, and an amendment's newest filing wins. The FEC's electronic intake checks format, not truth, so the register also carries fabricated and mis-keyed filings; totals therefore exclude filings flagged by reviewed mechanical detectors or ruled out in the public adjudication ledger, pending or after human review, and every total states its share of never-corrected provisional notices. Excluded rows are never deleted or edited: raw rows are always served beside the totals, exactly as filed. Support/oppose is the filer's own designation, and candidate identification is as filed. What the law cannot see is not here: paper filings outside the e-filing lane, issue ads that avoid express advocacy, sub-threshold spending by persons, and the donors behind a c4's IEs are invisible by statute, not by omission.

rendered on: cand_disb · cand_outside_face_shares · cand_receipts · cand_spent · outside_industry_shares · race_contest_disb · race_contest_receipts · race_contest_spent · race_dark_share · race_disb · race_outside_floor · race_receipts · race_spent · spender_face_shares · subject_outside_totals

22. The exempt-org file is a classification, not a disclosure

EO BMF rows are the IRS's registry of exemption status: subsection codes are tax classifications, and the financial figures are the extract of each org's MOST RECENT return, stale by construction (multi-month filing lags). Churches and sub-threshold orgs legally never appear: absence is not nonexistence. A c4's donors are not here and are not anywhere; that invisibility is statutory. An EIN joins a committee or a company only where a register files both keys (an 8871's election authority id, an SEC record's EIN); a name resemblance is never a link.

23. The FARA registry is a floor, never a census

FARA rows are who REGISTERED as foreign agents, verbatim from the DOJ's own exports. Registration reflects compliance choices, exemption claims (the LDA exemption above all), and enforcement climate, not the universe of foreign influence. The registry has been actively thinning; absence of a name is silence, not evidence of absence. Country labels are the registrant's own filing. Joins to lobbying or personnel data are exact-string only; a name match is never identity resolution.

24. Disclosure brackets are ranges filed under oath, not numbers

Annual FD rows are what the filer disclosed, verbatim. Asset values, income, and liabilities arrive as BRACKETS ("$1,001 - $15,000") and ship as brackets: no midpoint, no estimate, ever. Owning an asset is lawful and disclosed by design; a holding beside a committee seat is a correlation surface, never an accusation. Paper filings are listed but their contents stay in scans (an honest miss); e-filed reports cover only what the form asks, and thresholds hide small holdings by statute. Filer names are served as filed on EFD. A report is linked to a member only when exactly one senator on PACrat's roster has the filed surname (a suffix such as Jr. or III set aside) and a given name where one is the start of the other (Dave and David); when two senators fit, or none, the report stays unlinked, and a candidate who never served in the Senate is never linked.

25. A case is a proceeding, not a verdict on character

Enforcement rows are the Commission's own case records, dispositions VERBATIM. Most MURs close in dismissal or with no action, which renders with exactly the same weight as a conciliation agreement. A respondent name is as filed and never resolved to a person or committee by matching; the admin-fine lane's committee_id is the only exact join. Being named respondent is an allegation processed, not wrongdoing established. The Commission has lacked a quorum to fine since May 2025, so recent silence is institutional, not exculpatory or inculpatory, and this ledger's computed lateness arithmetic runs independently of the frozen agency.

26. An EC names a candidate; the law forbids reading it as advocacy

F9 electioneering communications are broadcast ads naming candidates inside the 30/60-day statutory windows WITHOUT express advocacy. That legal distinction is the form's entire reason to exist, and these rows must not be read as support or opposition (the form has no such field). Amounts and candidate names are as filed; amendments coexist under new file_nums, so sums double-count. Who funded the communication is usually invisible: donor itemization (F92) applies only to non-segregated accounts, and c4 general treasuries stay dark by statute. Online-only ads never file an F9 at all.

rendered on: race_outside_floor

27. 8872s are the organization's own reports, with statutory holes

527 rows come from the IRS POFD full dump verbatim: amounts, names, employers, and the organization's own totals as filed; amended reports coexist with originals (the amended flag rides each row) so summing across filings double-counts. What the statute cannot see is not here: qualified state/local organizations that report to states instead of the IRS, organizations under the filing thresholds, contributions aggregating under $200, and expenditures under $500 are invisible by law. An EIN joins an FEC committee only where the organization files the committee's id on its Form 8871; nothing here crosslinks by name matching.

28. A platform ad library is the platform's self-classified record

Google's and Meta's ad libraries hold the ads the platform itself classified as election or political ads, for advertisers it verified; an ad it did not classify is not there, and the platform can revise or delete history (PACrat archives every pull for that reason). Per-ad spend and impressions are published as buckets and served as buckets: the sum of lower bounds and the sum of upper bounds are two honest numbers, and their midpoint is not one. Advertiser-level spend is the platform's exact cumulative figure across every election it lists. None of it is an FEC filing; where both exist the page shows both and names the residual.

29. A political file shows the order, not the airing or the dollars

OPIF rows are stations' own uploads to the FCC's public file, and the regime covers broadcast, cable, and satellite only. Online platforms keep their own libraries (Google and Meta publish theirs; see platform_ad_library), the platform's self-classified record and not a government file; streaming and connected-TV political spending has no public record anywhere, and what campaigns know about it comes from paid tracking the public cannot audit. Upload time is not air time. Folder labels are the station's own words ("GREEN-R-CONGRESS") and resolve to candidates only through the folder crosswalk, a reviewable registry of unique mechanical matches; ambiguous and unmapped labels stay unlinked, and anything built on the crosswalk serves its mapped share beside the number. Completeness equals station compliance, which the FCC enforces sparsely. The dollar amounts live inside the uploaded documents; PACrat links the documents, and where its contract parse has read an order it serves the flight dates only, so this lane serves no amounts at all (see fcc_order_dollars).

rendered on: disclosure_lag

30. Order dollars are parsed but not served

PACrat's contract parse reads the printed totals on stations' political ad orders, and none of those amounts is served until the parse passes its validation corpus, a pinned set of orders checked by hand. Until then a page shows counts of uploads and the flight dates the parse read, flagged where the parser refused an order's totals, and links every order to the filed paper at the FCC, where the amounts can be read directly. The flight dates carry the same caution: they are what the parse read, not yet validated.

31. Lobbying filings are registrant-reported estimates

LDA quarterly filings report income/expenses as the registrant's own estimate, rounded under the statute ($10,000 buckets), and often left blank when below threshold. PACrat serves each filing's figures as filed and never sums across filings (amendments re-report the same quarter, so a sum without amendment dedup double-counts; the versioned client_year_lobbying measure now resolves exactly that dedup with a published formula and is served beside the raw rows, never in place of them). Sub-threshold lobbying, grassroots activity, and state-level lobbying are invisible by law. Issue codes are the registrant's self-selection from the statutory list.

rendered on: bill_face_money · cand_face_shares · cand_outside_face_shares · client_year_lobbying · outside_industry_shares · race_face_shares · registrant_year_ld203 · registrant_year_lobbying · spender_face_shares · subject_cycle_totals · subject_outside_totals

32. Zip codes resolve through the Census ZCTA approximation

A zip code becomes a place through the Census Bureau's ZCTA relationship files, which approximate USPS zips with populated areas. PO-box-only and building-only zips have no ZCTA and honestly do not resolve to a district. A zip split by a district line shows every overlapping district with the Census overlap areas, biggest share first. The crosswalk carries its congress vintage; after redistricting, older cycles' maps differ from the vintage served.

33. A race's field is FEC registrations; the state's list decides the ballot

Race pages list candidates who filed with the FEC for that seat and cycle. Ballot access is state law: someone can be on the ballot without an FEC filing (below the spending threshold) or file with the FEC and never reach the ballot. Where the state election office's own candidate list is loaded, each filing is read against it by a name rule: a candidacy the state lists as withdrawn, not qualified or defeated leaves the race, and a party whose names on the list all match a filing has its side decided by the list. A name on the list that matches no filing is shown as the state filed it, and its party's side is ranked by money, behind the sides the list decides. Sides rank by what each candidacy raised this cycle, primary and general together. Party, incumbency, status, and any general-election result column come verbatim from FEC files; where FEC files record no result, PACrat shows none rather than inferring one.

34. The Record is what Congress prints, not a transcript

The Congressional Record is not a verbatim transcript. Members may revise their remarks before printing, and material can be entered without ever being spoken aloud; the Extensions of Remarks section is entirely that. The print edition marks some inserted passages typographically, and that marking does not survive into the plain text served here. The speaking members on each item are the Government Publishing Office's own attribution for that item as a whole, not a per-sentence record of who said what. Quote this text as what the Record prints, never as proof words were said on the floor.

35. Congressional trades disclose ranges, not amounts

STOCK Act reports bracket every trade ("$1,001 – $15,000"); PACrat serves the bracket verbatim and never sums, midpoints, or averages it. Blind-trust and exempt holdings are invisible; owner attribution (self/spouse/dependent) is as filed; the statute allows up to 45 days between trade and disclosure. Reports filed from 2020 on are read. Senate trades are parsed from e-filed reports; Senate paper filings are listed with no trades. House trades are parsed from the Clerk's PDFs that have the standard e-filed layout; a scanned House PDF is listed and marked paper with no trades, and a PDF in any other layout is listed and left unparsed, never guessed.

rendered on: trades_counted · trades_filed_late_share

36. An amendment pair is two filings, compared

Amendments are routine and the law provides for them. A pair on this ledger is two versions of one report by one committee, lined up by the report's own coverage window and compared field by field under a published matching ladder (data/amendment_deltas.yaml). Lines that match by transaction id compare directly; the rest match one to one by name, date and amount; anything still unmatched is counted as residue, shown, and never guessed to a side. Only e-filed reports can be compared line by line, so paper filings never appear here. Address fields are not compared or served on this ledger. The comparison states what differs between the two official records and claims nothing about why.

37. A request for more information is not a finding

The FEC sends Requests for Additional Information in the ordinary course of reviewing reports; thousands go out every year. A letter shown here is paired with a report by committee and coverage window only, and with any amendment of that report by date order only. The dates sit side by side; whether an amendment answers a letter is a judgment this ledger does not make, and a letter is not a finding of wrongdoing.

38. One committee can serve two candidacies

The FEC's linkage file sometimes ties one principal campaign committee to more than one candidate id in the same cycle: a mid-race handover (the committee that was Biden's became Harris's in 2024), a shared ticket, a candidate re-registered under a new id, or an error in the linkage file itself. The FEC's summary file credits that committee's whole cycle to each id it served. Where the ids belong to one person, PACrat counts the pot once. Where they are different people, each person's page carries the FEC's own credit and says so in place; the same dollars sit on both pages by the FEC's attribution, so figures from two people's pages must never be added together.

rendered on: cand_burn_rate · cand_burn_rate_cycle · cand_runway_days · cand_velocity · money_above_district_median_share_candidate · money_above_state_median_share_candidate · money_zip_college_share_candidate · money_zip_income_quintile_shares_candidate · money_zip_income_vs_district · money_zip_income_vs_state · money_zip_median_age_candidate · money_zip_median_income_candidate

39. An age or a gender is what a public roster lists

A member's birthday and gender come from the congressional roster dataset (unitedstates/congress-legislators), and a state legislator's from the Open States dump, exactly as each lists them. An age is counted from a listed birthday on the day the page or the response was made, and that day is stated beside it. A person the roster lists no birthday or gender for is shown as not listed, and PACrat never infers one from a name, a photo or anything else. A candidate who has never served in Congress is not on the congressional roster, so none is shown for them. The chamber composition counts sitting members only: those whose latest term on the roster has not ended.

40. A report's cash arithmetic is checked as filed, not audited

For each committee report, cash at the start plus money received is compared with money spent plus cash at the close, using the figures on the report's own summary page exactly as filed. A difference is stated as arithmetic, never as a finding: a figure can be misread from an old form version, a report can cover a window that overlaps another, and an amendment replaces what it amends. When several filings cover the same window, the one signed last is used and the others are named. The gap between one report's starting cash and the previous report's closing cash is shown the same way. A figure the report did not file leaves the result blank.

41. A transfer's two sides are matched by a stated rule

Money one committee reports sending to another, and the money the other reports receiving, are paired when the two committee ids, the amount to the cent and dates no more than two weeks apart agree. The pair is shown beside both filed rows, which are never merged or changed. A side with no partner is not a finding: the other committee may file on paper, in a different window, or under a different amount, and some committees are not in these records at all.

rendered on: cand_outside_face_shares · outside_industry_shares · spender_face_shares · subject_outside_totals

42. A person over the registers is an inference

No register files a person. The FEC files candidacies by id, Congress keeps a roster by bioguide, and each state files committees and filer names under its own keys. PACrat joins those records into one person by a stated model: filed signals (names, seats, dates, a shared treasurer or address, an identifier another record states) scored against weights in a reviewed data file, vetoes that cut a join whatever the score, and a sampled audit that measures how often each band is right. Every joined record keeps its own id and page, and every join shows the signals and the sentence that made it. A join the audit has not cleared is served as review, not as fact, and a name alone never joins two records. A wrong join or a missed one is a model error, never a filing.

43. A figure added across registers is a computed claim

No register files a person's money across the others. The FEC files federal campaigns, each state files its own, and the rules for what counts as money raised differ by register. Where PACrat adds a person's money across registers or across levels, the figure is a computed claim: it comes from a stated formula with a version, every term that went into it is listed beside it with the register, the rule and the span it came from, and every term the formula refused is listed with the reason. The per-register figures stand beside the sum and are the filed facts; the sum is not. A term is never added twice, and two registers of one state whose spans overlap contribute only the primary register's term.

What this site does not build

No person-level donor forensics before the IDENTITY layer

Recurring-donor and "structuring" claims computed over as-filed name strings publish artifacts, not findings (DESIGN P2/P8/§2.8). Nothing person-clustered ships until a confidence-tiered resolution layer exists, federal or state.

No LLC/shell unmasking

The registries that matter disclose no members; every successful unmasking was bespoke journalism (DESIGN P5, discarded). The honest product is showing the wall.

No coordination inference

Legally untestable (VISION §7.7). Parallel spending is shown; conclusions are the reader's.

No person pages, no reverse donor lookup, no donor maps

Committees and candidates are the units of power; neighbors are not. Contributor search stays transaction-lists at fec.gov parity.

No composite scores, no conclusion-embedding names

A published number is non-editorial only if it is deterministic, symmetric, single-formula, versioned, and un-weighted, with its full distribution alongside (DESIGN §1.2). Features get descriptive names; leaderboards are user-applied sorts.

No hand-curated ad or evidence archives

Editorial hunting is the labor model that killed the predecessors. Only mechanically enumerable corpora are admissible (DESIGN P9).

No naked-number relays

No surface, including MCP tools answering AI assistants, ships a figure without its formula version and peer context.

Machine-readable: /v2/boundaries.