Methodology

What this site shows

This site tracks two kinds of federal primary source filings for a watchlist of entities: lobbying disclosures filed under the Lobbying Disclosure Act, and federal contract awards recorded by USAspending. Two kinds is the complete list; no other source, campaign donations, stock trades, insider filings, feeds this site yet. A confluence, in the definition this site computes, is a watchlist entity with at least one lobbying filing and at least one contract award, both entity resolved, inside a rolling 90 day window. The window reads one date per filing. For a contract award it is the Base Obligation Date, the date USAspending records the award was first obligated, so every award in a confluence is a new award; a later modification to an older award does not bring it into the window, however recently it happened. For a lobbying filing it is the date the Senate posted the filing, its dt_posted, since a lobbying filing carries no obligation date. That is the same definition, computed by the same shared database function, that the underlying system uses internally to decide whether an entity is worth a human’s attention, not a separate definition built for this site alone.

Two kinds present is not the same claim as two independent signals converging in time. A registered lobbyist files roughly every ninety days, the same length as this window, so a lobbying filing sits inside it for almost any entity that lobbies at all, nearly the whole year rather than only around one date. Checked weekly across 36 points from 2026-01-05 through 2026-09-07, all but one of the twenty three watchlist entities with any lobbying registration showed a filing inside the trailing ninety days in at least thirty five of those thirty six checks. What a confluence actually tracks, in practice, is the contract side: which watchlist entities that already lobby also show recent federal contract activity. The lobbying half of this test is a gate an active lobbyist clears almost automatically, not a timing signal in its own right, and this page’s own wording says so rather than implying a coincidence the underlying filings do not show.

Every count, total and date on this site is computed directly from the filings listed on the filings page. Nothing here is estimated, projected or modeled.

The filings page is not windowed the same way. It lists every filing on record for an entity regardless of date, since its job is to be the complete evidentiary record a confluence traces back to, not a second view of the same ninety day window.

The plain language sentence

The line under each entity on the confluence page is generated, not written by a person. It states who the entity is, how many lobbying filings and contract awards it has in the window, what those total, and the date span they cover, drawn directly from the same reconciled figures the table above it shows. Nothing in it is estimated or paraphrased from a filing’s own text: it is those numbers, in a sentence, and that is the whole of what it claims to be.

Every dollar figure and date the sentence states is checked against that row’s own totals before this site is allowed to build. A sentence that named a figure the row does not carry would fail that check and stop the build rather than publish a claim this site cannot stand behind, the same discipline confluenceSplit() already applies to the table itself, described above. This mirrors the internal system’s own version of the same sentence, written for its Money Signals board and checked the same way; the two are independent generators over two different stores, described further under what this site shows, and are not expected to read identically, only to both be traceable to their own filings.

Two source types are the whole of what a confluence claims here, never a timing claim: see the reasoning under what this site shows on why a lobbying filing sits inside the window for nearly any entity that lobbies at all, so what the sentence is really reporting, in practice, is the contract award.

This page shows the sentence and never a numeric score. The underlying system’s own Money Signals board carries one, Confluence Score, because a row there can show three or more source types and the score does real work telling those apart; every row here meets the two source floor by construction, so a score would only repeat what the sentence already says rather than distinguish anything. The public view deliberately shows the traced sentence rather than a score.

What a contract amount does and does not claim

USAspending’s award search, the one this site’s filings come from, returns an award’s current state: who received it, its current total obligated value, when it was first obligated, when the record was last touched. Every contract Amount on this site is that total as USAspending reported it at the moment this system asked, not a figure for activity inside any particular window.

What that total is made of is asked separately, per award, of USAspending’s transaction count and transaction endpoints, and stored in its own table, contract_award_actions, one observation per award per check. Its central field is the action count. An action count of one means the award’s total is a single obligating action, so the figure is a new award amount and nothing about it is cumulative. More than one means the total sums an original action and later ones; the same observation records how many of those actions obligated anything and what the most recent one did, which the filings pages show under Made of. What it does not give this site is the share of a multi action total that fell inside any window, which is why the confluence page labels a multi action total as the award’s current total and never as money that moved in the window.

Unmeasured is the third state, and it is a gap in this system’s coverage, not a claim about the award. As of this build, 16177 of the 16178 contract awards on file have been measured, and 282 of the 282 new awards in the current confluence set; both figures are read from the store when the site is built, not typed in. The check costs three requests per award, so it fills in over weeks rather than in one run. Awards in the confluence set are measured before any other backlog, so that table is the first to reach full coverage. Unmeasured never renders as anything else, and in particular never as a count of zero: an award in this store has at least one action by definition, so zero is not a possible answer and printing it would be a fabrication.

Before that endpoint was found, this site tried building the same label by inference, from the gap between an award’s own obligation date and when USAspending last touched the record, on the reasoning that a wide gap meant a later action. Checked directly against every award this store had itself observed on two separate occasions, actual dollar amount included, that gap did not predict whether the amount had changed: of the 518 such awards whose amount never moved between observations, 285 still carried a gap wide enough to have been labeled, and of the 7 whose amount did move, 4 carried a gap narrow enough to have passed unlabeled. Worse than chance, in both directions. The label was removed rather than recalibrated, and the record of that is left here on purpose: the fix that eventually worked was asking the source a question it already answers, not tuning a threshold on a signal that was never there.

Per action detail arrived later the same day, once the endpoint carrying it had actually been called rather than read about. Where an award has been checked, a row now also says how many of its actions obligated anything and what the most recent one did. The first award this was run against is the reason it is worth having: two actions, of which the base award obligated 2,496,000 dollars and the second obligated nothing at all, an administrative modification to incorporate a form. A count of two, on its own, invites the opposite reading, and since 441 of the first 500 awards checked carry more than one action, that reading would be wrong most of the time.

A latest action of zero dollars is therefore a real and meaningful value on this site, not a missing one. It says the award was touched and no money moved. Missing is still written as unmeasured, and the count of actions that obligated something is never shown as zero, since an award with a total obligated value has at least one such action by definition.

Since 11 September 2026 each filing’s own page lists every recorded action against a measured award, with its date, modification number, type and what it obligated, as USAspending returned them. Where an amount on the front page moved, the latest action is shown beside it when the award has been measured, which is usually the action behind the change, and its date says whether it can be. For an award checked before the per action fetch existed the count is known and the individual actions are not, and its page says so. Occurred, the award’s own obligation date, and Published, when USAspending last touched the record, remain on every row exactly as the source gives them.

What the front page counts as a change

The store this site reads is append only and versioned. A filing is never overwritten: the same filing reappearing with different content from the source is inserted as a new version next to the one before it. That makes two questions answerable from rows already here, with no new measurement of any kind, and the front page asks both. Which filings were seen for the first time, and which already recorded filings had their dollar amount change.

Two kinds of row are deliberately left out of that page, and both exclusions matter more than what is in it. First sightings that arrived through the historical sweep and the one time backfill are not changes: thousands of them landed on the single day this system fetched them, and presenting that as a day of federal activity would be the most misleading thing this site could do with its own data. They are counted by day at the bottom of that page, labeled by how they arrived, and listed in full on the filings pages.

Second, and quieter: an amendment whose amount did not move is not listed. An amendment here means the source response changed fingerprint, which for a contract award can be nothing more than USAspending touching the record. Of the 86 amendments this system observed live in the first week of September 2026, none moved a figure. A ledger that counted all of them would read as constant activity while reporting nothing at all.

The day a change is filed under is the day this system observed it, not the day the underlying event happened at an agency. Both dates are on every row, and they can be weeks apart: a contract action from mid August can reach USAspending in early September and this system the same night.

New to the record, and new only to this watchlist

A first sighting is the day this system first stored a filing. That is not always news. When an entity or an alias is added to the watchlist, the next daily poll asks about it for the first time and stores filings that had been public for months, all in one day. On 10 September 2026 that put 47 lobbying filings onto the front page as if they had just happened, 44 of them posted in April and July. It is the same mistake the page already refused to make with the historical sweep, arriving by a different door.

So every first sighting is compared with the date the source itself gives the filing: the day the Senate posted a lobbying filing, or the day USAspending last modified a contract award’s record. Seen within seven days of that date, it is new to the record, the only kind the front page presents as news and the only kind in the site wide feed. Seen later than that, it is new to the watchlist: public before this system was watching, listed separately, with the number of days it had been public. A filing whose source gives no date is undated and listed with the second kind, since nothing can say it was new. Seven days is the same tolerance the front page has always used for on or near the day the source published it. The classification is computed in the store, in signal_store.first_sighting_novelty, from dates on the rows, and nothing about it is estimated.

For a contract award the comparison is with the last modification, not the obligation, so an older award whose record was modified this week and first seen now counts as new to the record: the record changed at its source, and its Occurred date says how old the award itself is.

The watchlist, and who chooses it

The entities this site watches are chosen by the editor of Our Revolution Media from entities named in its reporting. The watchlist page lists every one, with the editor’s own one sentence reason for watching it, the series it belongs to and the published piece it ties to where there is one. Those reasons are statements of editorial interest, stored in signal_store.entity_watch_reasons and written by a person; they are not findings of this site and nothing here writes one. An entity with no recorded reason is shown as having no stated reason yet.

The activity baseline, and why it is uneven for Defense work

Beside each entity’s last 90 days, the site shows the same count for the 365 days before, scaled to 90 days by signal_store.entity_activity_baseline. The two are shown as two numbers. No ratio or index is computed from them, because either side can be moved by things that are not the entity.

The largest of those is at the source. USAspending publishes Department of Defense contract data on a 90 day delay. As of this build the newest Defense award in the store was obligated on September 17, 2026, and the last 90 days hold 1 Defense awards for the whole watchlist against 331 civilian ones, while the baseline year holds its Defense awards in full. An entity whose federal work is mostly Defense therefore shows a window far below its baseline for that reason alone, which is a lag in publication and not a fall in activity; the watchlist and each entity page show the Defense counts under every figure so the size of it is visible. The same lag is why the confluence page’s by agency table currently has no Defense rows.

Two smaller things. The baseline year always contains one federal fiscal year end, when obligations peak, and the window may not: 2,532 of the baseline year’s 4,596 awards across the watchlist were obligated in a September. And the baseline is this store’s own record rather than an entity’s whole federal history: contract awards reach it through the historical sweep back to 1 November 2024, which covers the baseline year, and lobbying filings begin in 2024.

Permalinks, and how to cite one row

Every filing has its own page, at /filings/, the entity, and the filing’s key at its source, or /filings/unmatched/ and the key for a filing not yet matched to an entity. The key is stable across amendments, so the address does not change when the record does. The page shows the current version, every earlier version this store holds with its amount and when it was retrieved, the source link, the archived copy and how it was made where there is one, the action detail for a measured contract award, and a cite this row block. Every table on the site that lists filings links to these pages, and changes.csv carries the address in its site_url column.

To cite a figure, cite the page, the source address, and the date it was observed, which the block gives together with the content hash. The source can honestly change after that date; the page keeps every version, so a citation made today still leads to the figure it quoted. The license and citation page has the attribution line.

What entity resolution does, and does not, claim

A filing names a client, registrant or contract recipient as a plain text string, exactly as the source agency recorded it. That string is matched to a watchlist entity by whole word comparison against a hand kept alias list, never by a fuzzy or approximate match. A match means the filing’s own text contains the entity’s name or a known alias of it as a complete word, nothing more.

A filing that cannot be matched this way is never dropped. It appears on the filings page marked unmatched, carrying the exact string the source agency recorded, and is queued for a human to review and decide whether it names an entity already on the watchlist under a spelling the alias list does not yet cover. An unmatched filing is not evidence of anything about an entity; it is a filing this site has not yet been able to place.

A match is a claim about text, not about ownership or corporate structure. This site does not resolve subsidiaries, mergers or parent companies beyond what a human has explicitly entered into the alias list.

The alias list is published, entity by entity, on the aliases page, with where each alias came from.

Matching a lobbying filing to a contracting agency

A lobbying disclosure names the government entities its lobbyists contacted, each with the Senate’s own numeric id and name. A contract award names its Awarding Agency, in USAspending’s spelling. Nothing in either source links the two, so this site uses a crosswalk kept by hand: one row per Senate entity id, giving the USAspending Awarding Agency it corresponds to. It is keyed on the numeric id, never on the name, so a change in how the Senate spells an office does not break a match, and it is edited only by a person, never by the daily ingest.

A crosswalk row with no agency means the office has no contracting counterpart in this store, such as either chamber of Congress, an Executive Office of the President office, or the Federal Reserve. That is a decision recorded against the id, not a missing value. A Senate id that appears on a filing and has no crosswalk row at all is unmapped: a new office nobody has placed yet. Unmapped ids are left out of the confluence page’s by agency table rather than guessed at, and counted there. As of this build the crosswalk holds 100 ids, 78 mapped to an agency and 22 recorded as having no contracting counterpart, and 3 ids seen on 4 filings are unmapped.

A match here says a filing named an agency and an entity holds awards from it. It does not say the lobbying concerned those awards, and a filing naming several agencies counts once for each of them.

Update cadence

The underlying store is fed by a daily automated poll of both source agencies’ public APIs, one narrow query per watchlist entity. This site itself refreshes on the same daily schedule, after that day’s poll completes successfully. A number on this site can lag the underlying agency source by up to one day for that reason, never longer under normal operation.

The archive.org column on the filings page moves on its own, separate schedule, roughly every thirty minutes, independent of the once daily refresh described above. A filing can show no archived copy there for longer than that even after one has actually been found or made, simply because this site has not rebuilt since; this site’s own refresh lags a background process that runs more often than the page does. There is no failure state to wait out either way, described further under archived copies below.

The correction path

Every filing on the filings page links to its original source document. If a figure here looks wrong, the source link is the place to check first, since this site never edits or reinterprets what a source filing says. An entity match that looks wrong is a question for the alias list behind entity resolution, described above, not a data error on any individual filing.

Correction, 9 September 2026: lobbying attribution

Anyone who read this site before 9 September 2026 saw a larger number of lobbying filings attributed to several watchlist entities than it shows now. Those attributions were wrong and have been withdrawn.

The Senate lobbying database searches by client name on its own side, and that search is approximate: a query for one company returns filings by other organisations whose names resemble it. Every result is supposed to be re-checked here against the entity’s own name and alias list before it is attributed to anyone. Because of a fault in the ingest code, a filing that failed that check was attributed to whichever entity’s query had returned it, rather than being left unattributed.

The effect was that 576 of 2,322 lobbying filings, 24.8 per cent, were recorded against an organisation that did not file them. Filings by unrelated hospital networks were attributed to Centene; filings by blockchain trade groups to Block; filings by Intuitive Surgical and Applied Intuition to Intuit; filings by the Amalgamated Transit Union and the Transport Workers Union to TransUnion; and filings by ICANN, Nammo and a medical centre to the National Association of Manufacturers. Federal contract awards were never affected, because that half of the ingest never had the fault.

All 576 have been corrected. 54 of them, filed under Accenture LLP, moved to Accenture, which is now tracked separately from Accenture Federal Services. The remaining 522 carry no entity at all and appear on the filings page as unmatched, which is the honest state: the filing is real, and this site cannot say whose it is. The filings themselves were never altered, added or removed, and each one still links to the same source document it always did. What changed is only who this site says a filing concerns.

The confluence view still lists nine entities; none dropped out, because each one held genuine lobbying activity alongside its federal contracts. The evidence behind three of them is smaller than it was: Accenture Federal Services from 46 filings to 41, TransUnion from 17 to 10, and UnitedHealth Group from 14 to 13. A figure cited from this site before that date should be re-checked against the figure shown now.

Legacy backfilled filings, and why some dates are missing

This store began with a one time backfill of filings already on file locally from before this site existed, marked backfilled on the filings page. That backfill carried over every filing exactly as it was originally captured, but it could not recover data the original capture never requested. Contract filings from the backfill carry no occurred date or published date for that reason: the original capture did not request the two USAspending fields those dates come from. Every filing captured since, marked live ingest on the filings page, carries both dates.

Lobbying filings are not affected by this gap. Every lobbying filing, backfilled or live, carries the date it was posted, since that date was always part of the original capture.

A third label, historical, covers filings fetched after the fact rather than on the day they happened, back to 2024-11-01, under the same corrected field list live ingest already uses. A historical filing is not the same claim as a live one: this system did not observe it the day it was recorded by the source agency, only later. It is a stronger claim than a backfilled one in exactly the respect the backfill could not manage, it carries real occurred and published dates rather than none.

What proves a filing, and what content hash checks

What this system captures at ingest is a record of what the source returned, and when: the exact response, stored verbatim, fingerprinted with a content hash, timestamped with when this system retrieved it. That is not a verification of what the filing says, only of what the source said and when this system saw it say it. It is the provenance this site actually claims for every filing here, whether or not an archived copy link appears next to it.

Every filing stores a content hash, computed once, at the moment the filing was first captured from the source agency’s own response. It is a fingerprint of that exact response, and it is what lets a repeated daily poll recognize a filing it has already seen without storing it twice.

That hash cannot be recomputed later from the stored filing to check it. The database reformats the stored copy of a filing on write, in a way that does not exactly match the format the hash was originally computed against, so a later recomputation will not match even for a filing that is completely unchanged. A mismatch on recomputation means the check was performed incorrectly, not that the filing was altered.

What an archived copy means, and the two ways one gets made

The archived copy link next to a filing is a snapshot from the Internet Archive’s Wayback Machine, and it can arrive one of two ways, marked so the difference is never hidden behind one plain link. A filing marked found had already been captured by someone else, at a time this system did not choose and was not present for; this system located it afterward. A filing marked captured was archived by this system itself, at or shortly after the moment the filing was recorded here. The date shown next to either label is when the archive was actually made, not when this system found or made it, which can lag behind by hours.

A found snapshot and a captured one are not the same strength of corroboration: found means someone else made it, at a time this system did not choose; captured means this system made it itself, at or near the moment it recorded the filing. But neither is the filing’s own provenance, which the capture described above already established at ingest, independent of whether any archived copy exists. A filing with no archived copy is exactly as evidenced as one with one.

Most filings will never show an archived copy, and that is the ordinary case, not a gap. Capturing one depends on a public archive this project does not control and, as of September 2026, cannot always reach reliably. A filing with no archived copy has not failed anything and nothing here is waiting on it: it simply has not picked up one additional, optional confirmation that was never a precondition for the filing’s own provenance in the first place.

Being findable, and the one page kept out of search

This site launched public but unindexed, on the reasoning that being linkable and being searchable are two different decisions and the second could wait. It waited on the entity review queue being clear. That condition was dropped on 2026-09-09, because it does not arrive: a store fed by a daily poll of two agencies will always have filings whose subject string has not yet been placed, and treating a permanently non empty queue as a gate meant never opening the gate. An evidence site that a citation, an article link or a grant reviewer cannot find is not doing the work of an evidence site.

One page is still kept out of search, and it is the page that reasoning was really about. The unresolved filings page is a list of raw subject strings this system could not place; those strings name real companies, and a search for a company’s name should not surface them stripped of the explanation of what an unmatched filing means. It stays public and linked from the filings directory. The downloads are kept out of search for a related reason: a multi megabyte table ranking on its own, without the page that says what a null entity or a cumulative amount means, is worse than not being found.

Reusing a figure, and citing one

The filings are federal records and carry no copyright; what this project made is the watchlist, the alias list, the observation history and the action counts, and that is published under a Creative Commons attribution license. Every table here is downloadable as a static file, rebuilt with the site once a day.

For a citation, a source link alone is weaker than it looks, and not because the agencies are unreliable: an award record is designed to change as money is obligated against it, so a link that showed one figure when you cited it can honestly show another later. Cite the filing’s key, the content hash of the response this system stored, and the date it was retrieved, all three of which are on every row and in every download. The license and citation page has the exact form, and the limits of what the hash proves are described under content hash above.

What is not on this site

This site does not compute or display a salience ratio, meaning a measure of filing activity against press attention. The underlying store has a column reserved for that figure, and it is left empty on every filing by design. No source this project has confirmed as free and stable currently supplies a press mention count to serve as its denominator, and this site will not substitute an invented, scraped or estimated one in its place. A ratio with a fabricated denominator would be less honest than no ratio at all, so none is shown.

Per action contract detail was on this list for one day, on the reasoning that a column built from documentation alone is the same class of thing as an invented denominator. The endpoint was then probed and the column built from what it actually returned, which is the difference between the two cases and the only thing that ever separated them.

Search on this site is a static index your browser fetches and filters, not a query against the database. That is a deliberate limit rather than an unfinished feature: every page here is generated once a day and served as a file, so visitor traffic never reaches the store at all, and a search box that broke that property would trade a problem this site does not have for one it would then have. The index is a published file, documented inside itself, if you would rather query it yourself.

Where else to look

This site holds lobbying and federal contract records for the entities named in ORM pieces. It does not yet hold member stock trades or campaign finance. For those, go to the primary sources first: House financial disclosures, Senate eFD, and the FEC.

Two outside projects present those records in a scrollable format. Quiver Quant shows member trades alongside lobbying and contracts. Integrity Index, run by Political Integrity PAC, a registered hybrid committee, shows per candidate finances and trades, and also grades candidates and tracks a candidate pledge. We link them as reference, not as sources; every row on this site cites its own filing.