# Attribution and upstream provenance

Every upstream this project touches, what it contributed, its license, and
whether its rows can appear in a published release.

The distinction that matters most in this file, and the one most likely to be
misread, is between a **source** and a **grader**.

- A **source** contributes rows that we publish. Its license reaches our
  output and its attribution obligations, if any, are ours to satisfy.
- A **grader** is an independent parse of the same public documents that we
  compare ourselves against. We publish the *disagreement rate*, never its
  rows. Its license does not reach our output. Several graders are
  unlicensed, and we still cite them, because citing a measurement is not
  redistribution.

If you take a table from this project and redistribute it, you inherit the
obligations in section 2 and nothing from section 3.

**Enforcement is code, not convention.** The exporter carries an allowlist,
`REDISTRIBUTABLE_PROVIDERS`, and every published row carries a `provider` tag.
A release build raises rather than publishing a row whose provider is not on
the allowlist. A second frozen set, `UNREDISTRIBUTABLE_PROVIDERS`, exists only
so the refusal message can name the reason ("this is house_stock_watcher,
whose license field is null") instead of being generic. Being absent from the
allowlist is already disqualifying; the deny list is documentation.

---

## 1. Primary publishers (US government, outside copyright)

These are the publishers of record. Their material is US public record. See
LICENSE-DATA section 1 for why no license is needed and none is granted.

| Upstream | What it contributed | Status |
|---|---|---|
| Clerk of the House of Representatives, bulk `{YEAR}FD.xml` index and disclosure PDFs (`house_clerk`) | The filing denominator, and the `Last-Modified` publication clock that is this dataset's differentiator | Source. US government work. The Clerk publishes no terms of use, no robots.txt, and no license statement; silence is not permission, and the statutory restriction in LICENSE-DATA section 5 binds regardless |
| Our own parse of those PDFs (`house_clerk_ptr`) | Transaction rows extracted directly from the government document rather than from another parser's table | Source. Our extraction of public-record facts |
| Senate Office of Public Records, Electronic Financial Disclosure (`senate_efd`) | Senate PTR index records, e-filed transaction lines, paper scan images | Source. US public record. The site's `prohibition_agreement` checkbox was retrieved in full and is a verbatim recitation of 5 U.S.C. 13107(c) with no license grant and no redistribution clause. Its terms of use and robots.txt have never been retrieved or read by a human on this project; no clearance is claimed on that basis |
| SEC EDGAR filings and DERA structured data sets (`sec_edgar`) | Form 4 transaction lines, 13F-HR information-table lines, acceptance timestamps | Source. US government work |
| SEC fails-to-deliver data (`sec_fails_to_deliver`) | Dated CUSIP-to-symbol crosswalk, the basis for as-of security resolution | Source. US government work |
| SEC ownership forms as a ticker source (`sec_form345`) | Issuer ticker symbols keyed to CIK | Source. US government work |
| Office of Government Ethics, `extapps2.oge.gov` | Executive-branch 278 and 278-T index and PDFs. **Not currently published in any table** | Not yet a source. OGE does make a public-domain statement, and it is narrower than it looks: it covers only information *generated by* OGE, and a filer's own 278-T is generated by the filer. Do not cite it as cover |

## 2. Third-party sources whose rows may be published, and what you owe them

Two permissively licensed third-party parsers of the same public documents.
Their rows may appear in congressional trade projections. **Their notices must
travel with copies and substantial portions of those tables**; that is the
whole of the obligation, and NOTICE satisfies it.

| Upstream | License | What it contributed | Obligation |
|---|---|---|---|
| [`kovagent/congresskit`](https://github.com/kovagent/congresskit) | Dual **MIT OR Apache-2.0**. Both `LICENSE-MIT` (Copyright (c) 2026 userFRM) and `LICENSE-APACHE` are present and `Cargo.toml` declares the dual grant. **We take the MIT arm.** | Year-partitioned Parquet of House trades with a sha256 manifest. Two fields we had open as separate work items: `bioguide_id`, the stable cross-source legislator key, and `notification_date` distinct from `txn_date`, which is the point-in-time input | Retain the copyright and permission notice. Reproduced in NOTICE |
| [`kadoa-org/congress-trading-monitor`](https://github.com/kadoa-org/congress-trading-monitor) | **MIT**, Copyright (c) 2026 Kadoa | House trade rows covering years congresskit's partitions start too late for. Also the executive-branch 278-T rows currently in the lake, which have no verified primary source and are therefore not published | Retain the copyright and permission notice. Reproduced in NOTICE |

**Why the MIT arm of congresskit.** Both arms require attribution and both are
compatible with our Apache-2.0 code license. MIT's obligation is a single
notice-retention condition, which a NOTICE file discharges exactly.
Apache-2.0 additionally requires stating significant changes in modified files
(section 4(b)), an obligation written for source trees and awkward to satisfy
meaningfully for reprocessed data rows. Choosing MIT also makes the obligation
uniform across both third-party sources, since kadoa is MIT-only. Taking the
MIT arm of a dual grant is expressly permitted by the grant.

**A licensed source is still not an unencumbered one.** A row sourced from
either upstream is a row we cannot fix, cannot extend before its earliest
partition, and cannot make more accurate than its parser. That is the argument
for our own parse (`house_clerk_ptr`), independent of licensing.

## 3. Graders: used to measure our correctness, never republished

**Nothing in this section contributes a published row.** These licenses do not
apply to this dataset's output. Do not assume otherwise from their presence
here.

| Grader | License | What we used it for | Why it is grader-only |
|---|---|---|---|
| `TattooedHead/house-stock-watcher-data` | **null** (GitHub license field is literally null, meaning no license at all) | Historical incumbent for House trades; now a comparison corpus only | No license means no right to redistribute. This is the highest-priority exclusion because it is the source behind the historical primary House projection, not a hypothetical. A release build fails on it by design |
| [`jeremiak/us-senate-financial-disclosure-data`](https://github.com/jeremiak/us-senate-financial-disclosure-data) | **CC-BY-SA-4.0** | Senate cross-check | Redistributable in principle, excluded anyway. Share-alike would propagate to the whole published dataset, which would then purport to impose share-alike on public-domain facts. A permissive-looking license is not automatically a compatible one |
| `alicelmre2705/Congress-Trading-Signal` | **none** | Method reference and page census for paper filings | Unlicensed. Method and counts only, never rows |
| `kevinnbass/khanna_occ_complaint` | **null** | Paired page images and machine-transcribed text for 114 paper documents, used in OCR scoring | Unlicensed, so its text must never ship. Also **not ground truth**: the reference text is `gemini-3.1-flash-lite-preview` output on 1,787 of 1,787 pages with zero human transcription, so any CER computed against it is an inter-engine agreement rate, not an accuracy rate |
| Capitol Trades (`capitol_trades`) | **CC-BY-NC-4.0** | Score against | NonCommercial. Excluded |
| congressflow, Seagull | various, not individually established | Score against | Verification only by policy |
| OpenFIGI (`openfigi`) | Bloomberg terms | CUSIP-to-ticker resolution on the non-public trading path | Bloomberg's terms irrevocably dedicate the FIGI *identifier* to the public domain but name "Related Security Descriptions" as a separate category and grant nothing on it. That category is exactly the ticker and issuer name. So a FIGI string may be republished and an OpenFIGI-sourced ticker may not |
| Alpaca (`alpaca`) | Licensed market data | Prices and bars on the non-public path | Licensed data. Prices are permanently out of scope for this dataset |
| `jamiegl/financial-disclosure-scraper` | not established | **Proposed, not yet wired.** Intended as a third grader specifically because the repo was created 2020-10-17 and is therefore genuinely pre-LLM | Would be grader-only. Its value is its date, not its stars: the two graders we have were both created in 2026 and may share design assumptions in a way two independently hand-written parsers would not |

## 4. Byte paths and reference data

Neither a source nor a grader. These constrain *how we obtained* something, or
supply reference data used in measurement rather than in published rows.

| Upstream | Position |
|---|---|
| Internet Archive Wayback Machine | Currently the primary source of *bytes* for archived House paper PDFs, reached via the CDX API. Its terms grant access "for scholarship and research purposes only." This does not change the copyright status of the underlying document, and it is a concrete reason not to make a web archive the distribution source for republished bytes. Retrieval is deliberately routed here rather than at the Clerk, whose rate limiting escalates to IP blocks |
| Common Crawl (`data.commoncrawl.org`) | A separate archive, holding 69 validated pre-2020 paper PDFs retrieved by HTTP Range with no auth. Overlap with the Wayback census has not been computed, so the two counts must not be added |
| End of Term Web Archive | Checked and yielded zero net-new documents. Recorded so nobody re-checks it |
| `unitedstates/congress-legislators` | **CC0-1.0.** Filer roster used to measure survivorship bias in the datasheet. Not shipped in any table |
| GLEIF LEI reference data | **CC0** per its publisher. Evaluated for entity resolution; not currently shipped |
| SEC N-PORT structured data sets | US government work. Used for CUSIP asset-class classification, which is what lets an unresolved CUSIP be distinguished from a genuine equity gap. Not shipped as rows |
| CourtListener / Free Law Project bulk financial disclosures | Judiciary lane. **Not published**: the judiciary is not in `PUBLIC_TABLES`. Its terms of use assert no copyright in the data and restrict only unlawful and FCRA/consumer-report use; that read is from a web archive snapshot because the live terms page returns 403 to our user agent. The underlying AO-10 filings are US government works |

## 5. How to cite this dataset

See `CITATION.cff`. Citation is requested, not required. Nothing in
LICENSE-DATA conditions your use on attributing us, because for the facts we
have nothing to condition it with.

If you cite a coverage figure, cite the release it came from. Coverage
percentages in this project are computed against a government index that is
itself demonstrably a lower bound, so every one of them is an upper bound on
true coverage rather than an estimate of it.