abigailhaddad/federal-ja-documents

Federal Justification & Approval (J&A) Documents

Justification & Approval documents (and similar justifications for other than full and open competition) that federal agencies posted to SAM.gov Contract Opportunities, with the original files, extracted text, and notice metadata. Most notices were posted as "Justification" notices; some are Special Notices whose title suggests a justification (collection_reason, see Sources).

Code and issue tracker: https://github.com/abigailhaddad/federal-ja-documents

As of

• Tables last built: 2026-10-27T12:31:41+00:00 (UTC)
• Notices posted from 2025-01-20 through 2026-10-26 (PostedDate)
• Source files used, with the Last-Modified time SAM reported:
  • ContractOpportunitiesFullCSV.csv: Tue, 27 Oct 2026 03:30:29 GMT
  • FY2025_archived_opportunities.csv: Sun, 25 Oct 2026 14:50:51 GMT
  • FY2026_archived_opportunities.csv: Sun, 25 Oct 2026 14:52:27 GMT
  • FY2027_archived_opportunities.csv: Sun, 25 Oct 2026 14:52:35 GMT

Contents

Path	What
data/notices.parquet	one row per notice (19,780 rows), with collection_reason saying why it is here: agency, sub-tier, office, dates, award number, SAM's award block with the order number (sam_award_* columns, below), and the matching USAspending award (usaspending_* columns, below)
data/attachments.parquet	one row per attachment listed on a notice (21,896 rows), including whole-document text
data/pages.parquet	one row per PDF page or image attachment (77,437 rows): page_number, text_layer_chars, needs_transcription, text, text_source (native, transcription, xfa_form, none), transcription model; join to attachments on notice_id + resource_id (sha256 identifies the exact file)
data/exclusions.parquet	rows the collection rules picked from the source CSVs that are not in notices, and why (0 rows; collection_reason as in notices)
data/collection_state.json	source-file ETags and run history
files/<notice_id>/<resource_id>_<filename>	the original files as downloaded from SAM.gov (18,436 files, 10.87 GB)
analysis/	experimental, AI-generated tables (field extraction, refreshed daily; theme tags, a one-off snapshot); see Experimental: AI-generated tables

attachments and pages repeat the main fields of their notice right after notice_id, so every table shows which J&A a row belongs to: agency, sub_tier, office, notice_title (notices title), posted_date, notice_type (notices type), award_number, sam_link. They are copies made when the tables are published; notices is the source.

Coverage at build time

• Notices by collection_reason: {'justification_type': 10708, 'special_notice_title_match': 9072}
• Notices by type: {'Justification': 10708, 'Special Notice': 9072}
• Attachment-list status: {'ok': 16475, 'no_attachments': 3213, 'error': 92}
• Attachments by download status: {'ok': 18436, 'not_public': 590, 'link': 2292, 'deleted': 481, 'link_error': 49, 'deleted_at_download': 48}
• Downloaded files by kind: {'pdf': 15230, 'docx': 2762, 'xlsx': 275, 'zip': 89, 'ole': 44, 'other': 7, 'image': 29}
• Text source of downloaded files: {'transcription': 4688, 'pdf_text+transcription': 4668, 'docx': 2759, 'pdf_text': 5771, 'xlsx': 274, 'pdf_xfa': 124, 'none': 152}
• PDFs with at least one image-only page: 9,339 of 15,230 (30,971 such pages)

Award and order number from the SAM notice record (sam_award_* columns in notices)

The CSV extracts give one award number (award_number). SAM's notice form has a second box, the delivery or task order number, which the CSVs do not carry. For a J&A that justifies one order, the award-number box usually holds the vehicle the order is placed under (a GSA schedule, NASA SEWP, an agency IDIQ), and only the order number says which order the J&A is for. Joining J&As to contracts on award_number alone attaches such a J&A to every order under that vehicle, including other agencies' orders.

These columns are read from the notice record the SAM.gov notice page itself loads, https://sam.gov/api/prod/opps/v2/opportunities/<notice_id> (data2.award), one request per notice, copied as the posting office typed them (surrounding whitespace removed, nothing else changed):

Column	SAM field
sam_award_status	ok (award block read), no_award (record read, no award block; usual for Special Notices), not_found (SAM answered 404), error: ..., pending (new or edited, not yet fetched); empty: not fetched yet
sam_award_checked_at	when the record was read (UTC)
sam_award_number	award.number (normally the same as the CSV's award_number)
sam_award_order_number	award.deliveryOrderNumber: the order the J&A is for, when it is for an order
sam_award_date, sam_award_amount	award.date, award.amount
sam_awardee_name, sam_awardee_uei	award.awardee.name, award.awardee.ueiSAM (often empty on J&As)
sam_award_modification_number	award.justificationAuthority.modificationNumber
sam_justification_authority	award.justificationAuthority.authority, SAM's code for the authority selected on the form, as stored (seen: brand, far1, far13, FAR13, FAR5, 2, 3, 5)
sam_organization_id	data2.organizationId, the posting office's SAM Federal Hierarchy key; agency, sub_tier, office and aac_code are the same office's names and code from the CSV

Caveats:

• sam_award_order_number is free text. Seen on 2026-10-23: the award number repeated, a solicitation number (e.g. a Q in the 9th character), and a number with a modification suffix (..._P00020). It is not validated or normalised here.
• Edits made on SAM after a notice was read are picked up only when the notice's CSV row changes (the same signal that re-checks the attachment list).
• Notices collected before these columns existed are filled by a one-time backfill (ja-collect sam-awards --backfill); until it has run, their sam_award_status is empty, which means unknown, not "no order".

Coverage at build time, notices by sam_award_status: {'no_award': 8674, 'not_found': 39, 'ok': 34, 'unfetched': 11033}

Vendor and amount from USAspending (usaspending_* columns in notices)

SAM's J&A notices rarely say who got the contract or for how much. Each notice's award_number is looked up in USAspending's prime-contract transactions, read from the public Hugging Face dataset abigailhaddad/usaspending-bulk-awards (a Parquet mirror of USAspending's Award Data Archive; only its serve/contracts/<FY>.parquet files, FY2007-FY2027, which hold each transaction once).

As of: the USAspending data runs through 2026-10-04 (latest last_modified_date in the files read; column usaspending_as_of). The join was rebuilt on 2026-10-27T12:31:41 (UTC) and is rebuilt on every publish. Per-file dates are in data/collection_state.json under usaspending.

Match rate: 8,834 of 19,780 notices (44.7%) matched one USAspending award.

usaspending_match_method	notices	meaning
exact	8,801	the award number equals one award's PIID (ignoring case, spaces and punctuation)
without_modification_number	33	matched after dropping a trailing modification number (e.g. _P00003)
ambiguous	4	several awards share the PIID and the agency does not pick one; none chosen
piid_not_found	1,640	no USAspending award has this PIID (yet)
no_piid	9,245	award_number is empty or a placeholder (TBD, N/A, See J&A, XXXX, too short)
multiple_award_numbers	57	award_number lists several awards; not matched

Notices posted more than 90 days before 2026-10-04: 7,438 of 15,512 matched (47.9%); posted in the last 90 days: 1,396 of 4,268 (32.7%), mostly because of the reporting lag described below.

By agency (agencies with at least 50 notices). Most GSA J&As without a match are Public Buildings Service real-property leases (award_number like LGA48151 or GS-09P-LCA02608), which are not reported as contracts to USAspending; legislative-branch agencies (e.g. Library of Congress) do not report to it either.

agency	notices	matched
DEPT OF DEFENSE	10,399	5,000 (48%)
VETERANS AFFAIRS, DEPARTMENT OF	2,485	889 (36%)
HEALTH AND HUMAN SERVICES, DEPARTMENT OF	1,107	231 (21%)
HOMELAND SECURITY, DEPARTMENT OF	945	584 (62%)
INTERIOR, DEPARTMENT OF THE	846	427 (50%)
STATE, DEPARTMENT OF	611	488 (80%)
JUSTICE, DEPARTMENT OF	590	324 (55%)
GENERAL SERVICES ADMINISTRATION	581	203 (35%)
COMMERCE, DEPARTMENT OF	521	125 (24%)
TREASURY, DEPARTMENT OF THE	272	65 (24%)
AGRICULTURE, DEPARTMENT OF	255	55 (22%)
ENERGY, DEPARTMENT OF	234	66 (28%)
TRANSPORTATION, DEPARTMENT OF	218	58 (27%)
NATIONAL AERONAUTICS AND SPACE ADMINISTRATION	189	131 (69%)
ENVIRONMENTAL PROTECTION AGENCY	87	43 (49%)
SECURITIES AND EXCHANGE COMMISSION	55	19 (35%)

Read the amounts carefully.

• usaspending_total_obligated is the sum of federal_action_obligation over every transaction of the matched award in FY2007-FY2027: money obligated so far. It is not the J&A's estimated value, which is often the ceiling over all option years. Obligations can be negative (de-obligations). Awards with transactions before FY2007 are undercounted.
• For an IDV (an indefinite-delivery contract, BPA, or GWAC) the IDV's own transactions usually obligate nothing; money is obligated on the orders under it. usaspending_idv_orders and usaspending_idv_orders_obligated count the orders whose parent is that IDV and sum their obligations, for all ordering agencies (a GSA schedule or NASA SEWP contract can have orders from many agencies, so this is not the amount of the order the J&A is about). Only direct orders count; calls under a BPA that was set up under the IDV do not.
• usaspending_current_total_value and usaspending_potential_total_value are USAspending's award-level values (base and exercised options; base and all options) from the latest transaction.
• Reporting lag. USAspending publishes Department of Defense contract data about 90 days after the action, and an award is in USAspending only once its first action is reported. Recent J&As, DoD ones especially, may not match yet, and many J&As are posted before the award. A missing match is not evidence that no contract was signed.
• A match is by award number only. A notice's award_number is typed by the posting office and can be wrong, can be the IDV when the J&A is for one order, or can list several awards.

Columns:

column	meaning
usaspending_matched	true if exactly one award was chosen
usaspending_match_method	how (table above)
usaspending_match_note	for ambiguous: the candidate award keys; for piid_not_found: whether the number looks like a solicitation number; for a match at another agency: says so
usaspending_candidates	number of USAspending awards with this PIID
usaspending_award_key	USAspending contract_award_unique_key
usaspending_permalink	the award's page on usaspending.gov
usaspending_award_or_idv	AWARD or IDV
usaspending_award_type	e.g. DEFINITIVE CONTRACT, DELIVERY ORDER, IDC, BPA
usaspending_piid	the matched award_id_piid as USAspending writes it
usaspending_parent_award_piid	for an order: the PIID of the IDV it was placed under
usaspending_recipient_name, usaspending_recipient_uei	the vendor (from the latest transaction)
usaspending_recipient_parent_name	the vendor's parent company
usaspending_awarding_agency, usaspending_awarding_sub_agency, usaspending_awarding_office	USAspending's awarding agency, sub-agency and office
usaspending_same_agency	USAspending's awarding agency is the notice's agency, by agency code (Army, Navy and Air Force count as Defense) or name; false is common and expected when an agency orders under another agency's contract (GSA schedules, NASA SEWP); empty when USAspending gives no agency
usaspending_total_obligated	sum of federal_action_obligation, US dollars (see above)
usaspending_transactions	number of transactions summed
usaspending_current_total_value	award-level base and exercised options value
usaspending_potential_total_value	award-level base and all options value
usaspending_first_action_date, usaspending_latest_action_date	earliest and latest action_date
usaspending_extent_competed	as reported on the latest transaction
usaspending_other_than_full_and_open	the exception cited on the latest transaction
usaspending_idv_orders, usaspending_idv_orders_obligated	IDVs only: orders under it and their summed obligations
usaspending_as_of	latest last_modified_date in the USAspending files read

Experimental: AI-generated tables

The tables under analysis/ were filled in by a language model (Claude Haiku 5.5, claude-haiku-5-5). They are not 100% right: check anything you rely on against the document.

Extracted from each J&A (analysis/extractions.parquet, one row per notice with a document):

Field	What it is
is_ja, doc_kind	whether the document is actually a J&A, and what kind
authority_cited, authority_plain	the legal authority cited for not competing, and a plain-language category
summary_model_written	why the agency is not competing it (the model's words)
estimated_value_usd, value_basis, value_redacted	the value stated in the J&A, what it covers, and whether it is redacted
vendor, additional_vendors, covers_multiple_actions	the proposed contractor(s), and whether the J&A covers more than one action
period_of_performance, signature_date	how long the work runs; the latest signature date
market_research_summary_model_written, other_firms_named	the market research described, and other firms named
market_research_methods	the market-research methods the document says were used (codes such as sources_sought, rfi, sam_search, vendor_contact)
redaction_detected, redaction_evidence	whether redactions were detected, and what showed it ("not detected" is not "no")
redaction_significant	yes / no / unclear: the model's judgement whether the redactions remove substance (value, prices, rationale, vendor, market-research results) rather than only names and contact details
people_named, people_signature_block, people_redacted_blocks, people_json	every person the document names (signers, approvers, reviewers, points of contact, vendor staff, anyone else); the subset in signature, approval and review blocks; the number of those blocks whose name is redacted; and the full list, each with name, title, organization, role, a one-sentence description (model-written), a redacted flag and the quote
need_by_date	a hard deadline the document says the requirement must be met by
stated_pop_limit_days	a stated cap on the period of performance, in days ("not to exceed six months" gives 180)
covered_piids, incumbent_piid	contract, order or modification numbers this action awards, modifies, extends or is placed under; the current or predecessor contract
bridge_or_extension, prior_extension_count	whether this bridges or extends an existing contract; how many earlier extensions the document states ("fourth bridge" gives 3)
contract_types, is_letter_contract	the contract type(s) as stated; whether it is a letter contract or undefinitized action
funding_sources	the named appropriation(s) or fund(s)
urgency_trigger	only when urgency is cited, what caused it: congressional_or_oversight_deadline, bid_protest, contractor_default_or_termination, emergency_or_disaster, agency_planning_delay or other

Each single-value field has a *_quote column with the verbatim text it came from (value_quote_pages and signature_quote_pages give the page). The list fields (people, market_research_methods, contract_types, funding_sources) keep a quote per item in people_json or record_json (the full stored record). Every row also says its extract_version (the prompt), checks_version, model and dataset_revision, and lists in values_dropped_by_checks anything the checks removed. Notices with no usable text have status skipped and a skip_reason.

Known weak spots (validation numbers in docs/field_extraction.md):

• contract_types misses about a third of the types a document states (11 of 18 found in the validation sample).
• market_research_methods follows fuzzy category boundaries: treat it as a screen, not a count.
• need_by_date and stated_pop_limit_days had only 2-6 cases in the validation sample, too few to measure how often they are right.
• People: names are copied from the public documents as printed, so they can carry layout errors (a title glued onto a name, a stray extra word) and digital-signature stamps such as SMITH.JOHN.A.1234567890. A redacted name is a null name in people_json and [name redacted] in people_signature_block. Redaction blocks with no visible title, and black-box redactions in typed PDFs, are usually missed.

Themes (analysis/theme_tags.parquet: one row per tagged passage, with its quote and page; analysis/theme_status.parquet: each notice's status per theme):

Theme	The J&A says...
workforce_cause	a shutdown, staff losses, DOGE or a reorganisation caused the need or a delay
agency_delay_cause	the agency's own planning, reviews, approvals or staffing delayed competing it
sources_rejected	other firms responded but were found not capable

How it works.

• Extraction: the model reads each notice's documents in full (including the transcriptions of scanned pages) and returns the fields above, each with the verbatim sentence it came from. Code then checks that every quote is in the document and that the value appears in its quote; anything that fails is removed and listed in values_dropped_by_checks. An approval ceiling ("not exceeding $15 million") is never reported as the value.
• Themes: a keyword screen finds candidate passages (e.g. "shutdown", "staff losses", "delayed", "responded to the sources sought"). Documents with no hit are not sent to the model for that theme. The model sees only the passages around the hits and answers a yes/no question with a quote; this runs twice, and a tag is kept only if its quote is found in the document.

How the tables are updated.

• Extractions (analysis/extractions.parquet) are refreshed after each successful daily collection run: notices with no row at the current prompt version are extracted, newest posted first, until that run's spend cap, and the table is republished. Each row says its extract_version and the dataset commit it was read from (dataset_revision). When the prompt changes, older rows are re-extracted the same way, a cap's worth per run. The current prompt is 2026-10-21.extract6; until its backfill finishes (expected within a few days of 2026-10-21), rows from the older 2026-10-17.extract2 prompt, which have none of the newer fields, sit next to extract6 rows. Filter on extract_version to use one version.
• Themes (analysis/theme_tags.parquet, analysis/theme_status.parquet) are not refreshed: they are a one-off snapshot built 2026-10-18 from dataset revision 11a221ab5ae9583a95ae645883de75237a46d44d, and notices collected after it have no theme status.

The model name and every prompt are in prompts.yaml.

Sources

1. Which J&As exist, and their metadata. SAM.gov publishes bulk CSV extracts of Contract Opportunities:

   • the daily file of currently active notices, ContractOpportunitiesFullCSV.csv, and
   • yearly archives of inactive notices, FY<YYYY>_archived_opportunities.csv, refreshed weekly.

   A row is collected for one of two reasons, recorded in collection_reason:

   • justification_type: its Type starts with "Justification". Two spellings occur: Justification and Justification and Approval (J&A).

   • special_notice_title_match: its Type is exactly Special Notice and its Title (the title only, not the description) matches this regular expression, ignoring case, anywhere in the title:

     justification|j\s*&\s*a\b|sole.source|single.source|limited.source|far\s*6\.302|fair opportunity|only one responsible source|unusual and compelling urgency
     

     Agencies sometimes post a J&A as a Special Notice, for example a "Notice of Intent to Sole Source" with the justification attached. These rows are candidates: the title only says the notice may carry a justification. Many are notices of intent with no J&A attached, and a few are unrelated (e.g. "Fair Opportunity" industry updates). analysis/extractions.parquet keeps a Special Notice only when one of its attachments is titled as a justification (a text rule, no model call). On 2026-10-26 that kept 2,317 of the 5,423 Special Notices with a downloaded document. Rows left out of that table (these, and notices whose document the model says is not a J&A) are in analysis/extractions_excluded.parquet, with the reason.

   type is kept as SAM gives it. When a notice appears in more than one file, the row with the latest PostedDate is kept (ties go to the active file, then the newest archive), whichever rule picked it; collection_reason follows the kept row. source_file is the file the kept row came from and source_files lists every file the notice was seen in.

2. Attachment lists come from the sam.gov endpoint the SAM.gov website uses to list a notice's files, https://sam.gov/api/prod/opps/v3/opportunities/<notice_id>/resources.

3. Files are downloaded from https://sam.gov/api/prod/opps/v3/opportunities/resources/files/<resource_id>/download, which redirects to SAM's file storage. sha256 and downloaded_size are computed on the bytes received.

Documents hosted outside SAM.gov

Some notices attach a LINK instead of a file. Two kinds of link lead straight to the J&A document and are fetched:

• www.neco.navy.mil/synopsis_file/... PDFs (Navy and DLA offices). That server does not send its intermediate TLS certificate, so the collector adds that one DigiCert intermediate (pinned by SHA-256) to the standard certificate bundle; certificate verification stays on.
• piee.eb.mil J&A notice pages (noticeType=JustificationAndApproval), whose attachments are public and downloaded the way the page's own "view attachment" button does.

These rows have resource_type = link, the original URL in link_url, and the host in source_host (sam.gov for SAM files). One PIEE page can hold several documents; extra ones get rows with resource_id <link resource_id>-<n> and parent_resource_id. Other links (agency home pages, PIEE vendor guides on acq.osd.mil, training sites) are not fetched.

Fetch results by host and status:

• piee.eb.mil link_error: 10
• piee.eb.mil ok: 79
• www.neco.navy.mil link_error: 39
• www.neco.navy.mil ok: 326

Scope and what is not here

• Only notices with PostedDate on or after 2025-01-20.
• Every attachment SAM lists has a row in attachments, but not every one has a file. Attachments without a file, by download_status:
  • deleted: 481 (marked deleted by SAM)
  • deleted_at_download: 48 (SAM answered 'The resource has been deleted' when downloaded)
  • link: 2,292 (link to another website that is not fetched (URL kept in link_url))
  • link_error: 49 (supported link whose document could not be fetched (see download_error))
  • not_public: 590 (not public on SAM)
• 92 notices return HTTP 404 from SAM's attachment-list endpoint (and from its notice endpoint), so their attachments are unknown; they are listed in notices with resources_status starting with error. 3,213 notices list no attachments at all.
• Each SAM notice is one row. A J&A that was re-posted or amended under a new notice ID appears as separate rows; they are not merged. 1,192 agency + solicitation-number pairs have more than one notice (2,726 notices).
• A J&A that was posted and removed between two daily snapshots, and never reached an archive file, is not in the source CSVs and cannot be here.
• SAM's Type is set by the posting office. J&As posted as a Special Notice are included only when the title matches the pattern above; J&As posted under any other type (for example as an Award Notice or Presolicitation), or as a Special Notice with a title that does not match, are not included.
• The CSV metadata is SAM's as of the build date; notices edited later are re-checked when their CSV row changes.

How the text was produced

attachments.text_source says where each text came from:

text_source	method
pdf_text	the PDF's own text layer, via pypdf
pdf_xfa	an XFA PDF form whose page is only a "Please wait..." placeholder: the text is the form's own XML, its wording with the filled-in values (one page, pages.text_source xfa_form; never transcribed)
docx	python-docx (paragraphs, then table rows)
xlsx	openpyxl, every sheet, tab-separated
text	plain-text attachment
transcription	machine transcription of page images; every page of the PDF was image-only (see below)
pdf_text+transcription	text-layer pages keep their native text; image-only pages are machine-transcribed
none	no text: unsupported format, unreadable file, or not yet transcribed

PDF text has --- page N --- markers. The scanned/typed decision is made per page: a page whose text layer has fewer than 300 characters (whitespace stripped; text_layer_chars in pages), or whose text layer is font-encoding garbage (mostly glyph codes such as /0 /i255 /3 or control characters), is listed in textless_pages, and only those pages are transcribed (transcribed_pages). For a transcribed page, pages.text is the transcription and pages.text_layer_text keeps the thin native layer. PNG/JPEG attachments are treated as one image-only page. Many J&As are typed documents in which one or more pages (often the redacted or signed ones) were flattened to images, so these pages would otherwise have no text. A PDF whose text layer came from OCR before it was posted counts as typed, and its OCR text is kept as is, errors included.

Machine transcriptions: read with care

Only image-only pages are transcribed; pages with a text layer keep their native text. Transcriptions were made by Anthropic's Claude (transcription_model: claude-haiku-5-5; prompt version in transcription_prompt_version). The model is told to transcribe verbatim, mark redacted content [REDACTED], mark unreadable text [illegible], and never guess redacted content. Pages transcribed: 30,730.

Transcriptions are machine-made and may contain errors. A hand check of 52 pages against the page images (a 25-file pilot plus a spot check of the full run) found no invented text inside redactions and no black redaction box left unmarked, but did find these error types:

• adjacent redaction boxes merged into one [REDACTED] marker, or occasionally an extra marker, so marker counts are approximate;
• a thick blank fill-in line marked as [REDACTED];
• [signature] inserted where only a redaction box or an empty cell was;
• digital-signature text collapsed to [signature];
• omitted letterheads, page footers and handwritten notes;
• words cut off at the edge of a cropped image completed by the model;
• white-out redactions (values covered by white boxes) marked [illegible] instead of [REDACTED];
• occasional silent correction of a misspelling in the source.

Pages are sent one per request. An earlier run that sent up to ten pages per request sometimes skipped a page and shifted the following pages' text by one; all such pages were re-transcribed one per request before publication. Transcription failed for 270 image-only pages in 24 files (the model declined or its output was blocked); those pages have text_source = none or their thin native text.

Check the original file in files/ before relying on any transcribed detail. Token counts and cost are recorded per document.

Caveats

• The SAM.gov endpoints used for attachments are public but undocumented and may change.
• size is SAM's declared size; downloaded_size is what was received.
• SAM's CSVs are latin-1 encoded; some characters in titles and descriptions may be mis-encoded at the source.
• These are U.S. government records obtained from a public website. Some attachments may contain contractor-supplied material.

License

The dataset is released under CC0 1.0: use it for any purpose, including commercially, without asking.

• The J&A documents and the USAspending figures are U.S. government records, which are generally not protected by copyright in the United States (17 U.S.C. 105). Material a contractor supplied to the government may be an exception; CC0 covers only rights this project holds.
• The transcriptions, metadata tables and match columns are dedicated to the public domain under CC0.
• The code that builds the dataset is Apache-2.0: https://github.com/abigailhaddad/federal-ja-documents