abigailhaddad/dc-procurement-raw

DC procurement: raw source snapshots

Work in progress. These snapshots are being used for an analysis that is still being checked. The files here are raw source material, not findings.

Raw, unmodified responses from the District of Columbia's Contracts and Procurement Transparency Portal (https://contracts.ocp.dc.gov), plus the source documents and other agencies' published files used alongside them.

Nothing here is derived. No classification, no extraction, no OCR output, no joins. Those stay out of this dataset on purpose: they are model output and should not be mistaken for what the District published.

What is here

Path	What it is
raw/	Portal API responses, pulled 2026-09-05
raw_baseline_20260824/	The same endpoints, pulled 2026-08-24
archive/	Every row the portal has ever returned, kept after it stops returning it; refreshed weekly
h1b/raw/	US Department of Labor LCA disclosure files
calc/raw/	GSA CALC API responses (federal awarded hourly rates)
dcas_sources/	Saved source documents for the DCAS procurement history

Each snapshot directory carries a manifest.json recording what was pulled, when, and how many rows each file held.

Two snapshots, on purpose

raw_baseline_20260824 is kept because it cannot be re-fetched. The portal serves current state only, so an August pull is not reproducible from a September request. Twelve days moved real numbers between these two: contract rows 7,113 to 7,168, payments 866,823 to 869,839, distinct contracts 3,983 to 4,013.

The archive, and why it exists

The portal is a rolling window, not an archive. When FY2027 began on 1 October 2026 the contracts search stopped returning 1,297 of the 7,165 option-period rows in raw/: every period of 647 contracts, and the oldest periods of 635 contracts it still lists. Payments in the current table begin on 6 October 2016.

archive/ keeps every row ever pulled. Each file is the portal's response with four fields added and nothing else changed:

Field	Meaning
listed	Whether the latest pull returned the row. Absent on rows last pulled before the archive began; treat absent as listed.
first_listed	First pull that returned it, from October 2026 on.
last_listed	Last pull that returned it.
recovered_on	Set on rows lost on 1 October 2026 and put back: from raw/ (the 2026-09-05 pull) where it held them, otherwise fetched from the portal's detail endpoint, which still answers for delisted contracts. Every period fetched that way that raw/ also holds matched it field for field.

archive/parquet/ holds the same rows as Parquet, one file per table, for the explorer at https://dc-procurement.pages.dev and for anyone who wants to query them directly. Amounts and dates are parsed from the portal's text, and each parsed column sits beside the text it came from; columns.json says where every column comes from. A few columns are not the portal's, and columns.json marks each one: agency names filled in from agency_names.yaml, and on contracts period_kind, base_amount, corrected_amount, same_award_as and correction, from corrections.yaml, and of_interest and of_interest_detail, from of_interest.yaml, both in https://github.com/abigailhaddad/dc-software-procurement. The portal's own values are never overwritten; they sit beside these.

contract_details_older_periods.json holds periods the detail endpoint still serves but the search had stopped listing before this project's first pull. They are kept apart so that contract_details.json stays what the search listed.

Payment ids repeat (879,641 rows, 865,562 ids in October 2026), so do not de-duplicate on id. A payment is kept as delisted only when its id stops appearing at all; when the portal revises a payment, the new rows replace the old ones rather than sitting beside them.

Quirks worth knowing before you use this

The payments endpoint ignores every filter. FiscalYear, PaymentFiscalYear, CommodityCode, AgencyName and ContractNumber were each sent alone and each returned the identical complete table. There is no way to request a slice, so payments.json is the whole table, unpaginated, in one response.

A contract row is an option period, not a contract. 7,168 rows are 4,013 distinct contract numbers. Summing ceilings across rows double counts.

Most payments carry no contract number. 481,031 of 869,839 rows have one. Any join from payments to contracts covers about 45% of the rows.

paymentAmount and voucherAmount are different things. The first is the check total and repeats across every line it covers; the second is the line. Summing the wrong one gives $114.7B instead of $33.5B.

The commodity codes are assigned by the buyer and applied loosely. Class 918, "consulting services", holds Medicaid managed care contracts. Class 920, "data processing, computer, programming and software services", holds a plumbing contractor's $14M contract.

Provenance and license

These are public records published by the District of Columbia, the US Department of Labor, and the General Services Administration. They are reproduced here as fetched, without modification. The District's portal is the authoritative source; where this copy and the portal disagree, the portal is right.