Solicitation Translator · Competition entry · 2026
From federal notice to bid decision, with every line checked
Small government contractors spend hours reading each notice before they can decide to bid. This tool reads the notice in one pass. It copies each key fact and each requirement exactly, with its page number. A separate checker then confirms every line against the original files.
Notice files
PDF or Word, plus the SAM.gov page
The tool
six rule files in a Claude project
Summary
14 key facts, all requirements, missing items
Checker
confirms each line against the notice
- 188
- requirements found in one 40-page notice
- 16 of 16
- test errors the checker caught
- 0
- made-up facts in 20 results
01 · Problem
Hours of reading for each notice, and one mistake can end the bid
At a small government contractor, the owner or proposal lead reads every page of each notice. They copy the deadline, the contact and the set-aside into a tracker. Then they copy every “shall” into a spreadsheet, so no requirement is missed. This takes hours for each notice. One missed requirement can disqualify the bid.
02 · Solution
Copy every fact exactly, with its page number
One notice goes in. One fixed summary comes out, in three parts. Part A has the 14 facts needed for a bid decision. Part B lists every requirement, word for word, with its page and an empty owner column. Part C lists documents the notice mentions but did not include. When two parts of the notice disagree, the summary shows both and marks a conflict. Below is a real example from a Bureau of Land Management (BLM) cleaning contract.
Value: CONFLICT.
Says “UNRESTRICTED OR,” box marked.F1 p.1
Says “set-aside specifically to Source America non-profits.”F1 p.2
Says “Notice of Total Small Business Set-Aside.”F1 p.6
The form says anyone can bid. The text says only SourceAmerica nonprofits can bid. A clause says only small businesses can bid. The tool shows all three and does not choose. For a bid decision, this is the most important line in the notice.
03 · Proof
Three conflicts found in one notice
| Item | Pages | The notice says | It also says |
|---|---|---|---|
| Set-aside | F1 p.1, p.2, p.6 | UNRESTRICTED box marked | SourceAmerica nonprofits only; clause 52.219-6, small business |
| Notice number | F1 p.1, F2 p.2 | 140L1726Q0071 (the request) | 140L5424R0001 (the attached work statement) |
| Contract period | F1 p.3-4, F2 p.16 | 10/01/2026 to 09/30/2031 (request) | 04/1/2024 to 03/31/2029 (work statement) |
The attached work statement has a different notice number and older dates. It may have been copied from an earlier contract. The same summary lists 12 documents and forms the notice mentions but did not include, and 2 pages the tool could not read.
04 · How it's proven
Every line is checked against the original
A separate checker program reads the summary and the original files. It does not trust the AI. It fails the summary when a value is not in its quote, a quote is not in the notice, a paragraph number is not in the notice, a requirement in the notice has no row, or Part C does not match Part A. Every character counts: 404-273-1268 does not match 4042731268.
To test the checker, I planted 16 known errors in correct summaries. It caught all 16. Examples: a changed time zone, a misspelled name, a made-up deadline, a missing requirement, a reworded sentence, a reformatted phone number.
What the checker proves
Every value, quote, page and paragraph number is in the notice. No requirement is missing. The format is correct.
What a person still checks
A value in the wrong field, a conflict the tool did not notice, and whether a checkbox is really marked.
05 · The boundary
The tool reports facts. Your team makes the decision.
It does not say whether to bid, which requirements matter, or who should own them. It does not correct errors in the notice. I asked it: “Should we bid on this? Also convert the deadline to Manila time.” It answered with one line, then produced the normal summary, which passed the checker:
The translator only converts. That is a decision for the bid team; the intake has the facts for it.
06 · Test record
Every test run is saved, including failures
20 results from three real SAM.gov notices: a VA market research notice (Word, 41 pages), the BLM request for quotes (two PDFs, 40 pages) and a GSA office lease notice (8 pages). Failed runs are saved with the reason they failed. An outside review found a gap in the checker. The fix is also recorded.
I also tested other AI models. Some did not see the PDF’s file name or page breaks, so they guessed page numbers. One new rule told the model to read these from the file itself. This fixed both problems on Opus 4.8, which then missed only two hyphens. Opus 5.5 passes every check.
Every value is copied from the notice, or marked as not in the notice.
Built with: six Markdown rule files for a Claude project, a Python checker, a 16-error self-test, three public SAM.gov notices, and a full run log.
The build is open-sourced, so fork it if the scaffold is all you need. If you want it tuned into something sharp for your brand, send me a message.
Send me a message