Pre-submission document verification.
Your bid package is checked against everything the notice actually requires — mandatory documents, names, prices and each stated requirement — with every finding grounded in the source text. How it verifies, how accurate it is, and what it costs: all on this page.
First verification free · 1 pass ₩110,000 = up to 3 runs of one notice until it closes · Buy passes →
Not bidding yet? Start with matching new notices, free by email →
Not the missing form — the proposal that misses the bar.
We measured six months of KONEPS award data: 73,373 bids across 13,220 companies in negotiated contracts (협상에 의한 계약). The weight of failure is not where the paperwork story puts it.
| Cleared evaluation (정상) | 79.96% | |
|---|---|---|
| Failed the evaluation bar (협상평가부적격) | 17.91% | never reached price opening |
| Proposal not submitted (제안서미제출) | 0.60% | |
| Documents not submitted (서류미제출) | 0.35% |
How we know this label is a threshold, not a ranking cut-off
How we know the label means what we say: bidders marked 협상평가부적격 carry no opening rank, while bidders who lost but qualified stay 정상 — and all 5,420 winners in the window are 정상. Controlling for the number of bidders, the count of qualified bidders spreads out (in 4–5 bidder notices: 4 qualified 32%, 3 qualified 23%, 2 qualified 22%), which is a threshold, not a ranking cut-off.
We do not claim your score will rise. Technical evaluation scores are not published — neither you nor we can verify them afterwards. What we show you is which of the notice’s requirements your proposal does not answer, with the source text beside each one.
Source: KONEPS public award data (공공데이터개방표준 낙찰정보), openings 2026-02-11 – 2026-08-10, negotiated contracts only. Counted per bid, not per company.
Requirement extraction
Requirements are pulled from the tender package — tables included — into a structured, reviewable checklist.
Compliance check
Your bid is matched against each requirement. Gaps and deviations surface before submission — not after a rejection.
Mandatory-document check
The required-documents list is extracted from the notice and checked against your package — a missing form surfaces before it disqualifies you.
Cross-document consistency
Company name, representative, and dates are compared across every document in the package — a stale year or a mismatched name gets flagged.
Price consistency
Line items are summed against totals and checked against the stated budget — arithmetic slips and budget overruns are caught before submission.
Grounded, machine-certain findings
Each finding is backed by quotes from the source text. Only high-certainty results are marked certain; the uncertain are set aside for a person to review.
The LLM proposes. A deterministic layer decides.
Miriboa never lets a language model have the last word. Extraction and judgment are proposed by AI, then confirmed by a deterministic layer that scores certainty and grounds every finding in the source.
- 01
Parse
HWP/HWPX, PDF, DOCX and XLSX packages are parsed and normalized — tables preserved, not dropped, including the requirements most parsers lose inside HWP tables.
- 02
Propose, then critique
The AI proposes the requirement set; a critic pass filters it into a structured golden checklist.
- 03
Match, then grade
Each requirement is checked against the bid and graded — machine-certain, or needs-review.
- 04
Report
A pre-submission report lists met, missing, and deviating requirements, each grounded in the source.
Measured in August 2026 on a frozen exhaustive golden set — every document of three real tenders enumerated, 2,013 items (1,985 valid). The tender notices are real and published; the bid packages are synthetic, written for the benchmark. These figures describe benchmark performance and are not a guarantee of results on your own documents.
- 87.4%
- requirement-extraction recall — how completely requirements are found (frozen exhaustive golden set of 2,013 items, 1,985 valid; 92.0% on critical ones)
- 96.5%
- machine-certain precision — how often findings marked "certain" were right
- 18 / 19
- defects caught in the adversarial judgment benchmark
Miriboa is a vendor-operated verification service — send the notice number and your bid package, and the report comes back. Nothing to install, nothing to integrate. Request verification →
Reading a notice is not the same as proving your package meets it — the paid verification cross-checks every document, price and name to produce that proof.The reading half, before that, is free — our Claude connector reads notice documents with tables intact →
Miriboa checks documents against a notice’s stated requirements. It is not legal advice, not a legal determination of eligibility, and not a guarantee of winning or passing evaluation — final responsibility for your submission is yours.
How is this different from other bid-support services?
Alert feeds, award histories and bid-price analytics help you choose tenders and set a price. Miriboa covers the one step after those — checking, before submission, that your finished package meets the notice’s requirements. It works alongside those tools, not instead of them.
How is this different from pasting the notice into a chat AI?
Chat is useful for a draft-level look. But a chat answers differently every run, easily misses requirements inside HWP tables, and can assert things without evidence. Miriboa returns the same verdict for the same documents, quotes the source for every finding, and never asserts what it cannot confirm.
What files do I send?
Your bid package as HWP/HWPX, PDF, DOCX or XLSX — tables are preserved, including inside HWP. For the notice, the 나라장터 notice number is enough; we fetch the package ourselves.
How long does a verification take?
It is a machine judgment, so usually under an hour. A run that ends undeterminable returns the pass.
How is my bid package protected?
An NDA takes effect the moment you submit a request — documents are used only to perform the verification, access is limited to the person performing it, and files are purged within 15 days (immediately on request). See Security & Confidentiality in the footer for the full policy.
What does it catch?
It checks required-document completeness, name/date consistency across files, price arithmetic against the budget, and each requirement in the notice, all with verbatim evidence.
How do passes work, and why three runs?
Your first verification uses the free pass every new account gets; after that you buy passes one at a time (listed prices include VAT). One pass covers a notice up to three times because the point is the fix cycle — the first run finds the gaps, you patch the package, and the re-run confirms it, so the version you actually submit is the verified one.