What 17.47 million bids say — six months of bid-opening data, observed
One pass over the KONEPS bid-opening data Miriboa uses for verification and alerts (the procurement agency's award-information API; construction and services; February–August 2026). Three observations, reported as the data shows them — not market interpretation — with the data's own caveats (empty days, unconfirmed awards, collection timing) printed next to the numbers.
Observation 1 — the shape of competition is set by the contract method
| Type | Award method | Notices | Median bidders | Top 10% | Max | Single bidder |
|---|---|---|---|---|---|---|
| Construction | Small-sum quotation | 46,500 | 56 | 297 | 6,224 | 0.2% |
| Services | Negotiated contract | 38,110 | 1 | 4 | 35 | 54.4% |
| Services | Small-sum quotation | 30,429 | 8 | 57 | 2,799 | 8.5% |
| Construction | Qualification screening | 17,327 | 367 | 1,209 | 11,361 | 0.6% |
| Services | Qualification screening | 16,086 | 15 | 89 | 3,557 | 11.6% |
| Services | Direct negotiation | 15,530 | 1 | 1 | 2 | 100% |
Both are "services", yet negotiated contracts get a single bidder half the time while qualification screening draws a median of 367 bidders per notice. Bidder count is a function of the method, not of market interest. Any source quoting one "competition ratio" has collapsed this table.
Observation 2 — bids crowd the 1-point band above the floor, and nearly half fall below it
Qualification screening and small-sum quotations have a lowest-price floor: a bid below a set percentage of the reserve price is disqualified automatically. We measured how far each bid sits from that floor.
| Type | Bids | Below floor | Within 1 pt above |
|---|---|---|---|
| Construction · qualification | 9,398,745 | 22.0% | 41.3% |
| Construction · small-sum | 6,091,176 | 25.2% | 40.7% |
| Services · qualification | 772,795 | 42.1% | 41.0% |
| Services · small-sum | 749,875 | 42.4% | 40.6% |
In 0.1-point bins the 0–1 point band above the floor is nearly flat (3.4–4.4% per bin): bids fill the band evenly rather than cluster on one value. The reserve price is drawn by lot from several candidate prices at opening, so the floor cannot be known in advance. The 22–42% that land below it are the outcome of the draw, not a failed strategy.
Of 17,467,350 bids, 8,268,020 (47.3%) were disqualified, and 98% of those for being under the floor. All other grounds — missing documents, qualification score, negotiation evaluation — add up to 2%. Almost every elimination is price, not paperwork; paperwork failures are rare, which is why they are entirely the bidder's own (see the H1 elimination report).
Observation 3 — award ratios collapse to a point per method
Confirmed awards only (award price as % of the reserve price).
| Type | Confirmed awards | Median ratio | P10 – P90 | Most common 0.5-pt bin |
|---|---|---|---|---|
| Construction · qualification | 13,076 | 90.33 | 89.39 – 90.72 | 90.5 (56%) |
| Construction · small-sum | 42,437 | 90.33 | 90.00 – 90.75 | 90.5 (57%) |
| Services · qualification | 11,826 | 87.76 | 81.15 – 88.89 | 88.0 (41%) |
| Services · small-sum | 25,351 | 88.24 | 88.01 – 90.41 | 88.0 (51%) |
| Services · direct negotiation | 14,321 | 97.55 | 89.04 – 100 | 100 (30%) |
For construction qualification screening and small-sum quotations, 74–89% of award ratios fall in the 90.0–90.5 band. The award price is effectively fixed near 90% of the reserve; the contest is not how low you go but who lands lowest inside the 1-point band above the floor.
Read with the numbers — the data's own caveats
- 109 empty days. Of 439 daily files, 109 hold zero rows — mostly weekends and public holidays. Two are weekdays (construction 30 Jul, services 11 Aug): not holidays, so we suspected a collection failure, and once the API came back on 7 September we asked. The source has openings on both days (construction 30 Jul: 90,499 bid rows; services 11 Aug: 10,072). Both days are missing from the numbers in this report. Read them as "collection failure", not "none".
- Collection closer to the opening date captures less. The award-information API returns only notices whose award is confirmed. March–May were collected in August; July–August within the month. July–August looking smaller than March–April is collection timing, not the market. This report makes no monthly-trend claim.
- 17.8% of rows have no confirmed award — opened but not yet awarded. The ratio table counts confirmed awards only. Other fields are missing in at most 1.2% of rows.
Limits
- Construction and services only; goods and foreign procurement were not collected.
- One six-month window: no seasonality or trend can be stated.
- Observation, not interpretation. The crowding above the floor is explained by the rules (reserve-price lottery, lowest-price floor); the data says nothing beyond that.
- Agency and winner concentration were computed but not published — they would expose individual firms. All figures are aggregates.
Method and files
Source: the procurement agency's open-data award-information service (getDataSetOpnStdScsbidInfo), construction (3) and services (5), collected in daily windows. One row = one bid by one firm on one notice. Notices are distinct notice numbers (re-announcements with a different round, 4.8%, share the number).
Award and bid ratios are % of the reserve price. Floor gap = bid ratio − lowest-price floor (sucsfLwstlmtRt), for methods that have a floor. Two pure-Python passes of about five minutes each. Scripts and result JSON are published below; the 17,467,350 raw rows are public data available from the same API.
Files
- observe.py — scale, months, methods, bidders, award ratios, disqualification, re-announcements, missing fields
- observe2.py — collection dates per file, floor distribution, bid-to-floor gap histogram
- observations.json · observations2.json — results
- README — reproduction, field definitions, limits (Korean)
2% of eliminations are paperwork. That 2% is catchable before you submit.
Price is a lottery; paperwork is not. Missing required documents, price-vs-budget arithmetic and name mismatches are judged by machine against the notice text, with verbatim evidence.