Yellow Bench launches paid beta for Subaward Radar

Yellow Bench has opened a paid beta for Subaward Radar, a post-award research service for European IT resellers and managed service providers. The beta pack costs USD $3.

The Prague-based business says the service is designed to help smaller business-to-business suppliers assess recently awarded public-sector IT contracts in Europe and decide which opportunities merit further manual research. It focuses on post-award analysis rather than live tender alerts and does not present awards as open subcontracting offers or sales leads.

Subaward Radar brings together the legal winner, public buyer, awarded scope, CPV codes, published value and the original Tenders Electronic Daily source, then ranks each record against a supplier's stated profile. The aim is to help resellers, managed service providers and specialist delivery firms rule out weak matches before investing time in company research.

A delayed public snapshot of the dataset through 20 July 2026 covered 1,351 eligible non-defence IT award notices across 33 buyer countries. Of those, 276 had no normalised result value or only a nominal value of 1 or less, while 237 named more than one winner.

The snapshot also showed that 312 records included at least one winner company website supplied by TED, and 137 had a positive processed subcontracting field. Yellow Bench says those figures show why winner names alone do not provide enough information for suppliers judging whether a contract is worth researching.

Scoring model

The first matching model uses a deterministic rather than generative approach. It assigns separate points for CPV overlap, whole capability terms found in the published scope, geography, contract scale, declared subcontracting and freshness.

Each selected result carries a score from 0 to 100 and a visible explanation of why it was selected. Close matches that do not pass the threshold are kept in a rejection log. The paid beta provides five fresh priorities for one capability profile, along with that rejection log and original source links.

Yellow Bench describes the product as a research tool rather than a bid-management system, active-tender alert, personal-contact database or automated prospecting tool. It does not publish personal email addresses or telephone numbers drawn from TED, and defence-related and incomplete records are excluded from the commercial signal set.

Quality check

Yellow Bench says it carried out an independently labelled, score-blinded quality check using 100 previously unseen notices for one multilingual application development and maintenance profile. In that test, the reviewer marked 62 records as relevant and 38 as not relevant without seeing any scores or selected-result flags.

The unchanged hidden key selected 10 signals, all of them relevant, producing what Yellow Bench describes as 100% selected-set precision against a predeclared 80% threshold. The company also drew clear limits around that result, saying it measured the precision of those 10 selected records for one profile only. It did not measure recall, every technology capability, subcontract availability, buyer intent or willingness to pay.

During that labelling work, direct TED pages were blocked by a source web application firewall, so the reviewer used only the published scope fields frozen before scores were disclosed. That matters because Yellow Bench is positioning transparency as part of the service, with users expected to verify the original notice before taking any action.

Beta terms

The delayed award feed, source pages and a no-email capability preview remain free to inspect. Access to the paid beta pack requires a verified business email address, and Stripe shows the final localised price and tax before payment.

The first paid purchase comes with a seven-day refund window under the published terms. Yellow Bench also says the project is bootstrapped and makes no claim about existing customers or revenue.

Subaward Radar is operated by Yellow Bench in Prague. The company says AI tools assisted data processing, software development and language quality checks, while the matching and scoring model remains deterministic.

"The unchanged hidden key selected ten signals, and all ten were relevant, producing 100% selected-set precision against a predeclared 80% threshold."