Find relevant tenders behind vague notice titles

Boundary

Use the title to open the record, not to decide relevance

Begin with two fixed objects: one procurement candidate and one approved capability brief. The capability brief should state a deliverable, the object or process it affects, the buyer outcome, delivery conditions, available proof and explicit exclusions. “Data services” is no better as a capability than “professional services” is as a notice title. “Convert paper planning files into indexed digital records, validate metadata and return an exception log” gives the review something testable.

Copy the title exactly as published, including language, acronym and punctuation. Do not repair it in the evidence field. Add a separate title diagnostic that explains why it may be a poor summary. A title can be broad because the buyer is naming an entire programme, terse because it repeats an internal package code, or precise but misleading because the named technology is only one component of a wider obligation. The diagnostic directs inspection. It never supplies missing scope.

TED’s eForms model makes the distinction concrete. Title and description are separate business terms, and both may occur at procedure, group of lots, part and lot level. The TED search field list also exposes separate procedure and lot fields for titles, descriptions and classifications. A search interface that shows one title has compressed a record with several possible scope layers. Preserve that compression as a known limitation of the candidate view.

Sources

Read scope through a fixed evidence ladder

Resolve the current official record before interpreting it. Capture the portal, notice identifier, procedure identifier, version, publication time, source language and public URL. If a change notice exists, follow its reference to the changed notice and work from the current consolidated text. TED’s change-notice specification permits changes to notice sections and procurement documents, so an earlier description or attachment set cannot be assumed current.

Next, move from broad to specific. Read the procedure purpose and description, then enumerate groups, parts and lots. At each level, collect the title and description without moving a statement to another level. Continue with main and additional classifications, quantities, contract nature and place of performance. These structured fields can confirm the object or location of the purchase, but their meaning depends on the accompanying text. A CPV code that names software does not establish which software task the buyer will award.

Open public attachments only through links exposed by the official record or buyer. Record the document title, URL, date, version, language and the section or page inspected. Technical specifications, schedules and statements of work often settle whether an attractive phrase is a deliverable or context. If the file requires registration, agreement to terms, credentials or a private request, stop and record the access state. Do not describe unread material from its filename.

The Open Contracting Data Standard provides a useful cross-portal mental model. It distinguishes a tender headline from its description, item descriptions, classifications, lots and documents. The SAM.gov opportunities API similarly returns the opportunity title separately from a link to the description, classifications, place of performance and resource links. Field names vary by jurisdiction, but the research question remains stable: which current official field carries the scope claim?

Evidence

Turn passages into scope anchors, including evidence against the match

A scope anchor is a small, attributable unit of evidence. It contains the source level, field or document location, exact passage or structured value, source language, a restrained interpretation and the capability predicate it supports or contradicts. The interpretation must be reviewable. “The notice mentions records” is weak. “Lot 2 requires scanning paper case files, creating indexed PDF files and returning an exception report; these three outputs match the approved conversion capability” explains the link without claiming that every lot or condition fits.

Extract positive evidence and counterevidence in the same pass. A procedure description may mention document conversion while the only relevant lot bundles conversion with twenty years of physical archive storage, which the capability brief excludes. Another notice may use the right code but state that the contract is solely for scanner hardware. Those passages deserve the same prominence as the matching words. Relevance depends on what the buyer will purchase as a coherent obligation, not on the number of semantic similarities.

Separate quoted evidence from analyst interpretation. Preserve the official phrase in its source language, then add a plain explanation. If a translated rendering is needed for review, label it as a translation and keep the original beside it. Do not silently replace a buyer term with supplier vocabulary. Acronyms, internal programme names and technical abbreviations remain unresolved until the notice defines them or another official public source does.

Never infer scope from buyer identity. A transport authority can buy catering, insurance, construction or software. Never infer it from a filename either. “Digital strategy.pdf” might be background material. The source has to state a deliverable, object, outcome or binding boundary that bears on the capability test.

Decision

Five evidence states are more useful than a mysterious relevance score

Assign the result only after recording source coverage and counterevidence. Relevant means current official evidence identifies a deliverable inside the approved capability boundary and no reviewed scope fact contradicts that narrow conclusion. It does not mean the supplier is eligible, has capacity or should bid. Those questions need the documents and a separate qualification decision.

Mixed is appropriate when relevant work is present but bundled with material work outside the capability, split ambiguously across lots, or accompanied by a conflicting scope statement. Adjacent means the buyer problem is related but the purchased deliverable is different. A supplier that provides archive digitisation may find a records-management advisory contract adjacent: the domain overlaps, but the buyer is purchasing advice rather than conversion.

Irrelevant requires affirmative evidence. The reviewed description or document places the purchase outside the capability, and the reason is recorded. Unresolved is not a weak version of relevant. It means a named gap prevents the evidence test, such as a missing lot description, an inaccessible specification, conflicting versions or an undefined acronym. Keep that state until the missing fact is available or the candidate expires.

A numerical similarity score can help order a review queue, but it cannot replace these states. If a score is retained, publish its inputs and use it only as a retrieval aid. The decision must still identify the decisive evidence, counterevidence and missing facts in words.

Hierarchy

A procedure can be relevant while a candidate lot is not

Write the object of the decision into the note. If the search result represents an entire procedure, the result can say that the procedure contains potentially relevant scope and name the lot or part that supports it. It cannot say every lot is relevant. If the candidate already represents one lot, procedure-level language is context until lot text or a linked document assigns the deliverable to that lot.

TED’s current information model places title and description at several levels and gives lots their own place, contract nature and requirements. OCDS uses related lot identifiers so items and documents can be attached to a particular lot; without a related lot, information may apply to the whole tender. These structures prevent a common analytical shortcut: copying one attractive sentence from the procedure into every child lot.

When a broad procedure contains a promising lot, stop the AN-019 task after naming that lot and citing the anchor. The next task is lot-level relevance inspection, which must bring down shared conditions and test the whole lot boundary. When the notice does not expose lots clearly, use mixed or unresolved. Do not invent a commercial division from headings, file folders or repeated codes.

Attachments need the same discipline. A scope schedule linked at procedure level may govern all lots, or only define common terms. A lot-specific pricing sheet may reveal purchased items but not the whole obligation. Record the document relationship stated by the portal or document itself. If that relationship is unclear, describe the evidence without promoting it to a lot fact.

Artifact

Publish a note that another reviewer can reproduce

The final note should be compact enough to review quickly and complete enough to rerun. Begin with identity and provenance: portal, official URL, notice ID, procedure ID, candidate lot or part, version, publication time, retrieval time and source language. Preserve the exact title and the separate title diagnostic. Link the approved capability brief by identifier and version so later edits to the brief do not silently change the old decision.

Add a coverage ledger. Each expected layer receives one explicit value: checked with a source, absent in the official record, inaccessible with the access condition, not applicable, or unresolved. Blank does not mean absent. Then list positive anchors, counterevidence and missing facts. Each row contains its level, locator, evidence, interpretation and capability predicate. Finish with the controlled state, decision reason, narrow next action, owner and refresh trigger.

Consider an illustrative candidate titled “Programme Northstar, package four.” The current procedure description says the programme replaces an estate-wide records platform. Lot 4 says the supplier will scan legacy property files, create indexed records and return an exception register. A public schedule adds permanent off-site storage, which the capability brief excludes, and the notice does not show whether bidders can price scanning separately. The defensible result is mixed, not relevant. The note cites the matching outputs, the bundled storage duty and the missing separability fact, then routes the candidate to whole-lot review.

That example is useful because it can be disproved. A reviewer can open the same sources, inspect the cited locations and challenge the interpretation. “Eighty-seven per cent match” offers no comparable route unless every input and rule is exposed. Prefer a short reason that names the evidence and limit: “Lot 4 buys the three conversion outputs, but storage appears inseparable in schedule B; separability is unresolved.”

Agent handoff

Let agents inspect public evidence, then make them stop at the boundary

This task suits an agent because the inputs, source order and outputs can be made explicit. Give the agent one candidate, one approved capability brief and a permitted public-source list. It may resolve identifiers, follow official public links, enumerate lots, extract exact fields, compare scope predicates and draft the relevance note. Require citations for every non-trivial scope statement and an explicit null state for each missing field.

The agent must not sign in, accept portal terms, request documents, contact the buyer, bypass access controls or infer the contents of an unread file. It also cannot convert relevance into eligibility or a bid recommendation. Stop conditions should be machine-visible: no official notice found, competing current versions, an undefined acronym that controls the scope, a required attachment behind access, conflicting procedure and lot text, or evidence that contains personal or restricted information.

Give the output a stable schema rather than a free-form confidence paragraph. Controlled states, source locators, timestamps and missing-fact fields make the note usable by another agent. Natural-language reasoning still matters, but it sits beside the evidence rather than hiding it. A downstream agent can search only relevant and mixed notes, reopen unresolved ones when a document appears, and exclude irrelevant records using the stored reason.

Refresh on evidence change, not an arbitrary rewrite cycle. Reopen the note when the official notice version changes, the portal marks procurement documents as changed, a missing attachment becomes public, the lot structure changes or the approved capability brief receives a new version. Preserve the old result and time. That history shows why a candidate moved from unresolved to mixed or relevant.