How to develop and assess source discovery in AI-enabled work

Begin with a discovery question

The UK National Screening Committee's evidence review process says searches should answer questions identified during scoping and should document databases, dates, limits, results, inclusion criteria, quality checks, and limitations. The U.S. Government Accountability Office's 2024 Government Auditing Standards likewise directs auditors to identify potential information sources and determine the amount and type of evidence needed for the audit objective.

For ordinary knowledge work, the useful principle is the same: define what must be found before choosing a search tool. A request such as "research AI adoption" is not yet a discovery question. "Find current records that can show whether completion changed after the support workflow was introduced, for this programme and reporting period" gives the search a population, event, measure, and date boundary.

Use the ROUTES discovery record

ROUTES is a six-part record that turns an open-ended search into an inspectable retrieval plan.

| Field | Discovery decision | Record to retain | | --- | --- | --- | | Required proposition | What fact, rule, comparison, or uncertainty must be resolved? | One bounded question per row | | Origin locations | Which systems or publishers are most likely to hold the original evidence? | Repositories, databases, record owners, and official sites | | User vocabulary | Which terms do people, policies, systems, and specialists use for the same concept? | Tested names, synonyms, identifiers, and date formats | | Traversal paths | How will the search expand beyond direct queries? | References, citations, document families, owners, and related records | | Eligibility boundary | What qualifies for inspection, and what is out of scope? | Inclusion and exclusion rules | | Stop and gap rule | What ends this search, and what missing evidence must remain visible? | Coverage test, cutoff, unresolved gaps, and next owner |

Complete the required proposition first. Then list likely origin locations before opening a general search engine or assistant. This reduces the risk that a convenient web result displaces an internal system of record, current policy source, primary dataset, or responsible document owner.

Search in layers

Use three layers rather than one long prompt.

First, search known systems and authoritative publishers. Use exact identifiers, controlled vocabulary, version labels, dates, and document types. If the task refers to a named programme, contract, policy, or dataset, begin with the place that creates or maintains that object.

Second, broaden the vocabulary. Search alternative names, acronyms, former titles, related outcomes, and terminology used by other functions. Record which term revealed which source. AI can help propose vocabulary, but each proposed term remains a lead until the worker verifies what it means in context.

Third, traverse the evidence lineage. Follow citations back to origins, inspect references used by a strong source, look for companion appendices or data dictionaries, and search forward for corrections or superseding versions. This layer often finds material that keyword search misses because the decisive source uses unfamiliar language.

Each layer should produce a short trail. Keep the query, location, date, useful results, exclusions, and next move. The trail is not an exhaustive browsing history. It is the evidence needed to explain why the final source set was reachable and why important gaps remain.

Worked example: finding evidence for a service decision

Suppose an operations lead must decide whether to change an internal support handoff after response times worsened. This example is synthetic and does not describe a Two Prune customer or measured outcome.

The initial AI search returns articles about customer-service automation and a vendor benchmark. Those results are topically related but do not answer what happened inside the service. The lead creates three required propositions:

For the first proposition, the origin locations are the ticket system and its metric definition. The lead searches for the dashboard field, underlying export, timezone rule, excluded ticket states, and denominator. For the second, the likely origins are the approved procedure, change record, and team announcement. For the third, the search expands to staffing rosters, incident logs, routing configuration, and seasonal volume records.

The vocabulary step reveals that the ticket system calls the measure "first public reply" rather than "response time." The traversal step finds a configuration change linked from an incident record, while a reference in the handoff procedure points to an older approval that has since been superseded. The search trail records both discoveries and excludes the vendor benchmark from the causal analysis because it cannot describe this service.

The lead stops when every proposition has at least one inspected origin record or an explicit gap, the relevant period has been covered, and further searching is only repeating known lineages. The resulting artifact contains the ROUTES record, search trail, retrieved-source register, version notes, and two unresolved gaps. It supports the next analysis without pretending the search proved why response times changed.

• Response time changed during the stated reporting period.
• The handoff workflow changed before the observed shift.
• Another material operational change could explain the same pattern.

Practice discovery, not result collecting

A useful exercise gives a learner a decision brief with several possible evidence locations: an inbox, document repository, spreadsheet, policy library, public website, and an AI research tool. Ask the learner to write the ROUTES record before searching, then compare the plan with the actual search trail.

Change one constraint and repeat. A question about the current rule should shift effort toward the issuing authority and version history. A question about implementation should shift toward operational records. A question about stakeholder experience should add direct feedback and sampling boundaries. The exercise teaches that discovery routes follow the proposition rather than a fixed list of preferred websites.

Another drill is the missing-source challenge. Remove one decisive record from the pack. The learner should recognize the gap, search plausible origins, and narrow the conclusion if the record cannot be found. Inventing a source, substituting a secondary summary, or treating absence as evidence should lower the assessment.

Assess observable discovery behavior

A realistic assessment can provide a mixed workplace environment with incomplete filenames, competing versions, derivative summaries, and one unavailable record. Ask the participant to produce a decision-ready note plus the search record used to find evidence.

Look for evidence that the participant:

Do not score the skill by query count, browsing time, or number of saved links. A compact search can be strong when the proposition and origin location are clear. A long search can be weak when it repeatedly retrieves the same lineage or never checks the system of record. Retain the ROUTES record, search trail, opened-source history, retrieved artifacts, and gap log as assessment evidence.

• converts the task into bounded discovery questions;
• searches likely origin systems before relying on general summaries;
• varies vocabulary when the first query is too narrow;
• follows references and version lineages rather than counting repeated claims;
• records inclusion, exclusion, and cutoff decisions;
• distinguishes no result from evidence that something did not occur;
• exposes unresolved gaps and adjusts the conclusion accordingly.

Verify the discovery trail

Before delivery, take each material conclusion and trace it to a required proposition, then to the source-discovery route that produced its evidence. Confirm that the relevant origin locations, date range, alternate vocabulary, and version family were considered. Re-run the highest-value query and open the saved sources to confirm they are still the records described.

Then perform a negative-space review. Ask which source type could overturn the conclusion, where it would normally live, and whether that location was searched. If the answer is unknown, preserve that as a limitation or route it to a named owner.

Source discovery creates a defensible path to evidence. It does not establish that every retrieved item is relevant, authoritative, or sufficient, and it does not make the final judgment correct. In a Two Prune simulation, the discovery record and work trail can contribute evidence about Evidence Navigation and Context Framing. They support a task-bounded assessment signal, not a claim about all future work or an automated employment decision.

Sources

UK National Screening Committee evidence review process

• U.S. Government Accountability Office, Government Auditing Standards 2024 Revision

Sources: [1] [2]