Return to the original observation
A verification plan begins with a precise statement of the claim. “A possible enclosure appears in the terrain model” can be tested more clearly than “the AI found a settlement”. Record which parts of the shape are actually visible, which are inferred and what would contradict the interpretation. This helps reviewers avoid unconsciously searching only for confirmation of a compelling first impression.
Inspect the original measurements rather than only a presentation image. Compare alternative terrain visualisations, different dates of photography and any available survey metadata. An apparent bank may align with a data seam; a dark patch may be a shadow; a rectangular outline may correspond to a demolished modern building. Rejecting such candidates is a successful outcome because it protects time and prevents misleading records.
Read the existing record
Desk-based work should bring together historic mapping, previous investigations, geological information and relevant heritage inventories. In England, local Historic Environment Records can provide context that a national overview or a general web search misses. Historic England’s research guidance helps explain the strengths and limits of different investigation methods. Elsewhere, the appropriate local or national heritage body may hold records under different access arrangements.
Existing records are not complete ground truth. A blank area on a map may indicate limited survey rather than an empty past. Conversely, a feature absent from a public map may already be known to specialists and withheld for protection. Researchers should distinguish “not in the dataset consulted” from “previously unknown”. The difference matters when deciding whether a discovery claim is justified.
Local knowledge can also expose errors and enrich the question. A land manager may remember drainage works; a community historian may identify a former route; residents may explain a place name or a landscape practice. These contributions should be documented with permission and attribution where appropriate, not treated as anonymous raw material for a technical report.
Choose a proportionate field question
A visit requires lawful access, a safety plan and any necessary permissions. Publicly visible imagery does not confer a right to cross land, fly a drone or disturb deposits. Protected sites, environmental designations and local rules can impose additional requirements. Consultation with the responsible heritage authority should happen before work that could affect archaeological remains, not after a promising result attracts attention.
The first field question might be modest: does the mapped earthwork exist, and does it appear modern? A measured walkover, recorded observations or non-invasive geophysics may answer it. More complex questions about construction, use or date may require other approaches. The selected method should be capable of addressing the question under local ground conditions; collecting more data without a clear purpose is not automatically better research.
Excavation is not a universal verification button. It changes or removes deposits and creates obligations for recording, conservation, reporting and archiving. A limited intervention may be appropriate within a permitted research design, but leaving a site intact can also be the responsible decision. Machine-learning enthusiasm does not reduce those professional duties.
Keep observations and conclusions apart
Field records should state what was observed, how it was measured and how confidently it can be related to the candidate. A visible bank is an observation. Calling it a prehistoric boundary is an interpretation that requires additional support. Finding pottery on a surface may suggest activity, but displaced material does not necessarily date a particular earthwork. Relationships between deposits and features often matter more than an isolated object.
Good documentation includes coordinate reference information in the controlled project record, instrument settings where relevant, photographs, survey limitations and unresolved alternatives. Public versions can generalise sensitive location details without weakening the explanation of method. If field conditions prevent a meaningful assessment, the outcome should remain inconclusive rather than being forced into a simple confirmed-or-rejected label.
Review independently
Seek evidence that is not simply a re-rendering of the same original dataset.
Test proportionately
Match the field method to a specific question, permissions and site conditions.
Report the limits
Keep unresolved candidates visible in the research record without presenting them as discoveries.
Close the loop without rewriting history
Verification results should return to the model evaluation. Confirmed false positives reveal recurring confusions; missed features reveal gaps in detection. However, updating the training set must not erase the original test result. Keep versions so that a future reader can see what the model knew at each stage and whether reported improvements were measured independently.
The Chartered Institute for Archaeologists publishes standards and guidance relevant to professional investigation. The Archaeology Data Service supports the preservation and reuse of digital archaeological records. Together, these perspectives remind us that a discovery is not complete when a persuasive image is posted. Its durable value depends on a record that other people can understand, challenge and responsibly revisit.
Continue to heritage safeguards