Gains:
- Ability to maintain data integrity while digitizing and standardizing recording units such as context/locus, stratigraphy and Harris Matrix with artificial intelligence
- Ability to protect raw data and keep each transformation traceable against the risk of artificial intelligence completing missing data and making silent changes
- Understand why controlled dictionary, metadata, and open/FAIR principles are essential for the future usability of data.
The scientific value of an excavation lies not in the gold finds that emerge from it, but in the quality of its records. A find that is not well documented is almost irrelevant to science; Because it is an object that is unknown where, in which layer, and with what it is found, it is just a beautiful object. In this unit we will look at excavation logging (the systematic transcription and database of every context, find, and observation) and how AI can speed up data entry, organization, and standardization, but why the responsibility for data integrity remains with humans.
Basic principle: AI helps organize scattered records, standardize them and find inconsistencies; But which data is accurate, whether a record reflects the truth, and the integrity of the data are human responsibility. AI corrects the shape of the data, it does not guarantee its accuracy.
Basic units of the archaeological record
Context/locus. Each excavation divides the site into units of context (context/locus—an individual deposit, layer, pit, or wall; the smallest unit of meaning of the excavation). Each context receives a unique number and all finds are associated with that number.
Stratigraphy and Harris Matrix. Stratigraphy (the before-after relationship of layers relative to each other) is the basis of relative dating. The Harris Matrix (a diagram showing the ordering of contexts relative to each other) visualizes these relationships. AI helps detect inconsistent stratigraphic relationships (e.g. a cyclic “A is both above and below B” error).
Find inventory. Every find; The context is recorded with number, type, size, material, condition and photograph.
Data standards. Common standards are used to make records shareable and comparable. CIDOC-CRM (an international conceptual framework/ontology developed to describe cultural heritage data) enables data from different institutions to talk to each other. A controlled dictionary of terms (an approved list that ensures that everyone uses the same term for the same concept) is used.
Tip: Use AI to “organize and audit” data, not “interpret” it. “What context numbers are contradictory in these records?” It's a good question; "What period does this context belong to?" It is an expert decision. Where AI is strongest is in consistency checking.
Roles of AI in registration and database
- Digitization. Hand-written excavation notebooks, inventory cards and old drawings are imported into the database with AI-powered text recognition (this links to HTR in unit 6).
- Standardization. Different spellings ("byzantium", "Byzantium", "byz.") are mapped into the controlled dictionary; dates and measurements are brought into uniform form.
- Consistency check. Errors such as missing context number, contradictory stratigraphy, use of the same number in two different finds are marked.
- Query and summary. Queries such as “All glass finds in the following context” and summary tables are quickly generated.
Caution: When standardizing, the AI may silently modify data while trying to "fix" it; for example, he may round a measurement to a "reasonable" value or fill in an uncertain reading with his own estimate. Every automatic change must be traceable and the original record must be preserved. Raw data is never overwritten.
three mini cases
Case 1 — Digitization saved the archive. A museum was storing 40 years of handwritten inventory cards (about 22,000 cards) at risk of being lost. AI-powered text recognition and standardization imported the cards into a searchable database; Each card was checked on a sample basis by a human examiner, and unreadable areas were marked “unclear.”
Case 2 — Inconsistency check found errors. An excavation team had the AI audit the database at the end of the season; YZ marked that three finds had the same context number but conflicting layer information. The review revealed a data entry error; if not corrected it would have distorted the stratigraphic interpretation.
Case 3 — Silent change caught. While standardizing measurements, an assistant noticed that the AI was interpreting some “0” values as “zero” rather than “missing”; this distorted the broken piece lengths. The process was rebuilt to preserve the original data but keep the changes in a separate column.
Four copyable templates
1) Standardization (controlled dictionary):
Your role: data wrangling assistant. Map the values of [field:material/period/type] in the following records to the following controlled dictionary: [dictionarylist]. CHANGE values that do not have equivalents in the dictionary; Mark it as "no match". Report each change as an original-new pair; deleting raw data.
2) Consistency check:
Find inconsistencies in the following excavation records: repetitive context numbers, missing mandatory fields, conflicting stratigraphic relationships, awkward dates/measurements. List each problem with registration number and leave it to ME to fix it; don't change it yourself.
3) Harris Matrix logic control:
Below are stratigraphic relationships between contexts (A above/below B). Is there a logical contradiction (loop, two-way relationship)? Show which contexts, if any. Do not comment; just check for logical consistency.
4) Query and summary table:
The following extract from the database dump below: [e.g. find type distribution per context]. Give the result as a table, specifying which records each row is derived from. DO NOT ADD any value that does not exist in the data; If it is empty, write "no data".
Weak prompt / Strong prompt
Weak prompt:
Correct this excavation data and fill in the missing items.
“Fill in the gaps” allows the AI to fit data; The result is an unreliable database where fact and fiction mix.
Powerful prompt:
Mark ONLY formal inconsistencies (spelling, date format, duplicate ID) in the excavation data below. COMPLETE missing values; Leave it as "incomplete". List each proposed change along with the original; I will give the approval. Changing raw data.
Difference: the former corrupts data silently; The second one controls and leaves the decision to the human.
Good data model: fields, relationships and metadata
An archaeological database is not an arbitrary table, but a thoughtful data model (a blueprint of which tables, which fields, and what relationships the data will be organized into). The basic principle is that everything depends on context: finds to context, contexts to each other (stratigraphy) and to the field. Each record carries a unique identity (ID) and relationships are established through these identities; A find refers to a single context, a context may contain many finds.
Just as important as the recording is the metadata (data about the data itself: who recorded it, when, by what method, what version). A record that is unknown by whom, when and with what confidence was entered cannot be questioned in the future. AI is helpful in checking the consistent filling of metadata fields and flagging missing metadata.
Another critical principle is an open and sustainable format. Rather than locking data into proprietary, closed software, keeping it close to open standards (text-based tables, documented schemas) ensures that data can be read decades later. Archaeological data are produced not for a single publication but for future generations; That's why FAIR principles (Finding, Accessible, Interoperable and Reusable of data) have become increasingly standard. AI is useful in labeling your data in accordance with these principles and finding deficiencies; But it is up to you to decide which data will remain open and which will remain sensitive (confidential).
Tip: When you're setting up your database, ask yourself, "Can someone who doesn't know this open it and understand it in 10 years?" ask. Explain your abbreviations in a glossary, fill in the metadata fields, and back up the raw data in an open format.
Record/database task table
Quest
Role of AI
human decision
protection rule
digitization
text recognition
Vague reading confirmation
Raw scan is stored
standardization
Map to dictionary
Mismatched value decision
The original is preserved
Consistency
Contradiction marking
correction
Change is tracked
stratigraphy
Logic control
Comment
Matrix expert approval
Query/summary
Table production
Comment, choice
Fidelity to data
Common mistakes
- Completing missing data. AI makes up; The missing must remain "incomplete".
- Overwriting the raw data. The original is always preserved; changes are kept separate.
- Not noticing silent changes. Every automatic conversion should be reported and audited.
- Skipping the standard. Without a controlled dictionary and a common model, data cannot be shared.
- Having AI perform stratigraphic interpretation. AI finds logical contradiction; Comment is for the expert.
In summary
AI in excavation record and database; It provides great speed in digitization, standardization, consistency checking and query works. But data integrity is sacred: no missing parts are left intact, raw data is preserved, every change is tracked, and stratigraphic interpretation belongs to the expert. Good records are the most valuable legacy that excavation leaves to the future.
Application task
Prepare a small sample excavation data table (10-20 records); deliberately put a few inconsistencies in there (duplicate ID, misshapen date, conflicting layer). Have the AI find these with “consistency check” and “Harris Matrix logic check” patterns; then write a controlled dictionary and map the material field to the "Standardization" template and make sure to preserve the raw data.
checklist
- [ ] I stored the raw data separately, unmodified.
- [ ] I did not make the AI complete the missing parts; I left it as "incomplete".
- [ ] I have checked each automatic change with the original.
- [ ] I used controlled dictionary and data standard.
- [ ] I made the stratigraphic interpretation based on expert judgment.