The Gold Standard of Archaeology | Use-Wear Reference Collections, Analyst Blindness and Functional Attribution

1. Function is inferred from a trace-producing history

Back to contents

A polished edge does not carry a label saying “cut reeds.” It records contact between a tool surface and other materials under particular forces, motions, durations and environmental conditions. Functional interpretation reconstructs part of that history.

The same artefact may have been manufactured, hafted, used for several tasks, resharpened, deposited, moved in sediment, excavated, washed and stored. Each stage can add, remove or alter traces.

Start with a process question: which observed features are consistent with use, which alternative processes could produce them, and what comparison makes the distinction credible? A typological name such as scraper is not itself evidence that scraping occurred.

2. Build the reference collection around variables, not labels

Back to contents

An experimental reference collection should document raw material, blank production, edge angle, surface preparation, hafting, worked material, material condition, motion, force, duration, operator, environment, cleaning and observation method. A photograph without that history is an illustration, not a controlled reference.

Use replicated experiments where feasible. Wear develops along a continuum and can vary between operators and pieces of the same rock. A single ideal specimen encourages analysts to match a memorable picture rather than a distribution of traces.

Include unused controls and taphonomic controls. They show which features arise from manufacture, handling, sediment, bag contact or cleaning and make the collection useful for exclusion as well as resemblance.

3. Comparability begins with raw material and surface history

Back to contents

Polish and fracture expression depend on mineralogy, grain size, texture and surface topography. A reference made on fine-grained flint may not transfer directly to quartz, obsidian, shell or a coarse ground stone.

Archaeological weathering, patination, heat and conservation treatment can mask or transform wear. Record surface preservation before selecting the reference subset. If the experimental and archaeological surfaces differ materially, lower the resolution of the functional claim.

A good reference search therefore begins with compatibility filters rather than the desired activity. It asks which experiments share the relevant material and history before asking which look most similar.

4. Observation protocols control analyst degrees of freedom

Back to contents

Specify cleaning, lighting, magnification sequence, microscope type, scanned area, image acquisition and descriptive terminology before interpretation. Searching only the “interesting” edge can inflate positive findings.

Use a systematic traverse and record absent as well as present features: edge removals, rounding, striations, polish texture, distribution and orientation. Separate low-power observations from high-power or quantitative surface measurements.

Archive raw images and instrument settings where permissions allow. A final functional label cannot be re-evaluated if the path from surface to claim has disappeared.

5. Mask context when it can steer identification

Back to contents

An analyst who knows that a tool came from a hide-working area may unconsciously search for traces expected from hide processing. Masking contextual hypotheses during the first observation pass reduces that expectation channel.

Blind testing has two roles. Calibration tests ask analysts to classify experimental pieces whose use is hidden. Case masking asks analysts to describe archaeological traces before learning typology, residues or spatial interpretation.

Complete blindness is not always possible or desirable because raw material and preservation are needed to choose valid comparisons. Record exactly what was masked, what was disclosed and when interpretive information entered the workflow.

6. Worked example: six edges and a tempting story

Back to contents

Six stone flakes from one deposit have edge rounding and patches of polish. The excavator expects cereal processing because charred grain occurs nearby. The analyst first receives coded pieces with raw material and preservation information but no activity hypothesis.

Using a prespecified traverse, the analyst finds two pieces with oriented striations and polish distribution consistent with longitudinal contact on medium-soft plant material, two with weak non-diagnostic rounding, one dominated by sediment abrasion and one too altered to assess. A second analyst independently reviews images and records.

The report does not call all six sickle inserts. It attributes two pieces to a bounded plant-working class, treats two as indeterminate, excludes one from functional inference and reports one probable taphonomic overprint. Grain remains support a site activity, not the function of every edge.

7. Match claim resolution to test performance

Back to contents

An analyst may distinguish motion more reliably than exact worked material, or broad material classes more reliably than species. Report the level at which blind tests and reference coverage demonstrate competence.

Use confusion matrices rather than a single accuracy percentage. If bone and antler are frequently confused, a combined hard-animal-material category may be defensible even when exact identification is not.

Do not turn an uncertain trace into a precise task by combining several weak clues as if they were independent. Residue, wear and context can corroborate one another, but shared contamination or expectation can also make them fail together.

8. Changed-context transfer check

Back to contents

Move the worked example from flaked stone to shell tools used on plant, hide and mineral materials. The logic transfers: documented reference histories, unused controls, masked assessment and bounded functional classes remain necessary.

The trace formation and observation protocol do not transfer unchanged. Shell microstructure, fracture behaviour, cleaning tolerance and preservation require a shell-specific reference collection and terminology.

A flint analyst’s confidence cannot be imported as evidence of shell competence. Recalibrate with blind shell experiments and report which material-motion combinations the new method can actually discriminate.

9. Limits and common misconceptions

Back to contents

Reference collections demonstrate similarity under tested conditions; they do not recreate every ancient task. Absence of visible wear can reflect short use, resharpening, poor preservation or inadequate imaging rather than non-use.

Blind testing does not remove all subjectivity, and algorithmic classification does not remove it either. Training labels, image selection, preprocessing and reference coverage embed human decisions.

Do not infer intended function from actual use, or actual use from typological design. Do not call a surface “diagnostic” without comparative performance. Do not wash or consolidate an artefact before considering how treatment will affect traces and residues.

10. Audit the chain from experiment to assemblage claim

Back to contents

For each archaeological piece, retain catalogue identity, condition, cleaning history, scanned zones, observations, reference subset, candidate interpretations, exclusion reasons, analyst confidence and contextual information disclosed.

At assemblage level, report how many pieces were examined, assessable, positive, indeterminate and excluded. Sampling only formal tools cannot establish activity prevalence across all debitage.

Reanalyse a subset independently, include known controls where possible and publish disagreement. The strongest audit makes it possible to see whether the conclusion depends on one analyst, one reference image or one unusually well-preserved edge.

11. Frequently asked questions

Back to contents

How large must a reference collection be?

No single number is sufficient. Coverage should match the raw materials, motions, worked materials, durations and preservation states relevant to the claim, with replication and controls where variation is expected.

Should analysts always be blind to archaeological context?

A masked first pass is valuable when context could steer recognition. Necessary material and preservation information can still be supplied, and context can enter later as a separate evidential layer.

Can machine learning replace experimental archaeology?

No. Models still require labelled, relevant reference data and external validation. They can quantify surfaces or classification, but the causal link to tool use remains experimental and archaeological.

12. The gold standard is calibrated functional attribution

Back to contents

Use-wear becomes persuasive when the reader can follow the complete comparison: archaeological surface, alternative formation processes, relevant experiments, masked observation and tested level of discrimination.

The practical test is simple to state: could another trained analyst, using the archived protocol and reference set, understand why this claim was made and where it should stop?

If yes, function is an accountable inference. If the result rests on resemblance to a favourite photograph or knowledge of the expected story, the label is more specific than the evidence.