The Gold Standard of Archaeology | Replicate Dates, Discordance Tests and Distinguishing Context Risk from Laboratory Error

1. Define what the replicate is meant to repeat

Back to contents

The word replicate hides several designs. A laboratory can measure the same prepared carbon twice, split pretreatment fractions, date two tissues from one organism, date separate organisms from one event or send matched material to independent laboratories. Each design repeats some stages and leaves others shared.

Draw the chain from excavation to reported age: context, selection, cleaning, pretreatment, graphitisation, measurement, background correction and reporting. Mark which links are common between determinations. Two measurements sharing one pretreatment do not test pretreatment contamination. Two seeds from one deposit test contextual coherence more than instrument repeatability.

The research question should choose the replicate. If the concern is laboratory reproducibility, split and blind suitable material. If the concern is whether a deposit represents one event, select independent short-lived samples with secure associations. More dates without a defined target can multiply numbers while leaving the actual risk untouched.

Practice check: Diagram two proposed determinations and circle every shared stage that the pair therefore cannot independently test. Do the check before settling the conclusion, because a neat answer produced after the decision is not evidence that the rule guided it. State what would count as a different answer and who is responsible for acting on it.

Changed-context drill: Compare replicate measurements of one bone extract with dates on three separate seeds from the same floor. Keep the principle, then change only the parts that genuinely depend on the setting. If the method cannot say what changes and what stays fixed, it is a slogan rather than a transferable standard.

Evidence record: Keep sample identities, provenience, biological relationship, pretreatment splits, laboratory codes and the stated purpose of replication. Preserve enough detail for a careful reader to reconstruct the reasoning without relying on the memory or authority of the person who made the decision. Name uncertainty directly; a visible limit is more useful than silent precision.

2. Keep a conventional age distinct from a calendar date

Back to contents

A laboratory reports a radiocarbon age with uncertainty and a code tied to the measured material. Calibration translates that age through a calibration curve into a calendar probability distribution. Replicate consistency is often assessed in radiocarbon-age space before calibration because the measurement model is clearer there.

Do not compare only the overlap of calibrated ranges. Calibration curves are irregular; two compatible conventional ages can produce complicated calendar distributions, while visually overlapping calibrated intervals can mask measurement inconsistency. Retain the unrounded age and standard uncertainty for each determination, and state any additional uncertainty used.

Chronological models may combine determinations after appropriate checks. Combination is not clerical averaging. It asserts that the measurements estimate the same underlying radiocarbon concentration or event under the model. That assertion needs archaeological and material justification as well as a statistical calculation.

Practice check: Place the conventional ages, uncertainties and calibrated distributions side by side and explain why visual overlap is not the consistency test. Do the check before settling the conclusion, because a neat answer produced after the decision is not evidence that the rule guided it. State what would count as a different answer and who is responsible for acting on it.

Changed-context drill: Move the samples onto a calibration plateau and observe how similar ages can yield broad calendar ranges. Keep the principle, then change only the parts that genuinely depend on the setting. If the method cannot say what changes and what stays fixed, it is a slogan rather than a transferable standard.

Evidence record: Archive lab codes, conventional ages, uncertainties, calibration curve and version, software settings and unrounded outputs. Preserve enough detail for a careful reader to reconstruct the reasoning without relying on the memory or authority of the person who made the decision. Name uncertainty directly; a visible limit is more useful than silent precision.

3. Use weighted combination only after testing consistency

Back to contents

When independent measurements estimate the same underlying age and their uncertainties are credible, an inverse-variance weighted mean gives more weight to the more precise determination. But precision is not permission. Combining incompatible results can manufacture a narrow mean that describes none of the material.

A chi-square consistency test such as the Ward and Wilson approach asks whether scatter is greater than expected from the quoted uncertainties under a common-age model. Report the test statistic, degrees of freedom and threshold used. A non-significant result means the data did not show excess scatter at that sensitivity; it does not prove identity or correct context.

Small sets have limited power, and quoted errors may omit shared systematic components. Independence matters: duplicate counts from the same target or corrections based on the same standard are not equivalent to fully independent dates. Treat the test as one layer in a case, not a switch that turns several samples into one event.

Practice check: Calculate or review the consistency test before any weighted mean, then state the common-age assumption in archaeological words. Do the check before settling the conclusion, because a neat answer produced after the decision is not evidence that the rule guided it. State what would count as a different answer and who is responsible for acting on it.

Changed-context drill: Repeat with one uncertainty halved and ask whether greater weight reflects genuine information or optimistic error reporting. Keep the principle, then change only the parts that genuinely depend on the setting. If the method cannot say what changes and what stays fixed, it is a slogan rather than a transferable standard.

Evidence record: Retain the formula or software, covariance assumptions, statistic, degrees of freedom, decision threshold and combined result if justified. Preserve enough detail for a careful reader to reconstruct the reasoning without relying on the memory or authority of the person who made the decision. Name uncertainty directly; a visible limit is more useful than silent precision.

4. Diagnose discordance by stage, not by convenience

Back to contents

A discordant result can arise from mislabelling, instrument or background problems, incomplete pretreatment, conservation material, reservoir effects, old wood, intrusive roots, residual deposits, mixed contexts or a genuine sequence of events. The latest or earliest result is not automatically the error.

Start with the laboratory record and sample chemistry, then return to excavation documentation and formation processes. Ask what would produce the observed direction and magnitude of difference. Modern contamination tends to shift an old sample younger; inbuilt age shifts charcoal earlier than the target event; a marine reservoir can make material appear older without laboratory failure.

Investigation should be symmetrical. Researchers often discard the result that conflicts with expectation while scrutinising the convenient date less closely. Use blind identifiers where practical, predefine exclusion grounds and document all outcomes. An unexpected date is evidence about the sample-context relationship until a specific failure is demonstrated.

Practice check: List at least one laboratory, material and context explanation for the direction of the observed discordance. Do the check before settling the conclusion, because a neat answer produced after the decision is not evidence that the rule guided it. State what would count as a different answer and who is responsible for acting on it.

Changed-context drill: Switch the sample from a short-lived seed to unidentified charcoal and reassess the most plausible risk. Keep the principle, then change only the parts that genuinely depend on the setting. If the method cannot say what changes and what stays fixed, it is a slogan rather than a transferable standard.

Evidence record: Keep pretreatment data, quality flags, conservation history, photographs, stratigraphic sheets, exclusion rules and communication with the laboratory. Preserve enough detail for a careful reader to reconstruct the reasoning without relying on the memory or authority of the person who made the decision. Name uncertainty directly; a visible limit is more useful than silent precision.

5. Distinguish same-sample, same-organism and same-context replication

Back to contents

A duplicate measurement of one homogenised sample estimates analytical repeatability but shares material history and preparation. Different fractions from one bone can reveal pretreatment or contaminant sensitivity. Two seeds from the same cereal ear are biologically close; two seeds from a sealed hearth may represent one use episode but still be separate organisms.

These levels create a nested design. Variation within a prepared target points toward counting or target preparation. Variation between fractions points toward pretreatment or preservation. Variation among organisms in one context may reveal residuality, intrusion or duration. Variation between laboratories can expose process differences but also includes sampling heterogeneity if the submitted pieces were not equivalent.

Name the level in the article and dataset. “Replicate dates from the layer” is too vague. A reader should know whether replication tested the instrument, the specimen or the archaeological event. Without that information, agreement can be overclaimed and discordance misassigned.

Practice check: Classify each determination pair by shared sample, organism, context, pretreatment and laboratory. Do the check before settling the conclusion, because a neat answer produced after the decision is not evidence that the rule guided it. State what would count as a different answer and who is responsible for acting on it.

Changed-context drill: Compare two split charcoal fragments with two seeds recovered metres apart in the same stratigraphic unit. Keep the principle, then change only the parts that genuinely depend on the setting. If the method cannot say what changes and what stays fixed, it is a slogan rather than a transferable standard.

Evidence record: Use a nested identifier system linking context, object, subsample, fraction, preparation batch, target and measurement. Preserve enough detail for a careful reader to reconstruct the reasoning without relying on the memory or authority of the person who made the decision. Name uncertainty directly; a visible limit is more useful than silent precision.

6. Worked example: three barley grains and a duplicate target

Back to contents

A burnt floor produces three barley grains. Grain A is measured twice from the same graphite target and the results agree closely. Grain B is statistically consistent with A. Grain C is several centuries later and fails the group consistency test. Averaging all four numbers would narrow the answer while hiding the late grain.

Laboratory review finds normal backgrounds, yields and quality indicators. Excavation photographs show Grain C came from a narrow root channel that cut the floor but was missed during bagging. The discordance is best explained as contextual intrusion, not laboratory error. A and B may support the burning event, while C dates later activity in the channel.

Now change one fact: suppose Grain C came from a secure lump of fused floor material and its pretreatment yield was abnormal. The investigation shifts toward material chemistry and repeat pretreatment. The example shows why the statistical flag cannot name the cause. It tells the team where a common-age story is inadequate.

Practice check: Write the inference after each stage: before the test, after discordance, after lab review and after contextual review. Do the check before settling the conclusion, because a neat answer produced after the decision is not evidence that the rule guided it. State what would count as a different answer and who is responsible for acting on it.

Changed-context drill: Replace Grain C with oak heartwood that is earlier rather than later and diagnose inbuilt age. Keep the principle, then change only the parts that genuinely depend on the setting. If the method cannot say what changes and what stays fixed, it is a slogan rather than a transferable standard.

Evidence record: Preserve specimen maps, bagging notes, target relationships, test output, laboratory review and the reason each date is retained or separated. Preserve enough detail for a careful reader to reconstruct the reasoning without relying on the memory or authority of the person who made the decision. Name uncertainty directly; a visible limit is more useful than silent precision.

7. Use inter-laboratory comparison carefully

Back to contents

Sending material to two laboratories can test reproducibility only if the material is sufficiently homogeneous and allocation is controlled. One lab receiving clean inner collagen and another receiving surface-contaminated material creates a sample difference, not a pure laboratory comparison. Blind coding and split homogenised material reduce expectation and allocation effects.

Laboratories may use different pretreatments, backgrounds and uncertainty conventions for legitimate reasons. A discrepancy should trigger technical discussion, not a ranking based on which answer matches the chronology. Request process information relevant to the material while respecting laboratory quality systems and limits on proprietary details.

Shared calibration curves and reference materials also mean that inter-laboratory agreement is not total independence. Use comparison exercises and known-age materials where available, but keep the archaeological specimen problem separate. A laboratory can perform well in proficiency tests while a particular sample remains unsuitable.

Practice check: Design a blind split that holds material constant and states which laboratory stages are intentionally allowed to differ. Do the check before settling the conclusion, because a neat answer produced after the decision is not evidence that the rule guided it. State what would count as a different answer and who is responsible for acting on it.

Changed-context drill: Move from homogeneous charcoal powder to a heterogeneous bone with consolidant on one surface. Keep the principle, then change only the parts that genuinely depend on the setting. If the method cannot say what changes and what stays fixed, it is a slogan rather than a transferable standard.

Evidence record: Keep allocation, blinding, mass, fraction, pretreatment protocol, quality data, uncertainty convention and proficiency information. Preserve enough detail for a careful reader to reconstruct the reasoning without relying on the memory or authority of the person who made the decision. Name uncertainty directly; a visible limit is more useful than silent precision.

8. Changed-context transfer: a stratigraphic sequence

Back to contents

A sequence model contains dates from several layers. Two determinations in the same layer disagree, while each fits a different neighbouring phase. It is tempting to assign them according to the model that looks most orderly. That reverses the evidence chain: the archaeological relationship should inform the model, and the model should help reveal tension, not erase it.

Apply the replicate logic at the event level. Ask whether the layer represents a brief deposit, accumulated fill or redeposited material. Inspect refits, interfaces, residual artefacts and sample taphonomy. Model the determinations as separate events or outliers if justified, and compare results under explicit alternatives.

The transferable principle is to locate inconsistency at the correct level. In a sequence, context risk includes misordered relationships and long formation durations, not just one intrusive specimen. A coherent posterior can still be wrong if strong priors force uncertain samples into place.

Practice check: Fit or conceptually compare a single-event, accumulated-layer and intrusive-sample interpretation. Do the check before settling the conclusion, because a neat answer produced after the decision is not evidence that the rule guided it. State what would count as a different answer and who is responsible for acting on it.

Changed-context drill: Introduce a calibration plateau and ask which uncertainty comes from the curve rather than the deposit. Keep the principle, then change only the parts that genuinely depend on the setting. If the method cannot say what changes and what stays fixed, it is a slogan rather than a transferable standard.

Evidence record: Retain stratigraphic matrices, model code, priors, outlier settings, agreement indices and results under plausible contextual alternatives. Preserve enough detail for a careful reader to reconstruct the reasoning without relying on the memory or authority of the person who made the decision. Name uncertainty directly; a visible limit is more useful than silent precision.

9. Know the limits and common misconceptions

Back to contents

Agreement does not prove correct association. Two residual charcoal fragments can agree with each other and still predate the deposit. Discordance does not prove laboratory error. A passed chi-square test does not show that the samples are contemporaneous, especially with few imprecise determinations. A failed test does not identify which result is wrong.

More replicates improve diagnosis only when they add independent information at the relevant level. Recounting the same target can refine analytical uncertainty but cannot reveal that the object was residual. Dating several pieces of the same long-lived timber does not create multiple independent event dates.

Avoid “majority vote” dating. Three agreeing determinations are not automatically right if they share contamination or context risk. The conclusion should state what was replicated, which failure modes remain shared and why the retained dates answer the archaeological question.

Practice check: Take one reassuring pattern of agreement and name a shared error that could produce it. Do the check before settling the conclusion, because a neat answer produced after the decision is not evidence that the rule guided it. State what would count as a different answer and who is responsible for acting on it.

Changed-context drill: Consider a museum object with undocumented conservation and no excavation context. Keep the principle, then change only the parts that genuinely depend on the setting. If the method cannot say what changes and what stays fixed, it is a slogan rather than a transferable standard.

Evidence record: Document remaining shared risks, power limits, rejected alternatives and the claim boundary in the final chronological statement. Preserve enough detail for a careful reader to reconstruct the reasoning without relying on the memory or authority of the person who made the decision. Name uncertainty directly; a visible limit is more useful than silent precision.

10. Audit the full evidence chain

Back to contents

The audit begins in the field: provenience, photographs, relationships, material identification, selection rationale and sampling personnel. It continues through storage, conservation history, shipment, pretreatment, measurement, calibration, statistical comparison and chronological modelling. Codes must link each stage without ambiguity.

Create a replicate map showing shared and independent components. Add a discordance log that records when inconsistency was detected, who reviewed it, what evidence was checked and what action followed. Retain excluded determinations in the archive and publication table with reasons; disappearance is not quality control.

Finally, rerun the interpretation under plausible alternatives: no combination, alternative grouping, an outlier model, different contextual assignments and appropriate curve choices. Report which conclusions survive. The aim is not to find the model with the narrowest dates, but to show how archaeological claims respond to uncertainty in measurement and association.

Practice check: Choose one published date group and reconstruct every shared and independent stage from trench to model. Do the check before settling the conclusion, because a neat answer produced after the decision is not evidence that the rule guided it. State what would count as a different answer and who is responsible for acting on it.

Changed-context drill: Repeat after discovering that two supposedly separate samples came from the same object. Keep the principle, then change only the parts that genuinely depend on the setting. If the method cannot say what changes and what stays fixed, it is a slogan rather than a transferable standard.

Evidence record: Archive the replicate map, all raw results, calibration files, model versions, exclusions, review notes and robustness comparisons. Preserve enough detail for a careful reader to reconstruct the reasoning without relying on the memory or authority of the person who made the decision. Name uncertainty directly; a visible limit is more useful than silent precision.

A project-level replicate audit benefits from a sample genealogy. Begin with the deposit and draw every physical division: object, fragment, powder, chemical fraction, graphite target and repeated count. Next to each branch, record which uncertainties remain common and which new uncertainties are introduced. This prevents a duplicated count from being described as a second archaeological sample and prevents two different fragments from being treated as analytically identical. The diagram can also reveal that “independent laboratories” received material with different conservation exposure or tissue composition.

Use a discordance conference with a fixed order. The field archaeologist presents context and formation evidence without seeing which result is inconvenient. The laboratory presents yields, chemistry, backgrounds and process flags without being told the preferred chronology. The chronological modeller then shows the consistency test and model tension. Only after those views are recorded should the group compare explanations. This sequence cannot remove bias, but it makes retrospective selection harder and preserves alternative accounts that later evidence may revive.

For every excluded or separated determination, write a causal reason rather than a label. “Outlier” describes model tension; it does not explain it. “Intrusive root-derived seed identified from the microstratigraphic channel” is an archaeological explanation. “Low collagen yield with contaminant-sensitive fraction and failed quality threshold” is a material or laboratory explanation. When no specific cause is supported, retain the result as unresolved and show how it changes the model. Honest unresolved discordance is more informative than a clean chronology produced by disappearance.

A robustness table should compare at least four analyses: all determinations separate, only pre-specified eligible material, justified combinations after consistency tests, and a model allowing an explicit outlier or mixture process. Report the event range, ordering probability and any substantive historical claim for each. If the story depends on one grouping decision, that dependency belongs in the conclusion. If the claim survives all plausible analyses, replication has strengthened the inference in a way the reader can see.

Destructive sampling adds another responsibility. Before requesting more material, identify which uncertainty a new determination can reduce. A recount of an existing target may clarify instrument repeatability; a new pretreatment fraction may probe contamination; a separate short-lived specimen may probe context. If none addresses the suspected failure, more sampling consumes heritage without adding the required information. The decision record should include expected information, required material, permission, remaining archive and the stop condition for further testing.

Close the audit with a plain-language sentence for each date group: what physical event it most likely records, what statistical evidence supports grouping, what context evidence links it to the research event and what alternative remains plausible. This sentence forces laboratory precision and archaeological meaning into the same frame. It is also a useful check against reporting a narrow calibrated range for a sample whose relationship to the event remains broad.

11. Frequently asked questions

Back to contents

Should consistent replicate ages always be combined? Only when they estimate the same underlying quantity and independence assumptions are defensible. Should a failed test trigger remeasurement? It triggers investigation; remeasurement may help only if it targets a plausible stage. Can one lab result be preferred because its error is smaller? Not without checking comparability and quality.

Do calibrated ranges need to overlap? Visual overlap is not the formal test and can mislead. Can dates from different materials be replicates? They can test one event if their associations are justified, but material-specific offsets must be considered. Is an outlier model a substitute for excavation review? No. It quantifies tension under a model; it does not explain formation history.

How many replicates are enough? The number depends on the risk, desired precision and level being tested. What should be published? All determinations, lab codes, uncertainties, material and context links, test results, exclusions and model assumptions. Transparency lets agreement and discordance remain archaeological evidence rather than private troubleshooting.

Practice check: Answer one project-specific question without using the words “reliable” or “outlier” until the evidence is named. Do the check before settling the conclusion, because a neat answer produced after the decision is not evidence that the rule guided it. State what would count as a different answer and who is responsible for acting on it.

Changed-context drill: Explain the same answer to a curator deciding whether destructive resampling is justified. Keep the principle, then change only the parts that genuinely depend on the setting. If the method cannot say what changes and what stays fixed, it is a slogan rather than a transferable standard.

Evidence record: Maintain a dated FAQ with project decisions, authority for methods and links to the underlying sample and model records. Preserve enough detail for a careful reader to reconstruct the reasoning without relying on the memory or authority of the person who made the decision. Name uncertainty directly; a visible limit is more useful than silent precision.

12. The gold standard explains both agreement and disagreement

Back to contents

Replicate dating is an experimental design embedded in an archaeological argument. The team defines the level of replication, selects material that can test the relevant risk and preserves the identity of every specimen and processing stage. Consistency is assessed before combination, and the uncertainty model is stated.

When results disagree, the investigation crosses laboratory and field records. It does not blame the least convenient number. Explanations are judged by direction, material behaviour, formation process and independent evidence. Some discordance will remain unresolved; that uncertainty should widen or divide the chronological claim.

The final test is whether a reader can tell what agreement rules out and what remains shared. A replicate design that tests only measurement should not be advertised as proof of context. A contextual replicate that reveals mixture should not be averaged away. The gold standard treats disagreement as information about the chain that produced the date.

Practice check: Summarise the project in three sentences: what was replicated, what the test found and what risk remains. Do the check before settling the conclusion, because a neat answer produced after the decision is not evidence that the rule guided it. State what would count as a different answer and who is responsible for acting on it.

Changed-context drill: Imagine the result challenges the preferred historical narrative and ask whether the same inclusion rules still apply. Keep the principle, then change only the parts that genuinely depend on the setting. If the method cannot say what changes and what stays fixed, it is a slogan rather than a transferable standard.

Evidence record: Publish the sample hierarchy, consistency tests, investigation, final grouping, uncertainty and a plain-language limit on the conclusion. Preserve enough detail for a careful reader to reconstruct the reasoning without relying on the memory or authority of the person who made the decision. Name uncertainty directly; a visible limit is more useful than silent precision.