The Gold Standard of Archaeology | Chronological Model Conflict, Outlier Priors and Boundary Sensitivity

1. Separate measurement conflict from archaeological model conflict

Back to contents

A radiocarbon result can disagree with a chronology because the laboratory measurement is unusual, the sampled material is older than the event, the context is residual or intrusive, the stratigraphic order is wrong, the calibration curve is complex or the phase model is inappropriate. These possibilities have different remedies. Begin with the dated material, laboratory result and contextual association before changing the statistical structure. Ask what event the sample actually dates: organism death, timber growth, deposition, burning or later movement. A model conflict is useful because it reveals that at least one link between measurement and archaeological event needs attention. Calling every conflict an outlier hides that diagnostic value.

2. Write the prior model as an archaeological argument

Back to contents

List every ordering constraint, phase membership, boundary assumption, hiatus, offset and sample association in ordinary language before encoding it. Identify which claims come from observed stratigraphy, typology, historical evidence or analyst judgement. Priors are not automatically subjective weaknesses; they are the explicit route by which existing knowledge enters the chronology. The danger is invisible or overconfident prior structure. A phase boundary may imply a distribution of events that the excavation evidence cannot support. Keep a model diagram and versioned code so another archaeologist can reconstruct the argument and see which posterior precision comes from measurements and which comes from constraints.

3. Diagnose disagreement before removing data

Back to contents

Inspect laboratory quality, duplicate measurements, material identity, pretreatment, reservoir or old-wood effects, context integrity, calibration features, agreement diagnostics and convergence. Plot the unmodelled calibrated distributions beside the modelled results. A date that looks surprising in the posterior may be entirely compatible before a strong stratigraphic constraint is applied. Conversely, a high global fit can conceal one problematic association if the rest of the model dominates. Record diagnostic questions before opening alternative models so the investigation is not steered toward the preferred story. Deleting a measurement because it broadens or shifts a boundary is not sensitivity analysis; it is post hoc selection unless supported by independent evidence.

4. Use outlier priors to represent plausible failure processes

Back to contents

An outlier model assigns a probability and form to ways a dated sample may not represent the intended event. Different mechanisms matter: a charcoal sample may carry an old-wood offset, a residual object may predate deposition and an intrusive sample may postdate the context. Choose an outlier model because its direction and scale match the archaeological process, not because it improves agreement. Report prior outlier probabilities and posterior classifications without treating them as proof of contamination. Compare results with and without the model. If many samples require large adjustments, reconsider the phase or context model rather than allowing the outlier component to absorb structural failure.

5. Test how boundary estimates depend on phase assumptions

Back to contents

Boundaries are derived quantities shaped by the dates, calibration curve and assumptions about how events occupy a phase. Uniform, trapezoidal or other process models can imply different start, transition and end estimates. Vary plausible phase membership, ordering, hiatus and event-distribution assumptions one at a time and in combinations that reflect real alternatives. Compare posterior medians, credible intervals and the probability of historically relevant ranges. Do not report only the model that gives the narrowest boundary. A robust conclusion remains substantively similar across defensible models; a fragile one should be presented as model-dependent, even when each individual posterior looks precise.

6. Worked example: an enclosure with residual charcoal

Back to contents

Eight radiocarbon dates constrain the construction and use of an enclosure. One charcoal date is centuries earlier than the articulated bone from the same ditch fill. The first model forces all samples into a single uniform phase and produces poor agreement, while a generic outlier model downweights the charcoal and estimates construction at 820–780 BCE. Excavation notes show that the charcoal came from redeposited hearth material, making an earlier-only offset plausible. A second model uses an old-material outlier prior and separates construction from later infill; the construction boundary widens to 850–775 BCE. Removing the charcoal entirely gives 815–770 BCE. The report retains all three results and explains why the second is archaeologically preferred.

7. Compare a small family of defensible models

Back to contents

A model comparison should answer named questions, not produce an unlimited garden of alternatives. Predefine a base model and variants for the uncertain assumptions: context membership, outlier mechanism, boundary form, reservoir correction or stratigraphic link. Keep the data and calibration curve fixed unless those are themselves under study. Use diagnostics and posterior summaries, but also compare the archaeological conclusions that matter. If every model places an event before a known horizon, that conclusion may be stable despite different exact boundaries. If ordering reverses under one plausible association change, uncertainty should be communicated at that level rather than hidden inside a single interval.

8. Changed-context transfer check: luminescence dates in a sediment sequence

Back to contents

The same reasoning applies when optically stimulated luminescence dates are combined with stratigraphy. The changed context alters the failure processes. Incomplete bleaching can make a depositional age too old; dose-rate heterogeneity and post-depositional mixing can broaden or bias estimates. An outlier prior designed for residual charcoal is not transferable without modification. The analyst should map what each sample dates, encode stratigraphic ordering, test boundary and accumulation assumptions and compare models that represent plausible bleaching or mixing processes. The gold standard transfers the diagnostic logic, not a particular radiocarbon prior or OxCal command.

9. Know the limits and common misconceptions

Back to contents

Bayesian modelling does not make a disputed context true, and a high agreement index does not validate every archaeological assumption. Posterior probability is conditional on the model, data and calibration information. Outlier probability is not the probability that a laboratory made an error. Wider sensitivity results are not failed analyses; they reveal what the evidence can support. A uniform phase prior is not assumption-free, and boundaries are not directly observed dates. Several plausible models can coexist when contextual evidence is incomplete. Software defaults, convergence thresholds and calibration-curve versions also matter. Specialist dating advice may be required, and final claims should remain tied to the sampled material and excavated context.

10. Audit the evidence-to-boundary chain

Back to contents

Archive sample identifiers, material descriptions, context records, pretreatment, laboratory reports, calibration curve and version, reservoir or offset choices, prior specifications, model code, random seeds where relevant, convergence and agreement diagnostics, rejected runs and every reported alternative. Link each ordering constraint to field evidence. Preserve the reason for including, excluding or reassigning a sample before and after modelling. Export unmodelled and modelled distributions and a table showing how key boundaries move across variants. An independent reviewer should be able to rebuild the chronology and locate exactly where an archaeological judgement narrows, shifts or reverses the result.

11. Frequently asked questions

Back to contents

Does a poor agreement value identify the bad date? No; it signals tension that may arise from data, association or model structure. Should an outlier always be removed? No; model the plausible failure process and retain the measurement unless independent evidence supports exclusion. Are boundary estimates observed dates? No; they are posterior quantities conditional on the model. Is the narrowest model best? No; precision without defensible assumptions is misleading. Must every alternative receive equal weight? No; explain evidence and plausibility, then show sensitivity to credible options. Does model agreement prove archaeological truth? No; it only assesses behaviour within the specified framework.

12. The gold standard keeps chronology answerable to archaeology

Back to contents

A strong chronological model does not silence conflict. It uses conflict to test the route from sample to event, from context to prior and from phase structure to boundary. The gold standard makes those assumptions visible, assigns outlier processes for archaeological reasons and reports sensitivity across a limited family of defensible models. Precision is retained where evidence supports it and surrendered where conclusions depend on fragile choices. The final chronology becomes a transparent argument that can be criticised, reconstructed and improved when new dates or field evidence arrive.