The Gold Standard of Archaeology | Plateau-Aware Sampling, Curve-Resolution Forecasts and Sequential Dating Design

1. Begin with the decision the chronology must support

Back to contents

State the archaeological question before counting samples. The aim might be to distinguish two occupation episodes, place a destruction horizon within a historical window, estimate the duration of a building phase or test whether a deposit is residual. Each requires different precision. Translate the question into a decision such as whether two phases overlap by more than a generation or which of two ordered models is supported. Define the population of datable events, relevant calendar window and acceptable uncertainty. A programme cannot be designed from the laboratory standard error alone. If the calibration curve cannot resolve the requested difference in that window, the team must change the question, add independent chronological information or acknowledge that radiocarbon alone is insufficient.

2. Forecast curve-limited resolution before excavation samples are spent

Back to contents

Use the appropriate calibration curve and software to simulate determinations across the plausible age range. Explore laboratory uncertainty, curve shape, sample spacing and prior order constraints. On a plateau, simulated calendar distributions may remain broad even when the radiocarbon measurement is precise. Near a steep section or diagnostic wiggle, ordered samples can separate time more effectively. Report the software and curve version, assumptions and the resolution metric used. A forecast is not a promised result; real samples add contamination, reservoir effects and context error. Its value is to show where precision comes from and prevent a team from consuming rare material in pursuit of calendar detail the curve cannot deliver.

3. Build a sampling frame around events, not convenient objects

Back to contents

List the depositional or biological events that could answer the question, then identify materials directly connected to them. Prefer short-lived, securely identified material when the target is a narrow event, while recognising that even seeds can be residual or intrusive. Long-lived charcoal may predate burning because of old wood; bone can reflect dietary reservoir effects; shell requires suitable marine correction. Record context, association, taxon, life span, taphonomy, treatment and expected failure risk. Spread samples across independent contexts and the sequence positions most likely to constrain boundaries. Ten fragments from one ambiguous feature do not equal ten independent chronological observations. The sampling frame should reveal which events lack any trustworthy datable material.

4. Design combinations that can beat a plateau

Back to contents

A plateau reduces information from isolated dates, but archaeology can add order, spacing and independent anchors. Select a sequence of short-lived samples with well-supported stratigraphic relations rather than many interchangeable samples from one layer. If tree-ring sequences or known-age offsets are available, wiggle matching may exploit the curve’s structure. Samples on both sides of a transition can constrain a boundary even when the transition itself falls on a plateau. These gains are conditional on correct ordering and sample-event association. Simulate alternative sequences, including one uncertain relation removed. If the promised precision disappears when a single contested link is relaxed, the design depends more on that link than on the radiocarbon measurements and should say so.

5. Use a pilot round to learn, not to declare victory

Back to contents

Submit a small, preselected set that spans the most informative contexts and includes quality-control material. The pilot should test preservation, pretreatment yield, contamination, approximate placement on the curve and compatibility with the proposed sequence. Define decisions in advance: which result would trigger reservoir investigation, additional sampling, a revised context model or stopping. Do not choose the next sample merely because it will support the most attractive pilot mode. Update the chronological model with all pilot results, including failures and outliers, and compare what remains unresolved with the original decision. Sequential design is adaptive, but adaptation must follow documented information needs rather than hindsight-driven storytelling.

6. Worked example: dating a two-phase waterfront structure

Back to contents

A fictional excavation seeks to learn whether two timber-and-hearth phases were separated by less than forty years. Simulations show that the expected interval lies on a calibration plateau: four isolated charcoal dates would usually leave a century-scale range. The team instead selects short-lived seeds sealed below the first floor, an articulated twig from its destruction layer and short-lived material beneath the rebuilt floor, preserving their order. A pilot on the two outer contexts locates the sequence near the plateau edge but reveals one sample with a possible aquatic reservoir signal. The team dates the twig and a second terrestrial seed rather than repeating the suspect material. The final model narrows the phase order but cannot guarantee a forty-year separation, so the report answers the broader question and retains the sharper claim as unresolved.

7. Set stopping rules in terms of information and decisions

Back to contents

Stop when the posterior or calibrated evidence answers the prespecified decision to the required level, when plausible additional samples are unlikely to change it, or when remaining uncertainty is dominated by context rather than measurement. A budget ceiling alone is not a scientific stopping rule, though it is a real constraint. Define measures such as expected reduction in boundary width, probability that phases overlap, or discrimination between two models. Also define futility: on a flat plateau, another similar isolated date may add almost no calendar resolution. Stopping avoids needless destruction of material and false confidence from a large sample count. Continuing is justified when a new sample contributes a different event, stronger order information or an independent control.

8. Changed-context transfer check: a cemetery with uncertain sequence

Back to contents

In a stratified building, physical order may be strong. In a cemetery, burials can be spatially related without a reliable temporal sequence, and human bone may carry dietary reservoir effects. The transferable method is still to define the decision, simulate curve resolution, select event-linked material and adapt after pilot results. The changed context means that assumed grave order should not be imposed merely to sharpen posteriors. Independent artefact seriation, ancient DNA kinship or stratigraphic intersections may contribute probabilistic information, but each needs its own uncertainty. Paired terrestrial material can investigate reservoir offsets. A sequential plan should prioritise graves whose dates discriminate competing cemetery models, not simply fill an even grid across the site.

9. Know the limits and common misconceptions

Back to contents

A larger sample count cannot guarantee calendar precision, and a smaller laboratory error does not eliminate calibration-curve ambiguity. Simulation results depend on the assumed calendar window, sequence and error model. Bayesian modelling cannot rescue incorrect stratigraphy or residual samples. A plateau is not a defect in calibration and should not be bypassed by reporting an uncalibrated age. Outlier models do not automatically identify why a sample disagrees. Pilot-informed adaptation can inflate confidence if models are changed to fit attractive results and the changes are not disclosed. Some chronological questions require dendrochronology, luminescence, historical evidence or a wider uncertainty statement rather than more radiocarbon dates.

10. Audit the design from forecast to final model

Back to contents

Archive the research question, decision threshold, sampling frame, rejected candidates, material identifications, context photographs, taphonomic reasoning, curve and software versions, simulation code or settings, expected and actual laboratory errors, pretreatment and quality data, pilot decision rules, all determinations, model versions, outlier settings and stopping rationale. Compare forecast and realised resolution. Record why each new sample was selected after the pilot and whether the choice was made before or after seeing intermediate posteriors. An independent reviewer should be able to rerun the forecast, trace each determination to an archaeological event and see whether the final precision came from measurements, stratigraphy or unreported selection.

11. Frequently asked questions

Back to contents

How many dates are enough? Enough to answer a defined question under the curve and context, not a universal number. Can more precision beat a plateau? Often only modestly; sequence or independent anchors may matter more. Should pilot dates enter the final model? Yes, unless there is a documented technical reason to exclude them. Is an outlier a bad laboratory result? Not necessarily; it may reflect residuality, intrusion, reservoir effects or a wrong model. When should sampling stop? When the decision is answered, further information is negligible or remaining uncertainty comes from context. Can simulations prove the final precision? No; they forecast performance under stated assumptions.

12. The gold standard dates the question, not the sample count

Back to contents

A strong dating programme anticipates the information limits of the calibration curve and makes each destructive measurement serve an archaeological decision. It selects independent event-linked material, combines dates with defensible order, learns through a declared pilot and stops when additional work has little value. Plateaus then become design constraints to manage honestly rather than inconveniences hidden by averages or oversized sample lists. The resulting chronology is as precise as the evidence allows and no more.