The Gold Standard of Archaeology | Inter-Laboratory Offsets, Shared Standards and Reproducibility across Radiocarbon Facilities

1. Define what reproducibility means in the project

Back to contents

Inter-laboratory reproducibility asks whether facilities measuring comparable material under stated procedures produce results consistent with a common target and credible uncertainty. It is not identical to repeatability within one instrument, nor does it prove that the sampled object dates the archaeological event of interest.

Write the target before distributing material. It may be fraction modern for a shared standard, a conventional radiocarbon age for a homogenised material or agreement after a specified pretreatment. If laboratories apply different pretreatments to heterogeneous archaeological fragments, disagreement blends material, preparation and measurement.

Define the decision connected to the comparison: qualify a facility, investigate a suspected offset, choose whether to combine dates, or estimate an additional laboratory component. The design and required precision depend on that decision. A comparison too weak to detect an archaeologically important offset should not be described as proof of equivalence.

First-principles check: Name the measurand, comparison level and largest offset that would matter archaeologically. Write the answer before looking at the preferred outcome. Then identify the rule, observation or calculation that could change it. This keeps the method tied to evidence rather than to hindsight.

Worked practice: Rewrite “we will send samples to two labs” as a testable reproducibility question. Keep the intermediate steps visible. A reader should be able to locate the input, apply the same rule and understand why a different result would follow if one fact changed.

Changed-context transfer: Change the goal from routine quality assurance to resolving a disputed historical boundary. Preserve the underlying principle but replace the setting-specific details. State which parts travel, which do not, and which new authority or measurement would be needed.

Evidence record: Keep the target quantity, decision, tolerable offset, material scope and power rationale. Retain uncertainty and disagreement instead of compressing them into a clean label. The record should let another person reconstruct the decision without relying on the memory or status of the original decision-maker.

2. Build a sample genealogy before splitting material

Back to contents

Draw the physical chain from source material to every submitted portion: object, fragment, homogenisation, chemical fraction, graphite target and repeated count. The genealogy reveals which aliquots can be treated as exchangeable and which share preparation or source uncertainty.

A bulk homogenised standard can hold material close to constant, but it may not resemble fragile archaeological collagen, charred residue or sediment. Separate fragments better represent routine work but add biological and depositional heterogeneity. Neither design is universally superior; they answer different questions.

Record mass, particle size, storage, visible contamination and allocation. Randomise aliquot labels when feasible and retain reserve material. If a split is too small for each facility’s normal protocol, the comparison may test a special low-mass workflow rather than routine performance.

First-principles check: Locate every shared and independent stage before assigning an error term. Write the answer before looking at the preferred outcome. Then identify the rule, observation or calculation that could change it. This keeps the method tied to evidence rather than to hindsight.

Worked practice: Create a branching genealogy for one bone divided across two laboratories and two pretreatments. Keep the intermediate steps visible. A reader should be able to locate the input, apply the same rule and understand why a different result would follow if one fact changed.

Changed-context transfer: Replace bone with a mixed charred residue that cannot be fully homogenised. Preserve the underlying principle but replace the setting-specific details. State which parts travel, which do not, and which new authority or measurement would be needed.

Evidence record: Retain photographs, mass, homogenisation, allocation, codes, custody and reserve locations. Retain uncertainty and disagreement instead of compressing them into a clean label. The record should let another person reconstruct the decision without relying on the memory or status of the original decision-maker.

3. Use shared standards without hiding shared uncertainty

Back to contents

Primary and secondary standards anchor radiocarbon measurement and allow performance to be tracked. When all laboratories calibrate against the same consensus material, its assigned value is shared information. Treating that value’s uncertainty as a fresh independent error for every result can distort comparisons.

Separate uncertainty in the consensus value from laboratory-specific preparation, background correction, normalisation and counting. For a direct difference between two laboratories measuring the same standard, some shared components may cancel while others remain. The covariance structure should be stated rather than assumed to be zero.

Standards also have commutability limits. A clean oxalic acid standard does not reproduce collagen extraction, charcoal cleaning or conservation contamination. Use a ladder: instrument standards, process standards and relevant archaeological-type materials. Agreement at one rung does not guarantee agreement at the next.

First-principles check: Ask which uncertainty components are common to both results and which can create a between-laboratory difference. Write the answer before looking at the preferred outcome. Then identify the rule, observation or calculation that could change it. This keeps the method tied to evidence rather than to hindsight.

Worked practice: Write an uncertainty budget for two labs using the same consensus material. Keep the intermediate steps visible. A reader should be able to locate the input, apply the same rule and understand why a different result would follow if one fact changed.

Changed-context transfer: Replace a pure standard with homogenised ancient bone processed from raw material. Preserve the underlying principle but replace the setting-specific details. State which parts travel, which do not, and which new authority or measurement would be needed.

Evidence record: Keep standard identity, batch, assigned value, covariance assumptions, preparation and traceability. Retain uncertainty and disagreement instead of compressing them into a clean label. The record should let another person reconstruct the decision without relying on the memory or status of the original decision-maker.

4. Design a blinded and balanced intercomparison

Back to contents

Blinding sample identity and expected age reduces the chance that a facility treats a material differently because it knows the target. Use neutral codes, random allocation and sufficient material for routine procedures. Include duplicates within laboratories when repeatability must be separated from reproducibility.

Balance the design across material types and age ranges. If one laboratory receives only clean young material and another receives old, low-carbon samples, their performance cannot be compared fairly. Include process blanks and appropriate backgrounds, and decide in advance whether results outside reporting limits enter the analysis.

The design should prevent selective resubmission after an inconvenient result. A staged investigation can be legitimate, but stage-one outcomes and the stage-two trigger must be preserved. Quality learning improves when laboratories receive the full anonymised comparison after initial reporting.

First-principles check: Hold material and information differences under control before attributing a difference to a facility. Write the answer before looking at the preferred outcome. Then identify the rule, observation or calculation that could change it. This keeps the method tied to evidence rather than to hindsight.

Worked practice: Create an allocation schedule with blinded codes, duplicate structure and balanced material classes. Keep the intermediate steps visible. A reader should be able to locate the input, apply the same rule and understand why a different result would follow if one fact changed.

Changed-context transfer: Add one rare sample that cannot be divided evenly. Preserve the underlying principle but replace the setting-specific details. State which parts travel, which do not, and which new authority or measurement would be needed.

Evidence record: Archive randomisation, code key, dispatch, material balance, blanks, deadlines and resubmission rules. Retain uncertainty and disagreement instead of compressing them into a clean label. The record should let another person reconstruct the decision without relying on the memory or status of the original decision-maker.

5. Estimate offsets at the laboratory level

Back to contents

For each material, compute the laboratory contrast with its standard uncertainty under the correct covariance. Across materials, use a model that can estimate a persistent facility offset and additional dispersion. A single extreme result should not automatically define the laboratory, while averaging many technically repeated counts must not create false independence.

Report both the estimated offset and its uncertainty. A result statistically compatible with zero may still allow offsets large enough to matter; an offset statistically different from zero may be archaeologically negligible. Compare the interval with the predeclared tolerance.

Check heterogeneity. A constant offset may fit poorly if disagreement depends on age, material or pretreatment. Plot residuals by material class, carbon mass and age. A facility-wide correction is dangerous when the problem is actually a material-specific process.

First-principles check: Match the statistical unit to the independent laboratory-by-material comparison. Write the answer before looking at the preferred outcome. Then identify the rule, observation or calculation that could change it. This keeps the method tied to evidence rather than to hindsight.

Worked practice: Estimate a weighted mean offset and test whether material-specific contrasts are compatible with it. Keep the intermediate steps visible. A reader should be able to locate the input, apply the same rule and understand why a different result would follow if one fact changed.

Changed-context transfer: Introduce a shared standard batch that creates correlation among several materials. Preserve the underlying principle but replace the setting-specific details. State which parts travel, which do not, and which new authority or measurement would be needed.

Evidence record: Retain contrasts, covariance, weights, model, heterogeneity, tolerance comparison and diagnostics. Retain uncertainty and disagreement instead of compressing them into a clean label. The record should let another person reconstruct the decision without relying on the memory or status of the original decision-maker.

6. Worked example: three facilities date shared bone and wood

Back to contents

Three facilities receive blinded aliquots of homogenised bone and single-year wood. Each reports conventional radiocarbon results with uncertainties and process-quality measures. Facility C is 34 years older than A and B for bone, but the wood results agree. Its collagen yield and elemental quality are acceptable, while a pretreatment record shows a stronger ultrafiltration step.

A naïve pooled offset would blur the pattern. The material interaction suggests that the contrast may arise in bone preparation rather than instrument calibration. The team requests a prespecified second stage: reserve bone aliquots are exchanged as extracted collagen and as raw powder. Agreement on collagen but renewed difference on raw powder localises the issue toward pretreatment.

The outcome is not “Facility C is wrong”. Its procedure may remove contamination that the other protocols retain. Archaeological interpretation now needs material history and known-age or well-characterised reference bone. The comparison has narrowed the causal location of disagreement without yet declaring which result better dates the target event.

First-principles check: Use cross-material and cross-stage patterns to localise the source of disagreement. Write the answer before looking at the preferred outcome. Then identify the rule, observation or calculation that could change it. This keeps the method tied to evidence rather than to hindsight.

Worked practice: Draw the inference tree from stage-one interaction to the two stage-two exchanges. Keep the intermediate steps visible. A reader should be able to locate the input, apply the same rule and understand why a different result would follow if one fact changed.

Changed-context transfer: Reverse the pattern so all materials from Facility C are older by the same amount. Preserve the underlying principle but replace the setting-specific details. State which parts travel, which do not, and which new authority or measurement would be needed.

Evidence record: Preserve raw reports, quality indicators, stage trigger, exchanged fractions, new results and alternative explanations. Retain uncertainty and disagreement instead of compressing them into a clean label. The record should let another person reconstruct the decision without relying on the memory or status of the original decision-maker.

7. Distinguish an offset from excess dispersion

Back to contents

A laboratory can have no stable mean offset yet show more scatter than its quoted uncertainties predict. This is excess dispersion. Conversely, results can be precise but consistently displaced. The two failures require different remedies and should not be compressed into a pass-or-fail score.

Use standardised residuals or a hierarchical model to assess both location and scale. Investigate whether quoted uncertainties include preparation, background and long-term variability. A facility may report only analytical counting uncertainty while another includes a broader quality component, making visual comparisons misleading.

If extra dispersion is supported, incorporate it transparently in future combinations or revise the laboratory uncertainty model. Do not inflate every archaeological date indefinitely without locating the affected period, material or workflow. Quality-control records can show whether the problem was episodic or persistent.

First-principles check: Test mean displacement and unexplained variability as separate parameters. Write the answer before looking at the preferred outcome. Then identify the rule, observation or calculation that could change it. This keeps the method tied to evidence rather than to hindsight.

Worked practice: Compare two simulated laboratories: one with +25-year offset and one with doubled scatter. Keep the intermediate steps visible. A reader should be able to locate the input, apply the same rule and understand why a different result would follow if one fact changed.

Changed-context transfer: Add a known maintenance period that divides the data into two regimes. Preserve the underlying principle but replace the setting-specific details. State which parts travel, which do not, and which new authority or measurement would be needed.

Evidence record: Keep residuals, variance components, quoted uncertainty definitions, time periods and correction policy. Retain uncertainty and disagreement instead of compressing them into a clean label. The record should let another person reconstruct the decision without relying on the memory or status of the original decision-maker.

8. Changed-context transfer: intercomparison of marine shell

Back to contents

Marine shell adds reservoir effects and species-specific ecology. Two laboratories can agree perfectly on measured fraction modern yet the calibrated archaeological ages can differ if they apply different reservoir corrections or curve choices. Reproducibility must therefore be assessed at clearly separated stages.

First compare laboratory outputs on the conventional radiocarbon scale using shared conventions. Then compare any correction inputs and calibration models as a distinct inferential layer. A local reservoir value estimated from uncertain paired samples may be common to all calibrated results and should not be mistaken for laboratory performance.

The same principle transfers: trace the measurand, shared inputs and decision. What changes is the environmental model. Agreement across facilities does not eliminate ecological uncertainty, and disagreement after calibration may arise even when measurement was reproducible.

First-principles check: Separate physical measurement from environmental correction and chronological modelling. Write the answer before looking at the preferred outcome. Then identify the rule, observation or calculation that could change it. This keeps the method tied to evidence rather than to hindsight.

Worked practice: Create parallel audit columns for laboratory age, reservoir correction and calibrated age. Keep the intermediate steps visible. A reader should be able to locate the input, apply the same rule and understand why a different result would follow if one fact changed.

Changed-context transfer: Move from shell to freshwater fish affected by a hard-water reservoir. Preserve the underlying principle but replace the setting-specific details. State which parts travel, which do not, and which new authority or measurement would be needed.

Evidence record: Retain species, habitat, conventional result, correction source, curve, covariance and stage-specific comparisons. Retain uncertainty and disagreement instead of compressing them into a clean label. The record should let another person reconstruct the decision without relying on the memory or status of the original decision-maker.

9. Know the limits and common misconceptions

Back to contents

One agreed standard does not prove performance across all materials and ages. Two laboratories are not independent if they share pretreatment chemistry, graphite preparation, calibration standards or training. A small difference is not evidence of equivalence unless the design had power for the chosen tolerance.

A consensus value is not infallible. It may be estimated from participating laboratories and carry uncertainty or method dependence. Excluding “outliers” can narrow the consensus while hiding real heterogeneity. Use prespecified robust methods and retain all submitted results with reasons for any exclusion.

Inter-laboratory agreement also cannot repair poor archaeological association. Several facilities can reproducibly date old wood, intrusive roots or conservation material. The final claim must keep measurement performance and context validity separate.

First-principles check: State what the comparison can rule out and which risks remain shared. Write the answer before looking at the preferred outcome. Then identify the rule, observation or calculation that could change it. This keeps the method tied to evidence rather than to hindsight.

Worked practice: Take a reassuring pattern of agreement and name three common-mode failures that could produce it. Keep the intermediate steps visible. A reader should be able to locate the input, apply the same rule and understand why a different result would follow if one fact changed.

Changed-context transfer: Apply the same logic to a museum object with undocumented conservation. Preserve the underlying principle but replace the setting-specific details. State which parts travel, which do not, and which new authority or measurement would be needed.

Evidence record: Publish power, exclusions, shared risks, material limits and the archaeological association statement. Retain uncertainty and disagreement instead of compressing them into a clean label. The record should let another person reconstruct the decision without relying on the memory or status of the original decision-maker.

10. Audit the full intercomparison chain

Back to contents

The audit joins sample selection, genealogy, homogenisation, allocation, shipping, pretreatment, measurement, standards, normalisation, reporting and analysis. Codes must map without ambiguity while blinding is preserved until the reporting deadline.

Reproduce every contrast from original laboratory certificates. Verify units, conventions, uncertainty level and correction fields. Keep the first submitted result, not only a later corrected certificate. Corrections may be valid, but the reason and timing are part of quality performance.

Close with a decision table: accepted routine agreement, material-specific qualification, temporary investigation, revised uncertainty or no combination. The action should follow the predeclared tolerance and evidence, not the reputation of a facility.

First-principles check: Trace each reported value to physical material, process record and comparison calculation. Write the answer before looking at the preferred outcome. Then identify the rule, observation or calculation that could change it. This keeps the method tied to evidence rather than to hindsight.

Worked practice: Reconstruct one laboratory contrast from dispatch code through final model. Keep the intermediate steps visible. A reader should be able to locate the input, apply the same rule and understand why a different result would follow if one fact changed.

Changed-context transfer: Repeat after a code-key error is discovered during unblinding. Preserve the underlying principle but replace the setting-specific details. State which parts travel, which do not, and which new authority or measurement would be needed.

Evidence record: Archive custody, certificates, revisions, standards, code, model, review and action. Retain uncertainty and disagreement instead of compressing them into a clean label. The record should let another person reconstruct the decision without relying on the memory or status of the original decision-maker.

A robust comparison plan should include a detectable-difference calculation. Choose an archaeologically meaningful offset, plausible within-laboratory variability and the number of independent materials. Simulate the full hierarchy rather than treating repeated counts as separate samples. Report the probability of detecting the chosen offset and the interval width expected under no displacement. When power is low, “no significant difference” must not be translated into equivalence.

Use control charts carefully for long-term quality records. A stable reference material can reveal drift or a step change, but limits should reflect the facility’s verified historical process and material. Recalculating limits after a problematic period can normalise the failure. Mark instrument upgrades, chemistry changes and standard batches. A project comparison should ask whether the current archaeological samples fall within the validated operating regime rather than relying on an all-time facility average.

Unblinding deserves its own protocol. Freeze submitted results, uncertainty definitions and exclusions before revealing facility identities or consensus values. Then circulate a discrepancy table and allow factual corrections with version history. Corrections for transcription or unit errors differ from rerunning a sample after seeing the target. Both can be useful, but they answer different quality questions and should appear as separate stages.

When a correction factor is proposed, validate it outside the materials used to estimate it. A leave-one-material-out analysis or later proficiency round can reveal overfitting. Propagate uncertainty in the correction and restrict its scope by time, material and workflow. The cleanest action may be to keep results separate rather than force comparability.

The archaeological publication should connect the laboratory findings to chronology without overreach. If a 25-year offset is small relative to a calibration plateau, it may not change the calendar range; in a steep curve segment or wiggle-matched sequence, it may matter greatly. Run the chronological model with and without the supported offset or extra dispersion. Report which ordering, duration or historical claim changes. This makes the quality result decision-relevant.

Before publication, run a leave-one-material-out review. Re-estimate the facility offset after removing each shared material in turn and compare the result with the full model. A large swing shows that one material is carrying the conclusion. Inspect whether that material has unusual carbon mass, age, pretreatment, heterogeneity or consensus uncertainty. Report the influence rather than averaging it into silence. Repeat the review after excluding within-laboratory technical duplicates so that precision is not borrowed from pseudo-replication.

Build a causal discrepancy matrix for every large contrast. Rows should include source heterogeneity, storage and conservation, pretreatment, combustion, graphitisation, background, fractionation correction, measurement, data reduction and transcription. Columns should show the predicted direction, supporting evidence, contradicting evidence and next discriminating test. This prevents the investigation from defaulting to the most familiar explanation. It also shows when several explanations remain observationally equivalent.

The stop rule for investigation matters. Additional destructive work is justified only when it can distinguish live explanations or materially change the archaeological conclusion. Stop when the remaining contrast is below the declared tolerance, when a cause has been isolated with adequate evidence, or when new sampling cannot resolve shared uncertainty. Record remaining ambiguity. Reproducibility work is strongest when it knows what another measurement can add and when it cannot.

A final peer review should include both a laboratory scientist and an archaeologist who did not build the model. Ask each to state, independently, what physical event the reported age is claimed to represent and which comparison supports facility agreement. Differences in their answers expose vocabulary that is technically familiar but semantically unstable. Resolve those differences in the article, not in an unpublished meeting. Then check that tables, captions and data files use the same laboratory codes, material units, uncertainty convention and reference value. Small transcription inconsistencies can defeat an otherwise careful intercomparison because later researchers cannot tell whether a difference is physical or clerical.

11. Frequently asked questions

Back to contents

How many laboratories are enough? Two can reveal disagreement but give limited information about which facility or process is responsible; more facilities and shared references improve diagnosis. Must material be homogeneous? For a pure measurement comparison, as close as practicable; for routine-workflow comparison, heterogeneity must be measured or modelled.

Can results be combined if they pass a consistency test? Only if they estimate the same target and independence assumptions are defensible. Should a known offset be subtracted? Only with stable, externally supported evidence and propagated uncertainty. Does accreditation guarantee agreement? It supports quality systems but does not replace project-specific comparability.

What should be published? Material genealogy, blind design, facility codes or identities as appropriate, raw results, uncertainties, standards, contrasts, exclusions and limits. Why use shared standards? They help isolate measurement performance, provided their common uncertainty and commutability are handled honestly.

First-principles check: Answer by naming the comparison target and decision. Write the answer before looking at the preferred outcome. Then identify the rule, observation or calculation that could change it. This keeps the method tied to evidence rather than to hindsight.

Worked practice: Prepare a one-page explanation for an archaeologist choosing whether to combine dates. Keep the intermediate steps visible. A reader should be able to locate the input, apply the same rule and understand why a different result would follow if one fact changed.

Changed-context transfer: Explain the same result to a conservator deciding whether another destructive sample is justified. Preserve the underlying principle but replace the setting-specific details. State which parts travel, which do not, and which new authority or measurement would be needed.

Evidence record: Maintain reviewed answers linked to methods, evidence and current facility guidance. Retain uncertainty and disagreement instead of compressing them into a clean label. The record should let another person reconstruct the decision without relying on the memory or status of the original decision-maker.

12. The gold standard localises disagreement

Back to contents

A strong intercomparison begins with a measurable target, material genealogy and balanced blinded design. It uses standards at several levels, tracks shared uncertainty and estimates laboratory offset separately from excess dispersion and material interaction.

When results disagree, the team asks where the pathways diverged. Crossed exchanges of raw material, extracted fractions and targets can localise preparation or measurement effects. The investigation remains open to the possibility that a different procedure is better rather than merely different.

The final test is layered: are the physical measurements reproducible, are corrections comparable and does the dated material answer the archaeological question? Keeping those layers separate prevents a clean laboratory comparison from becoming false chronological certainty.

First-principles check: End with a stage-specific conclusion rather than a single reliability label. Write the answer before looking at the preferred outcome. Then identify the rule, observation or calculation that could change it. This keeps the method tied to evidence rather than to hindsight.

Worked practice: Summarise the comparison in three sentences: target, detected pattern and remaining uncertainty. Keep the intermediate steps visible. A reader should be able to locate the input, apply the same rule and understand why a different result would follow if one fact changed.

Changed-context transfer: Imagine the preferred site chronology depends on the laboratory judged less precise. Preserve the underlying principle but replace the setting-specific details. State which parts travel, which do not, and which new authority or measurement would be needed.

Evidence record: Publish design, results, model, localisation, action and the boundary of the archaeological claim. Retain uncertainty and disagreement instead of compressing them into a clean label. The record should let another person reconstruct the decision without relying on the memory or status of the original decision-maker.