<!-- ARTICLE_BODY_START -->
The decision is about agreement
In October 2026, the General Conference on Weights and Measures adopted a resolution on the future definition of the second. It did not select a new atom or replace the existing unit. Instead, it requested work toward a proposal for the 2030 conference, including a proposed implementation date. The resolution identifies unfinished work: some optical-clock comparisons remain inconsistent, some operational criteria remain unmet, and agreement has not been reached on a single species versus an ensemble of species. Those are substantive measurement questions, not merely a scheduling delay. CGPM Resolution 4, 2026.
A laboratory can demonstrate an extraordinarily small uncertainty while the international community still lacks enough evidence to make that device's transition the basis of a universal unit. The distinction is easier to understand if we imagine receiving a beautifully precise measuring instrument with no way to compare it against another independent instrument. Its display may be steady. Its operator may understand every component. Yet a shared standard requires other people, elsewhere, to recover the same result through their own apparatus.
For optical clocks, that comparison involves much more than sending a time reading over the internet. The frequencies being compared can differ by hundreds of trillions of cycles per second. The light carrying information between laboratories acquires disturbances along its path. Clocks at different gravitational potentials legitimately run at different rates. Measurements made on different days may depend on different operating conditions. Even an apparently redundant comparison can reuse the same atomic data or calibration assumption.
This is why the network matters. Its purpose is to expose disagreement, distinguish its possible causes and show that the proposed definition can survive outside one exceptionally well-controlled room. The most informative question is often not which clock has the smallest number in its uncertainty column. It is which combination of clocks, links and independent tests makes that number believable.
A unit and its physical realization
The second currently fixes the numerical value of the unperturbed caesium-133 ground-state hyperfine transition frequency at 9,192,631,770 hertz. That relationship is exact by definition. A real apparatus attempting to realize it has uncertainty. The word unperturbed matters because laboratory atoms are never simply abstract atoms isolated from motion, fields and their surroundings. BIPM definition of the second.
Consider the difference between defining a circle and drawing one. The mathematical definition is not rendered uncertain by a slightly imperfect drawing. Conversely, the perfection of the definition does not guarantee that a drawing is exact. Measurement standards have a similar division of labor. An agreed definition specifies the reference; an experiment demonstrates how closely a physical procedure realizes it.
An optical transition supplies a much faster oscillation than the microwave transition used in caesium clocks. In favorable circumstances, resolving the same absolute frequency difference against a higher carrier frequency gives a smaller fractional difference. But frequency alone does not determine performance. A broad, poorly controlled optical resonance would not automatically outperform a carefully measured microwave resonance. The useful combination is a high frequency, a narrow measurable transition, a stable interrogation system and well-characterized perturbations.
Redefinition would change which physical relationship anchors the unit. It should not make an everyday second noticeably longer or shorter. Continuity is essential because existing measurements, instruments and records already express quantities in seconds. The challenge resembles replacing a reference instrument without introducing an unintended scale change, except the replacement concerns a definition that underpins measurements throughout science and engineering.
This also explains why an optical frequency can be known remarkably well relative to another optical frequency while its numerical value in present SI hertz is less certain. The ratio can bypass the caesium realization. Expressing either frequency in the current unit reconnects the experiment to that realization and its uncertainty. There is no contradiction: the two measurements answer different questions.
The atom interrogates the laser
An atomic clock is often pictured as an atom emitting perfectly regular ticks that an electronic counter simply collects. A more useful picture is a controlled interrogation. A laser is tuned near an atomic transition, the experiment measures how the atoms respond, and feedback adjusts the laser toward the resonance. The continuous oscillator and the atomic reference perform different jobs.
If the laser frequency is slightly above resonance, an appropriately designed probing sequence produces a different response from the one obtained slightly below it. Repeated measurements create an error signal. A controller uses that signal to steer the oscillator. The response has statistical fluctuations, and the controller has finite bandwidth. Preparing atoms, probing them and reading out their state can leave intervals during which the reference is not being interrogated.
A hypothetical clock makes the division clear. Suppose its laser remains beautifully stable for one minute but drifts afterward. Atomic measurements can correct the long-term drift. Now suppose the atomic measurement is very accurate but available only briefly each hour. The laser must carry the frequency between those measurements. Neither component's isolated specification tells us the performance of the complete clock.
Long interrogation can sharpen frequency discrimination, but it demands that the laser remain coherent enough during the measurement. More atoms can improve statistical information, but atoms may interact or experience different environments. Stronger confinement can reduce motion while introducing light shifts. Clock design is therefore a sequence of tradeoffs, not a contest to maximize a single parameter.
Two broad architectures illustrate the alternatives. A trapped-ion clock can interrogate an individual charged atom held by electromagnetic fields. An optical lattice clock can interrogate many neutral atoms confined in a light-generated structure. Each architecture converts a difficult problem into a different set of controllable problems. The relevant comparison includes systematic effects, stability, reliability and the experimental evidence supporting all three.
A small uncertainty is an argument with components
In July 2025, researchers reported an aluminium-ion clock with fractional systematic uncertainty of 5.5 × 10⁻¹⁹. A magnesium ion assisted cooling and readout. The work connected improved laser stability, reduced unwanted ion motion and a better vacuum environment to the evaluated clock performance. Its uncertainty table separates effects rather than treating the headline number as a directly observed property. This is a dated example of a particular instrument, not a claim about the world record in October 2026. Marshall and colleagues, Physical Review Letters.
The distinction between a correction and its uncertainty is fundamental. Suppose a hypothetical environmental effect shifts a clock by 100 units, and repeated controlled experiments establish that correction with uncertainty of one unit. The corrected result can be much more accurate than the raw shift suggests. Another effect might shift the result by only two units but be poorly known, contributing comparable uncertainty. Sorting effects by their raw size would misidentify the limiting problem.
Combining uncertainty contributions also requires their relationships. If two terms are independent, their variances can be added. If they depend on the same temperature calibration, simply adding them as though unrelated can understate or overstate the total. The covariance is part of the measurement, not an optional statistical decoration.
An uncertainty budget is consequently an organized claim about what has been measured, corrected and left uncertain. It should tell another laboratory which assumptions to test. Does a shift change as predicted when the trap conditions change? Does the inferred correction survive a different measurement procedure? Is a supposedly negligible term bounded by an experiment or merely omitted because it was inconvenient?
A single well-documented budget provides strong local evidence. Agreement with an independently built apparatus adds a different kind of evidence. It tests whether something important escaped the budget entirely. Neither replaces the other: two inadequately evaluated clocks can agree, and two carefully evaluated clocks can reveal a discrepancy worth investigating.
The room remains part of the clock
The surrounding apparatus radiates electromagnetic energy. That thermal environment can shift the atomic transition. In a 2024 strontium lattice-clock study, the researchers reported total systematic uncertainty of 8.1 × 10⁻¹⁹ and revised the blackbody-radiation correction through a measurement of an atomic-state lifetime. They also evaluated a magnetic-field coefficient. The result shows that improving a clock may require better knowledge of atomic response as well as better control of the apparatus. Aeppli and colleagues, 2024.
A useful simplified thermal model writes a leading shift as a coefficient multiplied by temperature to the fourth power. This is only an illustrative leading-order model, not a complete correction for a particular clock. Differentiating it gives a fractional change in the shift of approximately four times the fractional temperature change. Around 300 kelvin, a one-kelvin temperature error would therefore change this simplified shift by about 1.33 percent. Whether that is acceptable depends on the shift's size and the desired frequency uncertainty.
The thermometer reading alone is not the whole environment. Imagine an atom inside a chamber whose walls are mostly near one temperature but which has a small opening toward a hotter component. The radiation reaching the atom depends on geometry and surface properties as well as a convenient wall sensor. A thermal model has to describe what the atom experiences, not merely what is easy to measure.
The same reasoning applies to electric and magnetic fields. Averaging two atomic states can cancel a leading sensitivity while leaving a smaller higher-order term. Operating at a special trapping wavelength can reduce the differential light shift without making every intensity-dependent effect vanish. A cancellation is a relationship with a domain of validity; it is not permission to stop characterizing the variable.
This is also why nominally identical clocks need not share identical corrections. Their atoms may be the same species, but their windows, trap geometry, field gradients, heating and operating sequences differ. Reproducibility means recovering the same unperturbed reference after those differences are accounted for.
Stability tells us how long a comparison needs
Systematic uncertainty and instability describe different limitations. An instrument can fluctuate substantially around a well-understood mean, or it can be exceptionally quiet while remaining offset. The first may improve with averaging. The second demands investigation of its bias. A comparison needs enough statistical precision to say something useful about the systematic budgets being tested.
For an explicitly hypothetical white-frequency-noise model, suppose fractional instability is 10⁻¹⁵ divided by the square root of averaging time in seconds. At 100 seconds it becomes 10⁻¹⁶; at 10,000 seconds, 10⁻¹⁷; at one million seconds, 10⁻¹⁸. That last interval is about 11.6 days. Improving the initial instability by a factor of ten would reduce the required averaging time for the same target by a factor of one hundred.
Those calculations are conditional. They assume the same noise behavior persists and that useful data are continuously available. Drift, environmental correlations, interrupted operation or a systematic floor can change the result. A line extrapolated across several orders of magnitude should not be confused with an achieved measurement.
The popular statement that a clock would lose only one second over an enormous number of years is another conditional calculation. A fractional frequency offset of 10⁻¹⁸, if constant, accumulates one second after 10¹⁸ seconds, roughly 31.7 billion Julian years. No laboratory has run that clock for such an interval. The conversion translates a fractional quantity into an intuitive duration; it does not establish lifetime, maintenance-free operation or immunity to changing conditions.
For a network comparison, the practical question is more immediate. How many hours of simultaneous usable data can the laboratories obtain? Which noise dominates during those hours? Does the statistical estimate remain valid after outages are removed? Can another campaign reproduce the result? These operational details determine whether an impressive local clock can actually test another one.
Counting light requires an intermediary
Optical frequency combs connect optical frequencies to frequencies that electronics can measure and can also compare widely separated optical transitions. In the standard comb relation, a tooth frequency equals an integer multiple of the pulse repetition frequency plus an offset frequency. Measuring and controlling these quantities gives an organized set of reference frequencies rather than a featureless band of light. NIST's frequency-comb program description, updated March 2025.
The integer matters. As a deliberately simplified arithmetic example, a repetition frequency of 100 megahertz and a tooth index of four million produce a frequency near 400 terahertz before adding the offset. A nearby optical signal can produce a much lower beat frequency against that tooth. The measurement must identify the correct tooth and the appropriate sign of the beat. A highly precise counter cannot rescue a mistaken frequency relationship.
This intermediary should not be imagined as an independent authority that certifies an atom. It transfers a relationship. If the reference laser has an uncorrected shift, a perfect comb can faithfully transfer that shifted frequency. Conversely, an imperfect transfer can make two sound references appear inconsistent. The entire chain must be checked.
One valuable design principle follows: compare by more than one route when possible. If two frequency-conversion arrangements produce the same ratio, their agreement constrains errors specific to one arrangement. If both depend on the same oscillator or electronics, the agreement may leave a shared error untouched. A block diagram is useful only when it reveals these shared dependencies.
Frequency-ratio work also changes what must be counted. Comparing two optical clocks directly can avoid requiring an exceptionally accurate numerical value in microwave hertz at every intermediate step. The ratio is dimensionless. That does not remove the need for careful engineering, but it separates the task of comparing optical references from the task of expressing their absolute frequencies in the current SI second.
The path can imitate a frequency difference
A link carries phase information through a physical environment. If its effective path length changes, the received phase changes. The time derivative of phase is related to frequency, so a changing path can look like a frequency difference between the endpoints. A fixed propagation delay and a varying propagation delay are therefore different problems.
A useful analogy is watching evenly spaced marks move past a camera through a flexible relay. If the relay stretches during observation, the arrival spacing changes even though the source marks remain regular. In an optical experiment the details are wave-based rather than mechanical, but the distinction between source behavior and transmission behavior remains essential.
In the 2021 Boulder Atomic Clock Optical Network study, aluminium-ion, ytterbium and strontium clocks were compared using optical fiber and a 1.5-kilometre free-space link. Reported ratio uncertainties were between 6 × 10⁻¹⁸ and 8 × 10⁻¹⁸. Using alternative transmission paths helped test the network separately from the clocks. These were measured comparison results, not a claim that every atmospheric link reaches that performance. BACON Collaboration, Nature.
An active link can use information from a return signal to estimate disturbances and apply a correction. The effectiveness depends on how well the outward and return measurements represent the relevant disturbance and on the time needed for information to propagate. Rapid changes cannot be corrected with advance knowledge that the system does not possess. Paths that differ between directions can leave residual effects.
The engineering interpretation of a successful link therefore includes length, environment, averaging time and operating conditions. A quiet laboratory fiber, a long installed route and a turbulent atmospheric path pose different challenges. Calling all of them optical connectivity conceals the quantities that determine whether the comparison is trustworthy.
A communication link that reliably transmits ordinary data is not automatically a precision frequency-transfer link. Recovering bits and preserving phase with a quantified uncertainty are distinct requirements, even when some physical infrastructure is shared.
International comparison changes the test
A manuscript revised in September 2026 reports a two-month European campaign conducted in 2023. Seven optical clocks at four national metrology institutes were connected through an optical fiber network. Reported ratio uncertainties ranged from 7.7 × 10⁻¹⁸ to 6.1 × 10⁻¹⁷. Two independently developed ytterbium-ion clocks agreed within the lower uncertainty. The study also discusses discrepancies involving other comparisons. Its publication date must not be mistaken for the date of the measurements. International Optical Clock Comparison Using the European Optical Fiber Network.
The international aspect matters because independence comes in degrees. Two clocks in one laboratory may share a laser, local environmental measurements, atomic coefficients or analysis software. Moving the comparison across institutions can diversify some of those dependencies. It also introduces new ones: longer links, different geodetic references, nonidentical schedules and more complicated data alignment.
Suppose two laboratories compare the same atomic transition and obtain a small difference. Several explanations remain possible. One clock may have a residual environmental shift. The transfer path may be biased. The gravitational correction may differ. The statistical uncertainty may be underestimated because successive observations are correlated. A good comparison is designed to separate these possibilities rather than assigning fault to whichever apparatus has the larger published uncertainty.
Changing one element at a time can help. Compare the laboratories through a second route. Repeat after modifying a suspected environmental condition. Add a third clock with different sensitivities. Reanalyze the same raw data with an independent implementation. None is a universal solution, but together they transform disagreement from an embarrassment into information.
A measurement community preparing a definition needs this ability to investigate. If every new discrepancy requires assuming the definition itself is correct and every dissenting instrument is wrong, the validation has become circular. Independent realizations must retain enough diagnostic freedom to challenge the proposed reference before it becomes the agreed anchor.
Ratios form a constrained network
For three ideal frequencies A, B and C, the ratios obey a simple identity: A divided by B, multiplied by B divided by C, equals A divided by C. This gives a closure test. Measurements along different paths should agree within their combined uncertainty after accounting for shared information.
Take an intentionally elementary example with ratios A/B = 2 and B/C = 3. Their product predicts A/C = 6. A direct measurement near 6 supplies a check only to the extent that it contains independent information. If the supposed direct result was calculated from those same two inputs, agreement is guaranteed and proves nothing additional.
Real clock networks operate near precisely specified reference ratios rather than these small integers. Small fractional deviations can be represented approximately as additive quantities. That makes loops in the network particularly useful: their signed deviations should sum consistently. But a network can have more measured edges than independent information if several edges reuse the same underlying clock record.
The analysis behind the CIPM's 2021 recommended-frequency update used an overdetermined body of published comparisons. Its authors describe three independent calculations using two algorithms and attention to correlations. This is a concrete example of why recommended frequencies are not obtained by averaging every published number as though it were an unrelated vote. Margolis and colleagues, account posted January 2024.
Here is a hypothetical covariance problem. Two reported ratios each contain an uncertainty of three units, entirely from one shared reference. Averaging them does not reduce that common contribution to three divided by the square root of two. The shared uncertainty remains three. Independent noise on each result can shrink with averaging, but the common reference term cannot.
A dense-looking network can consequently be less informative than a smaller network with genuinely independent references and routes. The scientific value lies in which uncertainties a new edge can challenge. Counting connections without examining their dependencies rewards redundancy that may be mostly cosmetic.
Gravity is part of the comparison
General relativity predicts a frequency difference between otherwise identical clocks at different gravitational potentials. Near Earth's surface, a useful small-height approximation gives the fractional difference as gravitational acceleration multiplied by height difference, divided by the square of the speed of light. The higher clock runs faster in this approximation. The calculation concerns a potential difference; geometrical height is only an approximate local proxy.
Using 9.81 metres per second squared and the exact speed of light, an illustrative one-metre height difference corresponds to about 1.09 × 10⁻¹⁶ in fractional frequency. One centimetre corresponds to about 1.09 × 10⁻¹⁸. A comparison approaching that level cannot casually ignore where the atoms sit relative to the relevant gravitational reference.
This is not a defect to remove from relativity. It is the predicted physical behavior of the clocks. Whether it is a correction or a signal depends on the measurement's purpose. To compare atomic transitions as references, the potential difference must be accounted for. To infer a potential difference from clocks, the atomic and instrumental effects must instead be controlled well enough to isolate the gravitational contribution.
A strontium experiment reported a frequency gradient consistent with gravitational redshift across a millimetre-scale atomic sample. The result concerns differential measurements within that sample, not two indefinitely operating field clocks with identical absolute uncertainty. It demonstrates how spatial variation can become measurable even inside an apparatus once differential sensitivity is sufficiently high. Bothwell and colleagues, manuscript posted September 2021.
The practical implication extends to the definition of the measurement location. A clock's surveyed reference point may not be exactly where the atoms are interrogated. A transportable apparatus may settle differently after installation. Local mass distributions and time-dependent environmental effects can matter at sufficiently demanding levels. Precision requires specifying the physical point and the reference model, not simply attaching a street address to a frequency record.
A clock can measure potential, with qualifications
Chronometric geodesy reverses the usual correction problem. Instead of obtaining gravitational potential information to compare clocks, it uses a comparison between suitable clocks to infer potential difference. That can complement geometrical surveying and gravity measurements. It does not convert every frequency discrepancy into a trustworthy elevation change.
Imagine two hypothetical sites with a measured fractional difference of 2 × 10⁻¹⁸ after instrumental corrections. Under the constant-local-gravity approximation, the equivalent height difference is roughly 1.83 centimetres. That number is a model-based interpretation. If an unrecognized thermal shift contributed half the frequency difference, the inferred height would be wrong by roughly half. A sensitive clock is not automatically a selective gravitational sensor.
This makes reference choice important. If one site changes, a comparison reports a relative change between the sites. It cannot, by itself, establish which site moved or whether the potential changed because of redistributed mass rather than simple vertical displacement. Additional observations supply the missing context.
The distinction between height systems also becomes consequential. A geometrical coordinate measured relative to a reference ellipsoid and a height related to gravity are not interchangeable labels. The clock responds to potential, including the relevant relativistic framework, rather than to whichever elevation number happens to appear on a map. Combining clock data with surveying therefore requires a shared reference convention.
There is a useful experimental strategy in this limitation. Measure the same sites with more than one method and ask whether the methods agree after their distinct sensitivities are modeled. A disagreement may reveal something about the clock, the link, the survey or the environmental model. The goal is not to declare the new instrument superior by definition, but to exploit the different physics each method brings.
For public-interest applications, repeatability under field conditions is at least as important as a striking equivalent-height calculation. An inferred change is valuable only when the supporting record establishes that the instrument's own environment and calibration did not change in a confounding way.
Transporting a reference creates another experiment
A 2025 study described a transportable strontium lattice clock with total systematic uncertainty of 2.1 × 10⁻¹⁸. Its blackbody contribution was controlled at 4.0 × 10⁻¹⁹ using a characterized cold shield. The apparatus was transported and operated at different locations. Its absolute frequency comparison against caesium fountains had a larger uncertainty, 1.9 × 10⁻¹⁶, illustrating the distinction between local optical evaluation and anchoring to the current SI realization. Nosske and colleagues, July 2025 manuscript.
Transport can connect laboratories without relying exclusively on a permanent high-performance link between them. But it exchanges one set of difficulties for another. Components experience acceleration and temperature changes. Alignment may need recovery. A field site may provide different power, vibration, thermal conditions or access to surveying. The transported reference must be evaluated as an operating system after arrival.
A useful hypothetical campaign compares the traveling clock against a home reference, moves it to a second site, performs the target comparison, and then returns for another home comparison. The return measurement can reveal a persistent change incurred during transport. It does not guarantee that no temporary shift existed only at the destination. The destination's environmental record remains necessary.
Similarly, a stable transport case is not the same as a demonstrated transportable standard. The evidence should include recovery time, usable operating periods and the conditions under which the uncertainty budget was re-established. A device that performs superbly after several weeks of expert rebuilding answers a different operational need from one that resumes validated operation rapidly.
For an international system, transportability can broaden participation by reducing dependence on a particular fixed route. It can also place substantial demands on staffing and local infrastructure. Evaluating those demands is part of metrology, because a reference that only a few specialists can recover under ideal conditions may have limited value as a widely reproducible realization.
A frequency reference must support a continuous timescale
A frequency measurement tells us a rate over an interval. A timescale also requires continuity: keeping track of elapsed time while instruments are unavailable, serviced or compared intermittently. The most accurate frequency standard does not automatically provide the best uninterrupted time output.
A 2019 demonstration combined a stable optical carrier with a strontium lattice clock and reported an estimated time error of 48 ± 94 picoseconds over 34 days. Its longer-term performance discussion included projections, which should remain projections rather than being converted into subsequent operational achievements. The experiment illustrates an architecture for connecting intermittent reference information to a continuing timescale. Milner and colleagues, 2019.
Suppose, hypothetically, that an oscillator's rate is uncertain by 10⁻¹⁵ during a day without calibration. Multiplying by 86,400 seconds gives an accumulated time uncertainty scale of 86.4 picoseconds if that rate uncertainty acts as a constant offset across the day. Different noise processes produce different time-growth laws; this simple multiplication is not a general stochastic model.
A brief highly accurate calibration at the start and end does not reveal every fluctuation during the gap. A timescale algorithm needs a model of the intervening oscillator behavior and enough evidence to validate it. Short outages and long outages can have very different consequences even when the total fraction of operating time is the same.
For example, two instruments may both operate 90 percent of a month. One loses a few minutes regularly; the other loses three consecutive days. Their nominal uptime is equal, but their contribution to a continuous timescale need not be. Recovery transients and the uncertainty of extrapolation across the long gap can dominate the second case.
This is the practical meaning of asking optical clocks to contribute regularly to international timekeeping. The question extends beyond whether they can produce exceptional laboratory measurements. It asks whether those measurements can be incorporated reliably into a service that must continue when a particular apparatus stops.
Agreement in rate does not set the hands
Two clocks can run at the same rate while displaying different times. Conversely, two displays can agree at one instant and immediately diverge because their rates differ. Optical comparison papers often concern frequency ratios, so reading them as demonstrations of perfectly synchronized displayed time can overstate what was measured.
Consider two hypothetical ideal counters. One begins at zero and another begins at ten seconds, but both advance at exactly the same rate. Their frequency comparison finds no difference, while their time offset remains ten seconds. To align the displayed times, an additional operation must establish the offset. That operation needs a convention for relating distant events and knowledge of signal delays.
Now let the second counter run faster by one part in a trillion. Over one million seconds it gains one microsecond. If it began one microsecond behind, the displays would coincide at the end of that interval despite having different rates throughout. A single agreement of displayed time would therefore conceal the rate error rather than demonstrate its absence.
This distinction helps explain why a link's specification must say what was transferred. A stable frequency relationship can tolerate an unknown constant path delay that would matter greatly to time synchronization. A changing delay can affect both. A report of excellent fractional-frequency stability should not be silently converted into an equally impressive claim about absolute time offset.
For a future realization of the second, rate agreement is central. For a continuously distributed timescale, phase continuity and time offset also matter. Keeping these quantities separate allows a network to demonstrate exactly what it has achieved without diminishing a genuine frequency-comparison result or promising an application the experiment did not test.
The sampling schedule belongs in the result
A network's measurements may overlap only during selected intervals. This becomes important when clocks drift, conditions change or a shared oscillator bridges gaps. Comparing average values from different windows can introduce a difference that would not appear in a simultaneous comparison.
Imagine a hypothetical reference whose fractional frequency increases linearly from zero to ten units over a day. Laboratory A observes only the first half, averaging approximately 2.5 units. Laboratory B observes only the second half, averaging approximately 7.5. Subtracting those means produces five units even if both laboratories measure the reference perfectly. The discrepancy arose from the schedule, not either instrument's calibration.
A correction is possible if the drift is known sufficiently well. But estimating it introduces uncertainty, and the estimate may depend on the same observations being compared. Alternatively, restricting the comparison to common intervals sacrifices data while reducing extrapolation. The best choice depends on the actual noise, uptime and purpose of the campaign.
Selection rules also need physical justification. Removing intervals with an identified loss of lock is different from discarding intervals merely because their frequency looks inconvenient. If a temperature excursion causes data rejection, the experiment should explain whether the remaining data still represent the claimed operating environment. Otherwise, the published uncertainty may apply only to a favorable subset without making that boundary visible.
These questions are unusually concrete in clock metrology because the target is so small. A campaign record should preserve the interval boundaries, operating-state flags and corrections required to reconstruct the comparison. Another analyst then has a chance to distinguish a genuine reference difference from the effects of when, and under which conditions, the measurements were made.
Choosing a definition leaves engineering choices open
The roadmap literature distinguishes definitions based on a single transition from approaches involving a set of transitions. It also treats comparisons, gravitational-potential knowledge and contributions to timescales as connected readiness questions. Its older milestone assessments should be read in their publication context, not substituted for the adopted 2026 resolution. Dimarcq and colleagues, Metrologia roadmap.
A single-transition anchor offers a conceptually direct reference. Other transitions can be related to it through measured ratios. An ensemble-based approach can distribute the role across transitions, but then the specification must make clear how their relationships and associated uncertainties are maintained. These are conceptual tradeoffs; they do not establish which option the international community should select.
An especially important question is what happens when knowledge improves. Suppose researchers discover a previously underestimated shift affecting one family of instruments. A robust system needs a way to update realizations and recommended relationships without disguising the scientific correction or unnecessarily disrupting continuity. The choice of definition and the procedures for practical realization must work together.
No definition can make every physical apparatus equally convenient. Laboratories will continue to differ in expertise, available lasers, environmental control, links and operating requirements. A universal unit does not require identical hardware everywhere. It requires defensible traceability and agreement within stated uncertainty across legitimate realizations.
For users, the benefit should be assessed at the point where improved reference performance reaches an application. An optical standard inside a national laboratory does not instantly confer its uncertainty on a remote device connected through a less accurate distribution chain. Each intermediate calibration and transfer matters. The network is therefore both a scientific comparison system and a route through which an improved unit becomes usable.
The October decision leaves a productive task: demonstrate agreement that is independent enough to be convincing, investigate the comparisons that do not yet agree, and sustain the measurements long enough to support a shared timescale. The next second will earn its authority through that collective evidence, not through a single spectacular display of digits.