Evidence checked October 31, 2026. All worked numerical examples and program scenarios are hypothetical. Proposed evaluation practices are editorial analysis, not current market rules or advice to alter a customer's electricity use.
1. The missing line on the chart
California ISO's notice accompanying its Demand Response Issues and Performance 2025 report says that about 77 percent of real-time demand-response capacity reported performing as scheduled on strained summer days. The wording matters. It describes a reported performance measure within a particular market and period, not a direct observation of the electricity each customer would otherwise have consumed. California ISO report notice.
Demand response turns flexibility in electricity use into a resource. A customer may postpone a process, adjust a building system or use another approved method to reduce demand during a requested interval. The grid can benefit when that reduction occurs at the right time and place. But measuring the benefit involves a difficult comparison: the meter shows actual consumption, while the no-intervention alternative is missing.
A chart can make this look deceptively simple. One line represents the measured load. Another line represents a baseline. The area between them is labelled response or savings. The measured line comes from an instrument. The baseline comes from a method that estimates what would have happened. Both lines may be useful, but they have different evidentiary status.
The problem is not unique to electricity. Any intervention that claims to prevent an outcome needs a credible account of the outcome that would otherwise have occurred. Demand response makes the issue especially tangible because small changes in the assumed baseline can alter payments, reported capacity and confidence in a resource used during system stress.
This article explains how those baselines work, why an apparently conservative method can sometimes understate genuine response, how rebound changes the wider energy picture and what a useful verification program should test. It distinguishes event-level delivery from permanent energy savings, and market settlement from causal program evaluation. These distinctions allow demand response to be judged on the service it actually provides.
The goal is neither to dismiss estimates nor to treat them as measurements. A counterfactual can never be observed for the same customer at the same instant under both treatment and no treatment. It can nevertheless be estimated with disciplined methods, tested against data and reported with appropriate uncertainty. That is the standard a resource should meet when the grid relies on its promised behavior.
2. Define the service before choosing the baseline
Demand response is not one uniform activity. A program may request a short reduction during an emergency, invite bids into an electricity market or encourage customers to shift consumption away from expensive periods. The relevant performance measure depends on the service and the rules under which customers participate.
PJM describes its demand-response program as compensating participating retail customers for reducing load when requested during high prices or threatened reliability, with payments through curtailment service providers. California ISO separately describes several participation models and requirements for market entry, metering and, for ancillary services, telemetry. These descriptions show why a result from one program should not automatically be transferred to another. PJM demand-response overview, California ISO participation overview.
Suppose a hypothetical factory agrees to reduce demand by one megawatt for two hours after a notification. The service includes a magnitude, a duration and a response deadline. A measured reduction beginning after the requested interval may still save energy, but it does not provide the promised timely service. Likewise, a large initial drop followed by a rapid return can fail a sustained commitment even if the event's average looks favorable.
A residential program might instead offer an incentive to reduce consumption voluntarily during selected evenings. Its objectives could include aggregate peak reduction, customer participation and manageable discomfort. Those objectives require different evidence from a tightly dispatched industrial resource. Counting enrolled households alone does not establish the amount of response available on a specific evening.
The unit also matters. Megawatts describe the reduction rate during an interval; megawatt-hours describe the accumulated energy difference. A program that reduces one megawatt for two hours produces a two-megawatt-hour event reduction under the stated baseline. It does not necessarily reduce total daily energy by that amount if consumption moves to another period.
A good evaluation begins with a service statement written before the event. It identifies the requested timing, the relevant meter boundary, the population, allowable operating methods and the comparison being estimated. Without that statement, analysts can select whichever outcome looks strongest after observing the data.
The baseline method should then fit the service. A method designed for predictable industrial schedules may perform poorly for irregular household behavior. A method adequate for annual energy evaluation may lack the time resolution needed for a fast grid service. Choosing the method is therefore part of defining the product, not an accounting detail added after participation has begun.
3. What a baseline is trying to estimate
A baseline is an estimate of the customer's load without the demand-response intervention during the event interval. It may use previous days, weather, operating schedules, a comparison group or a combination of information. The key issue is whether those inputs describe the counterfactual relevant to the event rather than merely producing a smooth-looking line.
The European Commission's Joint Research Centre published a 2022 review of customer-load baseline methods. Its summary emphasizes that suitable methods depend on several factors and compares approaches through literature, standards work and project experience. It does not identify one universally best formula for every customer and program. JRC baseline-method review.
Consider a hypothetical office normally using 500 kilowatts at a particular hour. On an event day it uses 400 kilowatts. A simple comparison suggests a 100-kilowatt reduction. But if the office would have used only 420 kilowatts because half the staff were away, the intervention's contribution is closer to 20 kilowatts. If unusually hot weather would have raised consumption to 600 kilowatts, the same measured 400 kilowatts corresponds to a much larger reduction.
The meter alone cannot resolve those alternatives. The baseline needs relevant information about conditions and a method chosen without knowing which answer produces the largest reward. That is why documentation of weather, schedules and exclusions can be as important as accurate metering.
A baseline can be unbiased on average while still uncertain for an individual event. Across many ordinary days, overestimates and underestimates may cancel. On the hottest or most unusual days, however, the errors may be systematically different. A program used during system stress should examine performance under those conditions rather than relying only on an average validation result.
The target also needs a clear causal boundary. If customers change behavior in anticipation of an event, that behavior may be part of the program's effect. A baseline adjustment that uses the altered pre-event consumption can accidentally absorb or amplify that effect. The analysis must decide which changes belong to treatment and which represent ordinary event-day conditions.
No formula eliminates the need for judgment about that boundary. The practical response is to state the assumptions, validate the method on appropriate data and show how results change under reasonable alternatives. A credible estimate can remain useful while acknowledging that it is not a directly measured quantity.
4. Similar days are similar only in specified ways
Historical-day methods construct a baseline from selected non-event days. Their appeal is transparency: a customer can often see which days entered the calculation and how their loads were combined. Their weakness is that the selected days may differ from the event day in ways that matter to consumption.
A weekday can differ from another weekday because of temperature, occupancy, production, holidays or maintenance. Selecting days by calendar category alone may therefore leave important variation unexplained. Selecting only high-load days can create a different problem if it systematically raises the baseline above ordinary expected use. The choice needs a rationale connected to the program's purpose.
Imagine a hypothetical warehouse with a highly regular weekly schedule. A baseline based on comparable weekdays may work reasonably well until a special inventory operation changes that schedule. An automated method can continue producing a precise number while missing the operational change. Precision in the formula does not guarantee relevance of the selected observations.
For a household, daily variation can be larger relative to total use. Cooking, laundry, travel and electric-vehicle charging may occur irregularly. Aggregating many households can smooth some idiosyncratic variation, but it does not remove common influences such as weather or a holiday. The analyst should distinguish individual unpredictability from systematic group-level bias.
A useful validation exercise applies the proposed method to non-event days as though they were events. Because no demand-response instruction occurred, the estimated reduction should have no persistent directional bias under suitable conditions. Large apparent reductions on these placebo days indicate that the method can manufacture a response from ordinary variation.
The placebo exercise must itself be designed carefully. Testing only mild, predictable days does not establish accuracy during extreme heat or unusual operating conditions. The test set should resemble the events for which the program intends to claim performance, while remaining separate from data used to tune the method.
Historical methods also need rules for missing or corrupted data. Silently filling gaps with favorable values can bias the result. Excluding too many days can leave a small and unrepresentative sample. The report should show how many candidate days were available, which were excluded and how those decisions affect uncertainty. These details make the method reproducible without pretending that history perfectly predicts the event.
5. Event-day adjustments can correct bias or introduce it
An event-day adjustment attempts to align the historical baseline with conditions observed before the response interval. If the day's ordinary load is higher than the historical pattern, an upward adjustment may improve the counterfactual. If it is lower, a downward adjustment may be appropriate. The adjustment is useful only if the pre-event information is a valid indicator of no-intervention demand.
California ISO's March 2025 report on 2024 demand-response performance discusses concerns that capped adjustments could understate response when event-day load departed substantially from historical levels. It describes the use of alternative adjustment factors for eligible resources. This is a concrete example of a measurement rule that can err in the downward direction, not only inflate performance. California ISO 2024 performance report.
Suppose a hypothetical baseline predicts 10 megawatts, while comparable pre-event hours suggest that the day's ordinary consumption is 20 percent higher. An adjusted event baseline of 12 megawatts may be more realistic. If actual event load is nine megawatts, the estimated reduction becomes three megawatts rather than one. The difference is entirely driven by the adjustment's interpretation.
Now consider whether the pre-event load itself was affected by the program. Customers might pre-cool buildings, advance a production task or charge equipment before the event. If the adjustment treats that deliberate increase as evidence of ordinary demand, it can raise the baseline in a way that overstates event reduction. This is an analytical risk, not an allegation about any participant.
The appropriate response is to define the adjustment window and test sensitivity. Excluding the interval immediately before an event may address some anticipation effects, but it is not a universal solution. Customers may receive notice earlier, and ordinary conditions may change during the excluded period. The method should be evaluated against the actual program design.
Caps can limit extreme adjustments, but they involve a trade-off. A tight cap can reduce manipulation opportunities while also suppressing legitimate correction during unusual conditions. A loose cap can improve fit while increasing exposure to contaminated pre-event data. The choice should be justified with evidence about error and incentives rather than presented as an inherently neutral technical setting.
A public performance report should therefore disclose whether its headline result uses an adjusted baseline, which adjustment regime applies and how materially the adjustment changes the estimate. Readers should not have to discover in a footnote that most of the claimed reduction depends on a rule whose behavior under stress has not been examined.
6. Weather models need tests beyond a good fit
Weather-sensitive demand invites statistical modeling. A model can estimate consumption as a function of temperature, calendar variables and other relevant information. This can outperform a simple average when the relationship is stable and the inputs are appropriate. It can also create a false sense of certainty if it is evaluated only on the data used to build it.
A hypothetical model fitted to several months of office consumption may reproduce ordinary days closely. If a demand-response event occurs during temperatures outside that historical range, the model is extrapolating. Its smooth curve does not establish that the building's equipment, occupancy or controls behave according to the same relationship in those conditions.
Validation should therefore separate fitting data from test data. The evaluator should examine errors by temperature range, day type and operating regime, not only a single overall score. A model that performs well most of the year can still be biased during the limited hours when the grid depends on the response resource.
The choice of weather measurement also matters. A distant weather station may not represent the conditions affecting the load, while a local sensor may have quality problems. The analyst should document the source, missing-data treatment and timing alignment. A model cannot correct for weather information that is itself mismatched to the electricity interval.
Operational changes require attention. A building that installs new equipment or changes its hours may no longer follow the relationship learned from older data. Updating the model can improve relevance, but the update rule should be defined before the evaluator sees which version produces a larger response estimate. Otherwise routine maintenance of the model can become an unrecognized source of selection bias.
The JRC review's central point that method choice depends on context applies here. A complex model is not automatically more credible than a simple transparent baseline. Its added variables should improve out-of-sample performance for the intended event conditions and remain interpretable enough to diagnose failures. JRC review.
A useful report presents uncertainty around the predicted counterfactual and explains the main sources of error. That uncertainty can then be carried into the estimated response rather than disappearing when predicted load is subtracted from measured load. The result may be less tidy, but it better reflects what the data can establish.
7. Comparison groups answer a different version of the question
A well-designed comparison group can help estimate what participating customers would have done without an intervention. Random assignment is especially useful because it can make treatment and comparison groups similar in expectation before the event. But the design still needs to specify who was assigned, what treatment meant and which population the result describes.
For a hypothetical program, randomly selected customers receive an invitation while others do not. Some invited customers enroll, some ignore the invitation and some enrolled customers do not respond to a particular event. The effect of receiving an invitation differs from the effect among customers whose participation was induced by it. Reporting these as the same quantity can mislead readers about expected program-wide delivery.
The comparison group also needs protection from spillovers. If a public appeal reaches everyone, customers outside the formal program may reduce consumption too. The measured difference between groups may then describe the incremental effect of enrollment above the public appeal rather than the total effect of all demand reduction that evening. That can be a valid question, but it should be named.
Attrition and missing meter data can weaken the design. If customers with particular response patterns are more likely to disappear from the analysis, the observed sample may no longer represent the assigned population. The evaluator should report exclusions and test whether they differ across groups instead of presenting only the final sample size.
A comparison-group estimate can be strong evidence of aggregate causal impact without being suitable for settling every individual customer's payment. Individual settlement may require a transparent rule applied to that customer's meter, while program evaluation can use experimental evidence across a population. These objectives should not be forced into one method merely for administrative convenience.
The same distinction applies to precision. A large sample can estimate an average effect accurately while leaving substantial uncertainty about a single participant's contribution. A program can therefore be useful in aggregate even though the counterfactual for one household remains noisy. The report should avoid treating individual uncertainty as proof that aggregate evaluation is impossible.
Conversely, a statistically significant average effect does not establish dependable delivery at every location and event. Grid planning may need information about the distribution of responses, common failure modes and the timing of availability. Causal evaluation and operational qualification complement one another; neither replaces the other.
8. Load reduction can move energy rather than eliminate it
Demand response often changes when electricity is used. A process postponed during an event may run afterward. A building may need additional heating or cooling to return to its normal condition. This rebound does not automatically erase the grid value of the event reduction, but it changes the claim that can be made about total energy savings.
A published research demonstration involving more than 300 heat-pump-equipped residential buildings examined both load reduction and rebound damping. The paper reports that predicting peak rebound was more difficult than predicting average reduction, in part because of synchronization when systems resumed operation. Its results concern that demonstration, not every heating system or climate. Residential-heating demand-response study.
Consider a hypothetical event that reduces demand by two megawatts for two hours, producing a four-megawatt-hour event reduction. If customers later consume three additional megawatt-hours relative to their no-event behavior, the net energy reduction across the wider window is one megawatt-hour. The event still shifted load away from the requested period, but calling the full four megawatt-hours permanent energy savings would overstate the result.
If all deferred equipment restarts simultaneously, the rebound can create a new peak. A staggered recovery strategy may reduce that peak, but it needs to respect customer needs and equipment constraints. The relevant metric includes the timing and shape of the return, not merely total post-event energy.
The evaluation window should therefore extend beyond the dispatched interval when the claim concerns energy or broader system effects. It may also need to begin before the event to capture preheating, precooling or other preparation. A narrow window can be appropriate for a market product defined by that interval, but it should not silently become a full-day environmental claim.
Rebound is also not necessarily a one-for-one repayment. The relationship depends on the process and conditions. Some deferred activity may be cancelled, some may become more efficient or less efficient, and some may occur at another site. The analyst should measure or model the relevant pathway rather than applying a universal assumption.
This is why load shifting and energy conservation deserve separate reporting. Both can be valuable, but they solve different problems. A program that uses the same total electricity at less stressful times can support the grid without reducing annual consumption. Honest labeling protects that genuine value instead of making it depend on an unsupported savings claim.
9. Aggregation can smooth noise and hide common failure
Combining many customers can make demand response more predictable because individual variations partially offset one another. It can also conceal important differences between participants. A portfolio-level success may include some strong responses, some nonresponses and some customers whose load increased. The aggregate number is useful, but it should not be the only diagnostic.
A hypothetical group of 1,000 homes may have highly variable individual loads while producing a stable combined pattern. That helps estimate the group's counterfactual. Yet if all homes experience the same heat wave or receive the same confusing notification, the errors or failures may be correlated rather than independent. Increasing the number of participants does not automatically remove a shared bias.
Location matters as well. A reduction in one part of the network may not relieve a constraint elsewhere. An aggregate national total can demonstrate broad participation without establishing value at a specific constrained substation or market zone. The program's geographic grouping should match the service it is offering.
Customer composition can change over time. New participants, departing customers and changes in equipment alter the portfolio. A baseline trained on an earlier group may need adjustment, but the adjustment must distinguish genuine population changes from selective exclusion of weak performers after an event. A transparent enrollment history helps make that distinction.
California ISO's participation overview recognizes aggregation and specifies market-entry and metering requirements. Those requirements establish an organized route for participation; they do not by themselves prove a particular portfolio's causal impact. Operational eligibility and measured delivery are separate pieces of evidence. California ISO demand-response overview.
A useful evaluation reports both aggregate performance and the distribution across meaningful customer categories, subject to privacy protection. It can identify whether a few large sites dominate the result or whether response is broadly distributed. That information matters for the risk that one equipment failure or one customer withdrawal substantially reduces future capability.
The analysis should also distinguish enrolled capacity from available and delivered capacity. Enrollment is a contractual or administrative status. Availability concerns the event conditions. Delivery concerns actual performance against the relevant measure. Treating all three as the same number can make a growing program appear more dependable than its operational record supports.
10. Metering and communications are different evidence streams
A successful dispatch begins with an instruction that reaches the right participant in time. It continues with an action and ends with a measurement. A failure at any point can reduce delivery, but the records needed to diagnose those failures differ. A meter trace alone cannot establish whether the customer received the message.
The evaluator should preserve dispatch timestamps, delivery acknowledgments where available, control commands and metered load. These records need synchronized clocks and a clear definition of the event window. A response that appears late may reflect a communications delay, a slow physical process or a timestamp mismatch. The remedy depends on which explanation is supported.
Meter resolution must fit the product. Hourly data may be sufficient for some energy comparisons but cannot establish the timing of a response expected within a much shorter interval. Conversely, high-frequency telemetry can show a rapid drop while revenue-quality settlement data establish the energy quantity over the event. The streams serve related but distinct purposes.
California ISO states that market participation requires certified revenue-quality meters and that ancillary-service bidding has additional certification and direct-telemetry requirements. This is a useful example of matching evidence requirements to the service, rather than assuming one data stream suits every product. California ISO participation requirements.
Missing data need a predefined treatment. A program should not count an unobserved interval as full performance merely because a control command was sent. Nor should every missing interval automatically become zero physical response without explaining that this is a settlement convention rather than an observation. The distinction between an accounting rule and a measured fact should remain visible.
Data corrections should preserve version history. If a meter reading is later revised, the reported performance and payment may change. A public summary can identify the revision without exposing individual customer information. Silent replacement of results makes it difficult to understand whether an apparent trend reflects behavior or improved data.
The broader lesson is that demand response is a coordinated service, not just an estimated difference between two curves. Reliable communications, appropriate controls and trustworthy measurement all contribute. A well-designed evaluation separates those functions so that a weak link can be improved rather than hidden inside a single portfolio percentage.
11. Incentives should reward the intended change
A baseline-linked payment creates incentives around both the event load and the information used to predict it. Participants may respond honestly to the program while the rule still rewards behavior that does not match the intended system benefit. Evaluating those incentives is a design task, not an accusation that customers are acting improperly.
For example, a hypothetical rule might reward reductions relative to recent high-use days. If customers can influence which days enter the baseline, they may have an incentive to increase consumption before future events. Even without deliberate manipulation, a method that selects unusually high days can produce rewards for ordinary regression toward typical load.
NIST's 2013 paper on demand-response measurement and verification frames the problem in terms of methods that are fair, simple and accurate. Those goals can conflict: a very simple method may be easy to understand but less accurate for irregular loads, while a complex method may be harder for customers to verify. The paper is a foundational discussion, not a current rulebook. NIST measurement-and-verification paper.
A useful incentive review asks whether the payment rule encourages the desired timing and magnitude of response, whether it rewards changes outside that objective and whether participants can understand the calculation. It also asks how errors are handled. A method that systematically underestimates response can discourage useful participation just as an inflated method can overpay for weak performance.
Testing on non-event days can reveal false apparent response, but it should be combined with analysis of participant behavior around notifications and baseline windows. If pre-event consumption changes after enrollment, the evaluator needs to determine whether that is ordinary adaptation, legitimate preparation or a source of measurement bias. The answer may depend on the program's stated objective.
The program should avoid relying solely on punitive detection after the fact. Clear rules, stable methods and transparent feedback can reduce confusion and improve behavior. Customers should know whether preconditioning, backup generation, storage dispatch or other actions are permitted and how they affect measured performance.
Public reporting should describe material incentive risks and the evidence used to manage them without publishing a recipe for exploiting a specific system. The aim is to show that the program pays for a defensible service. A large payment total is neither proof of success nor proof of waste until the underlying response and rules are understood.
12. Separate settlement, evaluation and planning
Three users may ask for three different numbers. A settlement team needs an amount calculated under an agreed payment rule. An evaluator needs an estimate of the program's causal effect. A grid planner needs a dependable contribution under future stress conditions. These numbers may be related, but they need not be identical.
A settlement baseline can be administratively practical and consistently applied while retaining statistical bias. That does not necessarily invalidate every payment; it means the settlement result should not automatically be presented as a causal estimate. Conversely, a rigorous experimental evaluation may identify an average effect that cannot be allocated precisely to individual customers for billing.
Planning adds another layer. A program's average response across mild events may not establish performance during extreme conditions. Customer availability, comfort limits, industrial schedules and notification lead time can change. A planner needs evidence about the relevant stress periods and uncertainty, not simply the largest response ever observed.
Suppose a hypothetical portfolio is enrolled for 100 megawatts, offers 80 megawatts in an event and is credited with 60 megawatts under its settlement method. The figures describe enrollment, offered capability and measured-or-estimated delivery under a rule. A causal evaluation might produce a different estimate with an uncertainty range. None should be substituted for another without explaining the transition.
A balanced scorecard can preserve these distinctions. It can show enrollment, event availability, dispatch instructions, settlement performance, independently evaluated impact and post-event behavior. It can also report data completeness and the number of events. A single percentage without these supporting quantities leaves readers unable to tell what succeeded or failed.
The California ISO notice's careful phrase about capacity reported as performing as scheduled is therefore worth preserving. Reporting language should retain the method and population behind the result rather than simplifying it into a universal claim that a fixed percentage of customers reduced use. California ISO 2025 report notice.
Separating the purposes also helps improve the program. A settlement issue calls for one remedy, a communications failure for another and a planning overestimate for another. Treating every discrepancy as generic underperformance prevents the institution from learning which part of the system needs attention.
13. Carry uncertainty through the subtraction
Suppose a hypothetical event has measured average demand of eight megawatts and a baseline estimate of ten megawatts. The point estimate of response is two megawatts. If plausible baseline uncertainty spans nine to eleven megawatts, the corresponding response spans one to three megawatts before considering meter uncertainty or other errors. Reporting precisely two megawatts without the uncertainty can imply more knowledge than the method supplies.
The calculation becomes more complicated across intervals because errors may be correlated. If the model systematically overestimates load throughout a hot afternoon, adding several interval estimates does not make that common bias disappear. A confidence calculation that assumes independent errors can become too narrow. The evaluator should examine the residual structure and choose an uncertainty method suited to the data and design.
Portfolio aggregation creates the same issue across customers. Independent household variation may average out, while a shared weather-model error can affect everyone in the same direction. The number of meters is therefore not enough to establish precision. The dependence among their errors matters, particularly when the program is dispatched during unusual system-wide conditions.
Uncertainty also should not be confused with a contractual tolerance. A market may define a payment rule or performance threshold for administrative purposes. That rule determines settlement under the agreement; it does not make statistical uncertainty vanish. A report can present the settled quantity and the evaluator's uncertainty separately without suggesting that either is the only legitimate number.
Finally, small estimated effects need proportionate language. If a plausible uncertainty range includes little or no reduction, the evidence may be inconclusive for that event even though the program performs well across a larger sample. Pooling events can improve precision, but the report should retain the distinction between an aggregate finding and a claim about every individual dispatch.
14. Build an evaluation that can be challenged
A credible demand-response evaluation should make its central assumptions inspectable. It should define the population, event windows, baseline method, adjustments, exclusions and treatment of missing data before presenting the headline result. It should identify which decisions were made before observing outcomes and which were exploratory changes afterward.
The analyst should test the baseline on suitable non-event data, examine error under conditions resembling actual events and compare reasonable alternative methods. A result that changes dramatically across plausible specifications deserves a wider uncertainty statement. A result that remains stable across well-justified checks supports greater confidence, though it still applies to the studied population and period.
The evaluation should follow the load before and after the event when broader energy claims are made. It should separate reduced consumption from shifted consumption and identify any measured rebound. If the program's objective is peak relief, the timing of that rebound may matter more than whether total daily energy falls.
Customer experience also belongs in the evidence. A technically successful response that cannot be sustained without unacceptable disruption may not be a reliable long-term resource. Participation, opt-outs and complaints should be examined with appropriate privacy safeguards. These records do not replace electrical measurement, but they help explain whether the observed response is repeatable.
A useful public summary can remain concise while preserving the essentials: the service requested, the measured load, the estimated counterfactual, the uncertainty and the wider-window effect. It should state whether the result is a settlement figure, a causal estimate or a planning assumption. Readers can then understand what the program has demonstrated without becoming specialists in every baseline formula.
Demand response has a genuine advantage: it can use flexibility already present in buildings, equipment and industrial processes. That advantage deserves careful measurement rather than exaggerated certainty. A program can provide valuable peak relief even when it does not permanently reduce energy use, and an honest uncertainty range does not erase a well-established aggregate effect.
The decisive question remains what would have happened without the intervention. No meter can answer it alone. Historical data, models, comparison groups and operational records each contribute evidence. The strongest program combines them in a way that can be tested, corrected and explained, so that the grid receives a dependable service and customers receive credit for the change they actually make.