Begin with a testable claim
“AI makes lighting better” is too broad to evaluate. A useful claim identifies the intended change, the affected displays and the conditions that must remain acceptable. For example, a pilot might test whether occupancy-based scene control reduces unnecessary exposure during quiet periods without increasing missed activations or reducing approved viewing conditions.
Write the acceptance criteria with conservation, access and operations staff before examining results. Separate essential requirements from desirable improvements. Respecting a lender's lighting conditions is an essential constraint; a reduction in energy use is not a substitute for it. Identify who has authority to accept the evidence and who can halt the trial.
Define the system boundary. State whether the evaluation includes daylight control, electric lighting, sensors, network services and associated computing. A narrow measurement of lamp electricity may be useful, but it should not be presented as the total operational footprint of a larger installation.
Build a baseline that can be compared
Record the existing arrangement before enabling new control. Document object positions, scenes, operating hours, daylight conditions, sensor locations and relevant maintenance issues. The baseline should represent actual practice, including out-of-hours activity, rather than a theoretical schedule that nobody follows.
Comparison periods can differ for reasons unrelated to the controller. School holidays, special events, weather and exhibition popularity can all change occupancy and exposure. Keep contextual records and use comparable periods where possible. A simple before-and-after total is not enough to attribute every difference to the new system.
When feasible, compare zones or alternate approved modes under a planned design, taking care not to introduce unequal conservation risks. Report practical limitations openly. A short pilot in one season may establish basic functionality without demonstrating annual performance.
Measure the object-relevant outcome
Use measurements that represent the relevant object planes and retain the assumptions connecting fixed sensors to them. Check selected points independently with an appropriate calibrated instrument. Document instrument uncertainty, orientation and spectral limitations so that differences between devices can be interpreted rather than hidden.
Accumulate exposure from time-stamped readings and mark gaps explicitly. A missing hour cannot safely be treated as an hour in darkness. If an estimation method is used, explain it and show how much of the reported total depends on estimates. Preserve peaks and threshold events alongside cumulative values.
Do not claim proven prevention of fading from a short trial with no visible change. Many changes are too small or too slow to detect in that period, and condition documentation has its own limits. Reduced measured exposure is a useful intermediate outcome, not a direct measurement of all future deterioration avoided.
Evaluate access and operational burden
Observe the complete visitor journey: arrival, label reading, sustained viewing and departure. Check seated and slow-moving visitors as well as typical walking patterns. A system can report high activation accuracy while still inconveniencing people whose behaviour falls outside its test assumptions.
Gather structured staff observations about abrupt changes, false activations, delayed scenes and difficulty finding controls. Record the effort required to investigate alarms, replace batteries, recalibrate sensors and maintain software. These tasks affect long-term reliability and should be included in cost comparisons.
Energy consumption can be measured separately using a clearly stated boundary and interval. If exposure decreases but equipment energy increases, report both rather than forcing them into one success score. Conservation, access, maintainability and energy performance answer different questions.
Test faults deliberately
Normal operation is only one part of acceptance. In a controlled test environment, introduce stale data, a disconnected sensor, network loss and a restart. Confirm that the approved fallback actually occurs, that alerts reach the right people and that recovery does not bypass current settings or erase history.
Check interactions between manual overrides and automation. A timed exception should expire as documented; a deliberate safe manual state should not be cancelled unexpectedly by a reconnecting service. Test configuration restoration with staff rather than relying on an undocumented supplier action.
For predictive functions, evaluate conditions outside the training sample, such as unusual occupancy patterns or a changed daylight distribution. If the model cannot recognize that its assumptions no longer hold, the supervisory controls must still constrain its actions. Document which behaviours remain untested.
Make a proportionate scale decision
A final evaluation should include the configuration version, measurement methods, complete comparison periods, missing-data treatment, observed failures and staff feedback. Report absolute outcomes as well as relative changes. A large percentage improvement from a very small baseline can otherwise obscure its practical significance.
Choose among expansion, revision, continued observation or return to the baseline. Expansion should be conditional on the new locations having comparable requirements; a successful gallery pilot does not automatically validate a different collection, sensor geometry or operating pattern. Define what must be recommissioned at each new site.
Finally, set review triggers. A new exhibition, replacement luminaire, software update or changed loan condition can invalidate earlier assumptions. Evaluation is not a one-time certificate of intelligence. It is a repeatable way to keep the relationship between conservation needs, measured evidence and operational decisions visible.
Use standards with their scope intact
CIE 157:2004 provides a relevant starting point for optical-radiation damage considerations. EN 16893:2018 addresses buildings and rooms intended for the storage or use of heritage collections, while EN 16163:2014 addresses the selection of appropriate lighting for indoor exhibitions. Refer to the actual documents and applicable professional advice rather than treating this page as a compliance assessment. Institutional resources from the Canadian Conservation Institute, the National Gallery and the Victoria and Albert Museum can also support a broader understanding of collection care.