Proving a safety control works: acceptable evidence and false confirmations

A safety analyst in a control room reviewing flight data to verify the effectiveness of safety controls.

A control is put in place, and a year later the safety review records that there have been no further occurrences of that type. The item is closed. The trouble is that nothing in that sentence is evidence about the control. It is evidence about a count, and a count of rare events over a short period is close to uninformative whichever way it falls. Proving that a mitigation reduced a risk requires a different kind of proof, and most organisations already hold the raw material for it without using it.

What « effective » has to mean before it can be measured

ICAO Annex 19 and the Safety Management Manual that accompanies it expect risk controls to be monitored. In the European rules the obligation to verify effectiveness is explicit: it reaches operators through the management-system requirements sitting alongside Part-CAT, and maintenance organisations through Part-145. None of these texts says how, and that is not an oversight — the method depends entirely on what the control does.

Before anything can be measured, the claim has to be stated in a falsifiable form. “This measure will improve safety” cannot be tested. What can be tested is a claim of the shape: when this specific hazardous condition arises, this control will interrupt the sequence before this specific consequence. That form does three things at once. It names the condition, so you know what to count. It names the mechanism, so you know where to look for a trace. And it names the consequence, so you know what a failure would look like.

Most controls that resist measurement resist it because the claim was never written that way. A control described as “increased awareness” has no demand event, no observable engagement and no failure mode, so there is nothing to measure and nothing to disprove. That is a design problem, not a measurement problem, and it is worth catching at the point the control is proposed.

To prove a safety control works, you need to define exactly what it’s supposed to stop, how it does it, and what failure looks like. Without these details, you can’t measure its effectiveness.

Why counting outcomes cannot answer the question

The events that safety controls exist to prevent are, by construction, rare. That rarity is what makes outcome counting useless over the horizons on which organisations actually review their controls.

Consider a hazard whose historical rate was such that a typical year across the fleet produced zero or one occurrence. After the control is implemented, a year passes with no occurrence. That result is the single most likely outcome under both competing hypotheses — that the control eliminated the risk entirely, and that it did nothing at all. It does not distinguish between them, because there was never enough exposure in the observation window for the two hypotheses to produce visibly different counts. Waiting longer helps, but the horizon required is usually years beyond the point at which the organisation needs an answer, and by then the operation has changed enough that the comparison is no longer clean.

The consequence is not that outcome data is worthless. It is that outcome data is a slow, low-resolution confirmation that belongs at the end of the chain of evidence, not at the start. An organisation that treats it as the primary proof will close items that should stay open and, more damagingly, will develop confidence in barriers that have never actually been tested.

The trap. “No occurrences since implementation” is not weak evidence that the control works. It is not evidence about the control at all, in either direction. Recording it as a positive finding converts an absence of information into apparent assurance, and that conversion is how organisations accumulate barriers they believe in and have never verified.

An airline implements a new cross-check procedure to prevent runway excursions. Over the next 12 months, no such incidents occur. However, the historical rate was already 0.5 incidents per year per fleet. Without counting how often the cross-check was actually used (demand) and whether it caught errors (performance), the airline cannot prove the procedure worked.

The evidence ladder

There are six distinct things one can establish about a control, and they are not interchangeable. Ranking them makes visible which rung an organisation is actually standing on when it declares a control effective.

Six levels of evidence about a risk control
Level Question answered Typical source What it does not prove
1. Design Does the control exist, specified and approved? Procedure text, modification record, training syllabus That it is present in the operation
2. Presence Is it in place and serviceable where and when it is needed? Unannounced sampling, audit, technical records That anyone uses it
3. Use Is it actually used, and not bypassed? Line observation, flight data, task sign-off patterns That the hazard ever arose
4. Demand Has the hazardous condition occurred with the control in place? Occurrence reports, flight data events, precursor counts That the control did anything about it
5. Performance on demand When solicited, did it hold? Event-by-event review of the demand cases That the overall risk has fallen
6. Outcome Have the consequences become less frequent or less severe? Accident and serious incident record Nothing — but it takes years to say anything

Levels one and two are what most audits produce. They are necessary and they are frequently mistaken for the answer. A finding that reads “procedure published, all crews trained, no findings raised” establishes level one and part of level two. It says nothing about levels three through five, which are where the question actually lives.

Demand and performance on demand

The measurement that carries real weight is a ratio: of the occasions on which the hazardous condition arose, in what proportion did the control interrupt the sequence as intended. Both halves have to be counted, and the numerator is worthless without the denominator.

Counting demands is the part organisations skip, and it is usually the part that is available. A demand is a precursor: the condition the control exists to catch, whether or not it went any further. A cross-check that catches a wrong entry is a demand satisfied. A guard that is struck is a demand satisfied. An alerting system that fires correctly is a demand satisfied. These leave traces in data the organisation already collects — flight data monitoring events, technical log entries, ramp reports, sign-off corrections, the free text of reports that were filed for another reason.

Once demands are counted, three separate quantities become visible, and they answer different questions. A high demand rate with high success says the control is working hard and holding — it also says the underlying hazard has not gone away, which matters for whether the control can ever be removed. A high demand rate with intermittent success identifies exactly where to invest. And a demand rate of zero says something important and easy to misread: either the hazard is not occurring, or the control is not being engaged, or nobody is recording the engagements. Those three have entirely different responses, and only fieldwork distinguishes them.

The right reflex. Before implementing a control, decide what a satisfied demand will look like in the data, and confirm that it will leave a trace somebody can retrieve. A control that operates invisibly can be complied with perfectly and can never be shown to work. Designing for measurability at that stage costs one meeting; retrofitting it afterwards usually cannot be done at all.

How a measure confirms itself falsely

Most false confirmations are not dishonest. They are ordinary statistical and organisational artefacts that look exactly like success.

A construction worker in a hard hat and high-visibility vest checks a clipboard at a nighttime worksite.
Safety checks are critical to preventing workplace accidents.
False confirmations, and the tell that exposes each one
Pattern How it appears The tell
Zero-event fallacy “No occurrences since the change” The pre-change period also contained long runs of zero
Regression to the mean A cluster triggered the action; the rate then fell The action followed an unusual peak rather than a trend
Reporting-rate shift Reports fall, or rise, right after implementation The change included briefings, campaigns or a new form
Definition drift The occurrence category count drops The taxonomy or threshold was revised in the same period
Exposure change Absolute counts fall Sectors, seasons, fleet mix or network changed too
Displacement The event disappears at one point A related category grows somewhere downstream
Risk compensation The barrier works; the margin does not improve Behaviour has adapted to the new protection
Observation effect Line observations show excellent compliance Compliance is measured only when someone is watching

Three of these deserve a note. Regression to the mean is the most common and the hardest to argue against, because controls are almost always introduced after a cluster, and clusters are followed by quieter periods whether or not anything was done. The defence is to compare against a longer baseline than the cluster and to look at the trend before the peak, not at the peak itself.

Reporting-rate shift cuts both ways and is frequently misread in the favourable direction. A control introduced with communication and training raises awareness, which raises reporting, which makes the indicator look worse while the risk is falling. The reverse also occurs: when a change is perceived as having settled the matter, reports of the same condition stop being filed because people assume they are no longer wanted. Neither movement in the count is about the hazard.

Displacement is the one that produces genuine harm rather than merely false comfort. A control that makes a task harder in one place often moves the work rather than removing the hazard: a step that becomes difficult gets done earlier, elsewhere, or by someone else. The occurrence category the organisation was watching improves, and a different one degrades, usually reported by a different department under a different heading. Looking only at the category the control targeted guarantees missing this.

Write the falsification criterion before the change

The single most effective discipline available here costs nothing and is rarely applied: before implementing the control, write down what result would count as failure.

The record should carry the mechanism claim in the falsifiable form given above; the demand indicator and where it will be read from; the performance-on-demand indicator; the specific result that would cause the organisation to conclude the control is not working; a review date; and the conditions under which the control would be removed. Setting out in advance what would count as failure removes the interpretive latitude that otherwise appears at the review, when the person assessing the control is often the person who proposed it and the evidence is ambiguous enough to support either conclusion.

The removal condition matters more than it appears. Controls accumulate. Each addition consumes attention, adds a step, and slightly degrades the controls already present by competing with them for the same finite operational capacity. An organisation that never removes a control eventually has a procedure nobody can execute in full, at which point the informal triage of which steps actually get done is being performed by the front line rather than by the safety process.

Indicators that measure the paper

Some indicators are popular because they are easy to collect, and are unfortunately measures of the administration rather than of the risk.

Weak indicators and stronger substitutes
Common indicator What it actually measures Stronger substitute
Percentage of staff trained Course attendance Observed correct execution on an unannounced sample
Procedure published and revised Document control Proportion of sampled tasks where the procedure is present and current at the point of use
Number of audits with no findings Audit scope and depth Time to restore the control after it is found unserviceable
Occurrences of the target type Reporting behaviour and exposure Demands on the control, and the proportion satisfied
Safety action items closed on time Project management Items reopened because the control did not hold

None of the left-hand column is worthless. Training completion and document control are genuine obligations and belong in the management system. The error is filing them under effectiveness, where they answer a question nobody asked.

What can be checked in the field this week

Five checks require no new data collection and can be carried out on any control that is physically or procedurally present.

  • Find it unannounced. Go to a random sample of the places the control should be, without notice, and establish whether it is there and serviceable. For an engineered barrier this is literal. For a procedural one, it means checking whether the document, the tool or the second person the procedure assumes is genuinely available at the moment the task is done.
  • Look for the workaround. A bypassed control leaves physical traces: a sensor obstructed, a guard removed and not refitted, an interlock defeated, a laminated summary card that contradicts the official procedure, a step consistently signed together with the following one. These traces are informative precisely because nobody produced them for the auditor.
  • Ask to be shown, not told. The question “what does the procedure say” tests memory. The question “show me how you did this one” tests the control. The gap between the two answers is the real measurement, and it only appears if the exercise carries no jeopardy — which means the conditions under which the observation is conducted are part of its validity.
  • Measure the restoration time. When the control is found unserviceable or unavailable, how long does it stay that way? A barrier that is deferred for weeks at a time is available less often than the organisation believes, and the deferral records give a directly usable availability figure without any new instrumentation.
  • Ask when it last got in the way. People who work with a control daily know exactly when it has been inconvenient, and the answer distinguishes a control that is engaging with the operation from one that is inert. A control that has never once been an obstacle has quite possibly never once been solicited.

When the honest answer is that you cannot yet know

Some controls cannot be shown to work within the review period, and saying so is a better outcome than manufacturing a confirmation. A record that says “the control is present, correctly used in the sampled cases, has not yet been solicited, and cannot be assessed for effectiveness before the next review” is accurate, actionable and safe. A record that says “effective, no occurrences” is none of those things, and it is worse than silence, because the next person reads it as verification and stops looking.

Occurrence reporting under Regulation (EU) 376/2014 supplies part of the raw material for this work, but only the part that crosses the reporting threshold. The demands that a control satisfies quietly are rarely reportable, which is why they have to be captured deliberately, and why an organisation that relies solely on the mandatory reporting stream will be permanently short of the evidence it most needs. Practitioners comparing how their organisations handle this find the discussion worth having on the FL330 professional forums.

Commentaires

Leave a Reply

Découvrez l’écosystème FL330

FL330, c’est toute l’aviation professionnelle — explorez nos autres espaces.