A trainee steps off a line flight sector convinced he flew badly: he counted four errors. The instructor logged only one, the only one that lasted more than a few seconds. Both counts are correct, and the gap between them sums up what threat and error management has established since the earliest line operations safety audits: error is an ordinary event, and it is not the number of errors that separates a routine flight from an undesired aircraft state. It is the share of those errors that are seen, understood, and closed out before they produce anything at all.
Errors happen all the time in flying, but what matters is catching them early. If you spot and fix an error quickly, it’s no big deal. But if you miss it, the problem gets worse over time. The key is having systems to detect errors before they cause real trouble.
A Boeing 737 crew on a 3,000-meter runway at 30°C sets the wrong flap setting for takeoff. If they notice immediately (within 3 seconds) and correct it, there’s no consequence. But if they only realize after rotation (30 seconds later), they’ve already committed to a longer takeoff roll, potentially exceeding runway limits. The same error costs nothing or everything depending on detection time.
The ordinary error, and the one that costs
Line operations safety audits, conducted on the flight deck by trained observers with no disciplinary purpose, produced a result that is counter-intuitive for anyone fresh out of training: among ordinary crews, error occurs on a very large share of the sectors observed, to the point that its absence is the exception. A reselected value, a skipped checklist item, an approximate readback, a computer page validated too quickly. This finding did not harden the hunt for errors: it changed the variable being observed.
The framework that emerged from this work distinguishes three things that are readily confused. A threat comes from outside: weather, a failure, an ATC constraint, demanding terrain. An error is a gap between what is done and what was intended. An undesired state is a position, a configuration, or an energy level that reduces margins, and it is the only one of the three with a direct cost. Between error and undesired state lies an interval, and that interval can be worked on.
The pedagogical consequence is immediate. A training scheme that scores the number of errors pushes people to hide them, even from themselves: the trainee learns not to name what he saw go by. A scheme that scores their management pushes people to state them, because stating an error is the first act of recovery. It is this shift, not an added dose of rigor, that sets a modern progress record apart.
The links between the act and its consequence
Between error and undesired state, each link has its own way of breaking. Commit. Detect, meaning notice that something does not match. Diagnose, meaning know what. Correct. Then observe the outcome: no consequence, undesired state, or an additional error.
The link most often missed is not the one people assume: diagnosis fails as often as detection, and more quietly. A crew that sees the aircraft depart in altitude and reselects the intended value has detected and corrected without having diagnosed; if the deviation came from a mode other than the one they believed was engaged, the same drift will resume a few seconds later, and the second observation will cost more because attention will have moved elsewhere. A correction that treats the symptom leaves the cause active.
The correction itself can be the next error. A poorly initiated go-around, a reconfiguration out of sequence, reprogramming during descent: the response to an error is an action like any other, executed under load and under time pressure, therefore at least as exposed. This is why chains that end badly almost always involve several linked errors. Detection nonetheless keeps a place apart: it is the only link whose failure removes all the following ones, since an undetected error has no chance of being diagnosed.
Delay sets the price
For an identical error, the cost of recovery does not depend on the initial severity but on the elapsed time. A wrong selection corrected within the second costs only one action. The same selection discovered two minutes later has produced a trajectory, hence an energy deviation, hence a negotiation with ATC and a reconfiguration. Discovered at the next phase, it has contaminated a performance calculation or an approach briefing.
Time does not just increase the physical deviation: it consolidates the false mental model. The longer a situation lasts, the more the crew builds a coherent interpretation around it that absorbs contradictory cues instead of registering them. A parameter that does not move as expected reads as an anomaly in the first few seconds, and as confirmation a few minutes later. Late detection therefore costs twice: a larger deviation, with a diagnosis that has become harder.
The trap. A spectacular recovery is praised even though it first signals late detection: people commend the save, never the chain that made it necessary. The twin trap is the silent correction, the reflex of quietly putting the right parameter back to avoid exposing one’s mistake. The deviation disappears from the aircraft but stays in the other pilot’s head, who will now be watching for a situation that no longer exists.
Who detects what, and in how much time
Detectors are not interchangeable and do not act at the same instant. Yourself first: it is the fastest channel, and it depends only on the verification loop you impose on yourself after every action. The other seat next, whose value lies in its independence more than in its experience. The machine, which only triggers on a threshold crossing, that is, once the undesired state already exists: these are safety nets, not watchers. ATC, which sees a trajectory but neither a configuration nor an intention. Finally the ground, which no longer detects anything useful for the flight in progress.

| Family | How it shows itself | Usual detector | Typical delay |
|---|---|---|---|
| Handling: trajectory, speed, configuration | The aircraft moves; the deviation becomes visible on the instruments | Yourself, then the deviation call from the monitoring pilot | A few seconds |
| Procedure: item omitted, page not validated | Nothing happens at the time; the omission is silent | Cross-checking and the checklist | From a minute to the next phase |
| Communication: incorrect readback, misunderstood message | The deviation appears only once the instruction is executed | The readback loop, the other seat, ATC | Immediate if the loop is done, very late otherwise |
| Mode: one mode engaged for another, incorrect entry | The display changes, the trajectory not yet | The mode annunciator call, then the trajectory | Until the phase change |
This hierarchy explains a rule that looks bureaucratic: standard callouts. A callout does not inform the other pilot, who has the same instruments; it forces the reading. Verbalizing the mode annunciator after every change turns available information into retrieved information, and it is about the only known mechanism for detecting a mode error before the trajectory reveals it.
Two pilots do not make two barriers
Cross-checking has value only if the two detectors are independent. Yet everything about a crew works against that independence: the shared briefing installs a shared expectation, fatigue is shared, workload peaks at the same moments, and the pilot monitoring heard the same clearance as the one who read it back. Expectation bias makes both sides of the flight deck hear the expected level in place of the delivered level, with the same conviction.
A simple design rule follows: cross-checking must be against a source, not a person. Comparing your selection to what the other pilot just called out verifies nothing; comparing it to the written clearance, the flight plan, the chart, restores independence. This is also why workload is a safety topic: monitoring is the first function abandoned when workload rises, at the exact moment error probability peaks.
The right reflex. State the doubt before you have resolved it. Saying at the third second that the mode obtained is not the one expected is worth more than a full diagnosis at the thirtieth: the callout puts the second detector to work while the first is still searching. Also name out loud the correction of your own error: the action puts the aircraft back in order, the statement puts the crew back in order.
Keeping your own detection log
The useful count after a sector does not fit in one number but in three. How many deviations were seen by yourself, in the action that followed? How many were flagged by the other seat, by an alert, or by ATC? How many were discovered only after the fact, at debrief, or never? The first number measures the personal verification loop, the second measures what you owe to external barriers, the third is the only one that should worry you.
Two distinctions keep this log honest. Correcting is not detecting: you can bring a parameter back by reflex without ever knowing why it moved, and this permanent compensation masks an active chain. And delay matters more than the fact itself: the same error seen in three seconds or in three minutes does not say the same thing about monitoring. Whoever keeps these three numbers sees their progress before the instructor does, because they are measuring the one variable that moves quickly with training.
What the operator sees, and what escapes it
At the scale of the airline, the same links reappear, shifted in time. Flight data analysis detects parameter exceedances, hence undesired states: it sees the outcome, never the error that produced it, still less the one that was closed out in time. The event report carries the error itself, with its context and intent, but only the one the crew detected and agreed to write down. Hence the central place of reporter protection: without it, the channel that carries errors is the first to dry up.
| System | What it captures | What escapes it | Timescale |
|---|---|---|---|
| Flight data analysis | The parameter exceedance, hence the undesired state once formed | The error closed out in time, and the reason behind the one that succeeded | Days |
| Event report | The error as the crew understood it, with its intent | What the crew did not detect, or chose not to write down | Days to weeks |
| Line operations safety audit | The error committed, detected or not, and its management | What does not occur in front of the observer | Campaign |
| Crew debrief | The fresh error, including the one that left no trace | Everything nobody states | Minutes |
A structural blind spot remains: the crew that detects well leaves no trace. Its flight produces no exceedance, no event, nothing to analyze. An airline’s indicators therefore measure its recovery failures and never its successes, which makes the debrief and the line audit irreplaceable: they are the only tools able to see an error that produced nothing.
What to go check, and what remains open
In your own documentation, four points deserve more than a skim. The standard callouts and deviation tolerances, which set the triggering threshold for detection: an unwritten tolerance produces none. The stabilization criteria and the point at which they are assessed, the only detection with a fixed deadline, independent of the vigilance of the moment. The cross-checking policy: what must be compared, against which source, and by whom. Finally task-sharing during a failure or a diversion, because that is where monitoring disappears first.
The gray area lies in measurement. No common system measures a detection rate; only the number of errors that got through is known. A crew can therefore see its monitoring competence improve without any indicator noticing. For anyone preparing a checkride or an interview, this changes how to present oneself: the competency frameworks used in training and checking make threat and error management a standalone assessment item, one that is demonstrated by recounting an error detected early, with the exact signal that revealed it, rather than by claiming never to have made one. This kind of account is honed in training and tested on the forum.
Applicable regulations, the operations manual, and the operator’s approved procedures prevail in all circumstances over this article, whose purpose is to explain a mechanism, not to serve as an instruction.
Questions fréquentes
Should an error that was detected and corrected with no consequence be reported?
Yes, whenever it reveals a reproducible mechanism: an ambiguous clearance, a computer page that invites the mistake, a checklist sequence poorly placed in the phase. The report is the only channel that carries the error itself, where flight data analysis will never see more than an exceedance. An error closed out in time is even the most instructive case, since it contains both the flaw and the barrier that worked. The reporter protection regime exists precisely so that this account costs its author nothing.
How do you know whether you truly detected something, or merely corrected it by reflex?
The test is simple: can you name the cause of the deviation at the moment you correct it? If the answer is no, you compensated for a symptom, and nothing guarantees it will not come back. A permanent compensation, a pilot who constantly holds a parameter at its value, signals that a chain remains active upstream. The useful reflex is to call out the deviation before correcting it, because stating it forces the diagnosis.
Isn't calling out the mode annunciator redundant since both pilots can see it?
Seeing is not reading. The information is displayed continuously, so it is easily absorbed by expectation: one reads the mode one requested rather than the one engaged, especially after an action believed to have succeeded. Verbalizing turns available data into retrieved data, and it timestamps that retrieval, letting the other pilot know how recent the last check was. It is about the only mechanism that detects a mode error before the trajectory takes care of it.
Does a two-pilot crew really form two independent barriers?
Rarely, and this is the most misunderstood point about cross-checking. Both pilots received the same briefing, heard the same clearance, and endured the same workload and the same fatigue: their reading errors are correlated, and so are their gaps. Independence is restored only by comparing against an external source — the written clearance, the flight plan, the chart — rather than against what the other pilot just called out. A barrier that queries the same representation as the previous one does not count as a second barrier.
Why do onboard alerts detect so late?
Because they are built on threshold crossings, not on intentions. An altitude deviation alert, a ground proximity warning, or a traffic resolution advisory all assume that a parameter has already left the expected range — that is, that the undesired state already exists. These are safety nets, sized to catch what has escaped everything else, not monitoring instruments. A crew whose alerts trigger regularly does not have poor machine barriers: it has failing human detection.
What should you say in an interview or a check when asked for an example of an error?
Choose a real error, detected early, and recount the chain rather than the mistake: what made it possible, the exact signal that revealed it, the delay, the diagnosis reached, the correction and its verification. This is what competency frameworks seek to assess under the heading of threat and error management. Claiming never to have made an error produces the opposite of the intended effect, since it mainly signals a lack of detection or a lack of honesty.

Leave a Reply
You must be logged in to post a comment.