Twelve questions on what a sensor really returns, why readings scatter, what a control reading and the integration time are for, and what makes a measurement series evidence. Every answer comes with its reason.
An AI assistant suggests doubling a sensor's integration time, "then the readings will get better". Why is this advice not enough?
Because the suggestion comes from general knowledge about sensors; only confirmation by an expert makes it reliable.
A second opinion is still just an opinion. Even the most knowledgeable person does not know this particular setup; only measuring on it makes the suggestion reliable.
Because the assistant does not explain why a longer window yields calmer readings; with a sound explanation the advice would be sufficient.
A sound explanation is no substitute for a measurement. It says why something should happen, not whether it happens in this setup and how strongly. That is precisely the fallacy of mistaking reasoning for measuring.
Because nothing but the setup itself knows how the setup behaves; a measurement decides that.
Exactly: the physical channel is the ground truth no language model has access to. Whether "better" is true only shows in a measurement series before and after the change.
Because "better" can only be judged once the assistant also predicts which values the sensor will deliver after the change.
A prediction is still just a claim, merely with numbers attached. Whether it holds is decided by the same measurement series that would have to test the original suggestion anyway.
A suggestion, an explanation and a prediction are all claims, and only a measurement on the setup itself turns one of them into evidence.
Between two measurement series, a group changes both the distance between sender and sensor and the integration time. The second series is clearly better. What do they know now?
That both changes helped, otherwise the second series would not be so clearly better.
How clear the improvement is says nothing about how it splits: it may have been one change alone, and the other may even have hurt without anyone noticing.
That it got better, but not what caused it.
Exactly the problem: two quantities changed at once can no longer be separated in the result. That is why you change one quantity per series.
That the distance was decisive, since it has the stronger physical effect.
That is a guess dressed up as a result. How strongly each quantity acts was precisely the open question.
That the series is unusable and has to be repeated before anything can be said at all.
That is stricter than necessary: the series does show that the combination helps. It just does not tell which part did what; two clean follow-up series with one change each will.
Change exactly one quantity per series, otherwise a difference cannot be put down to its cause.
Before measuring the colours of an LED with a colour sensor, you take a control measurement: readings while the LED is switched off. What for?
To check whether the sensor responds at all, before time goes into the real measurements.
A quick function check comes for free, but it would not need a recorded series. The purpose lies in the values themselves: they are the reference for everything that follows.
So that the sensor adjusts to the ambient light and afterwards subtracts it from the readings by itself.
The sensor learns nothing and subtracts nothing; it keeps delivering raw values including room light. Relating the readings to the baseline is your job, in your program and in your interpretation.
To know what the sensor measures from ambient light anyway; only against that can the readings with the LED on be interpreted.
Exactly: the sensor never reads zero, because room light is always there. The control measurement is the baseline every colour reading is compared against, and it reveals when the surroundings change.
To see how much the readings scatter without light; once the LED is on, the noise disappears in the signal anyway.
Recording the spread is right, the conclusion is not: noise does not vanish once there is a signal, it still sits on every reading. And the main purpose remains the baseline every colour reading is read against.
Every reading contains ambient light, the light of the LED and noise; the control reading measures the first part, so that the second can be read off.
You measure the same red LED five times in a row and get these red components: 198, 205, 201, 195, 206. What follows?
The sensor is broken, otherwise the same value would appear five times.
Spread is the normal case of any real measurement: noise from sensor, light and surroundings. Five identical values would be more suspicious.
Readings scatter; if you need a stable value, average over several measurements.
Exactly: single values jump, the average (about 201 here) sits still. That is why you read several times and average instead of trusting one reading. And the scatter itself is a result: it belongs in the record, because later it decides how many colours can be told apart.
The LED slowly gets weaker while being measured.
The values go up and down, not steadily downhill. That is noise, not a trend.
Only the last value counts, since the sensor had settled by then.
The last value is just one more draw from the same spread. It is no truer than the other four.
A single reading is a draw; the mean of several is stable, and the spread around it is a result of its own.
Which statements about the values of a colour sensor that reads the light of an LED are true? (Several answers may be correct.)
It delivers three values, red, green and blue, matching the three lamps inside the LED.
There are four: besides red, green and blue there is the clear channel, which measures without a filter. The symmetry with the LED is tempting, but the sensor is a different part with its own logic.
Besides red, green and blue there is a fourth, unfiltered channel (clear).
Correct: clear measures all incoming light without a colour filter and is therefore more sensitive than the three colour channels.
The measured values do not simply match the RGB numbers the sender puts into its LED.
Correct: overlapping colour filters, distance and ambient light shift the values. A transmitted (255, 0, 0) never arrives as (255, 0, 0).
With the LED off, all channels read zero, because without a sender no light reaches the sensor.
No: the sender is not the only light source. Ambient light is always present, and that is exactly why the control measurement is needed as a baseline.
The sensor returns four numbers, r, g, b and c, and none of them is the number that was sent.
A sensor can read the same light with long, medium or short reading windows. To transmit faster, a group shortens the sensor's integration time drastically. What should it expect?
The readings stay as calm as before and arrive more often; the window only sets how often the sensor delivers a new value.
The rate part is right, the calm part is not: the window also sets how much light one reading averages over. Shorter means less light per reading, so the noise hits harder.
The readings get calmer because the sensor has less time to be noisy.
The other way round: a short window averages over less light, so the noise hits harder, and single short readings cross into the neighbouring level.
More readings per second, but with more spread; single readings may cross into the wrong level.
Exactly the trade-off: short windows for a high symbol rate against long windows for calm readings. Where the limit of a setup lies, only a measurement series can tell.
The readings scatter more, but the number of readings per second drops too.
Half right, half wrong: the spread increases, but a shorter window allows more readings per second, not fewer. Otherwise there would be no trade-off.
Longer is calmer and shorter is faster: the integration time trades calm readings against speed.
A sender puts (255, 0, 0) into its LED, but the sensor reads r=203, g=41, b=57. How do you deal with that?
Complain about the sensor: an intact sensor should read (255, 0, 0) when (255, 0, 0) is sent.
The sensor is fine. Overlapping colour filters, distance and ambient light ensure that readings never equal the transmitted numbers.
Measure how each transmitted colour arrives, and match future readings to these learned patterns.
Exactly: calibrate and classify. Do not try to compute back what was sent; learn each colour's typical arrival pattern and match new readings to the nearest one.
Derive a conversion formula that computes the exact transmitted RGB numbers back from the readings.
Tempting but fragile: the distortion depends on distance, ambient light and noise, and changes with them. The back-computation that works today fails tomorrow.
Turn the LED brighter until the readings match the transmitted values.
The numbers live in different worlds: drive values of the LED against readings of the sensor. No brightness makes them coincide.
Calibrate, then classify: references of known colours, and every new reading goes to the nearest of them.
What belongs in the record of a measurement series? (Several answers may be correct.)
What was changed and what stayed constant.
Correct: without this you cannot later say what a difference was due to.
The setup: distance, ambient light, sensor settings.
Correct: only with the conditions is a measurement repeatable and comparable with others.
All readings, including the inconvenient ones.
Correct: recording only the nice values is not measuring but collecting evidence for one's own hopes. Outliers are often the most interesting information.
How many times the measurement was repeated.
Correct: whether a result rests on 5 or on 50 repetitions decides how much weight it can bear.
The record holds the conditions, the changed quantity, the number of repetitions and every value, so that someone else could repeat the series.
A receiver is supposed to tell red from blue. In a test it is right 5 times out of 5, and someone says: "Done, it works." What is the catch?
Nothing, five hits without a single miss are solid proof that the distinction works.
With two colours, pure guessing gets five hits in a row with a probability of about three percent. In a room full of groups that happens somewhere all the time.
Five hits can still be luck with two colours; only many repetitions separate skill from chance.
Exactly: guessing hits 50 percent per trial. 5 out of 5 happens by chance in roughly one run in 32; 50 out of 50 practically never. That is why receivers are tested with long series.
The test should have used more than two colours.
The task was red against blue; more colours are a different, later task. The catch is the small number of trials, not the number of colours.
Five hits only show that the threshold happens to fit; what matters is how far the mean of the readings sits from the threshold.
The margin to the threshold is a useful quantity, but a mean from a few readings is itself just one draw. The catch remains the small number of trials, not where the threshold sits.
The chance of five hits in a row by guessing is 0.5⁵ = 1/32, the chance of fifty is 1 in 2⁵⁰.
Why do you decide before a measurement series what result you expect, instead of just measuring and then having a look?
Because the sensor settings can only be chosen sensibly once it is clear beforehand which values are supposed to come out.
The settings follow the setup and the ambient light, not the hoped-for result. The sensor measures with any setting; the problem arises later, when interpreting the values.
So the series can be stopped as soon as the expected result is reached, avoiding superfluous repetitions.
That would be exactly the bias to avoid: stopping as soon as it fits collects evidence for your own hopes. The number of repetitions is fixed beforehand and carried through.
Without an expectation set beforehand, almost any result can be interpreted as a success afterwards.
Exactly: measuring first and deciding afterwards what counts as "good" silently bends the interpretation towards the hoped-for result. An expectation written down beforehand turns the measurement into an honest yes-or-no test.
Because a record without a written expectation does not count as a measurement series, only as a collection of single values.
What makes a measurement series is repetition and recorded conditions, not the expectation. It serves not the form but the protection against fooling yourself.
The expectation written down beforehand protects you from reading your own hopes into the data afterwards.
What makes a measurement series, as opposed to a single measurement?
That the measuring is done with special care: a steady setup, a properly configured sensor, and every value written down immediately.
Care is part of it but does not make a series: a single careful measurement is still one draw. The difference lies in repetition, recorded conditions and the single changed quantity.
Several measurements under equal, recorded conditions, with at most one quantity changed on purpose.
Exactly: repetition captures the spread, the recorded conditions make it repeatable, and the single changed quantity makes differences interpretable.
As many measurements as possible at as many different settings as possible, so that every relationship can be read from the data afterwards.
Measuring a lot by itself yields a pile of data. If several quantities changed between two values at once, no relationship can be attributed cleanly afterwards; hence one quantity per series.
The same measurement at different times of day, so that random fluctuations cancel out on average.
Repeating is right, but at different times of day the ambient light changes without plan. The conditions then shift along, and nobody knows whether a difference comes from chance or from the light.
Repetition, recorded conditions and one changed quantity turn single readings into a measurement series.
An AI assistant suggested increasing the sensor's gain, "that improves recognition". Your clean measurement series shows the opposite: the recognition rate drops. What do you do? (Several answers may be correct.)
Document the measurement series including its conditions.
Correct: without recorded conditions the contradiction can neither be retraced nor repeated later.
Distrust the measurement; the AI knows such sensors better than you.
The wrong way round: the AI knows sensors of this kind, but not your setup. The measurement on the real device is the ground truth no language model has access to.
Briefly check that the series was clean: one quantity changed, enough repetitions, a control measurement in place.
Correct: before filing a suggestion as refuted, your own measurement deserves scrutiny. Only a clean series can carry the contradiction.
Add the case to a log of mistakes: suggestion, measurement, outcome.
Correct: that is exactly what such a log is for. Documented cases in which a measurement refuted a suggestion are among the things you learn most from.
Keep the higher gain anyway, since a single contradicting measurement series carries less weight than a well-reasoned recommendation.
The weighting is upside down: a rationale says why something should hold, the clean series shows what holds in your setup. Deciding against your own clean measurement means giving up on measuring.
Check your own series first; if it was clean, it outweighs the suggestion, and the case goes into the log of mistakes.