The datasheet is the contract
In the previous guide you saw the design as a loop: pin down the spec, sketch a topology, do the back-of-envelope sums, then choose the actual components, simulate, build, and iterate. This guide lives inside that one innocent-looking step — choose the components — because it hides the difference between a circuit that works on the bench once and a circuit that works on ten thousand boards, in July and January, for ten years. A schematic full of round numbers is a wish. A real circuit is built from parts that come with a datasheet: the manufacturer's binding promise about what that part will and will not do.
Open any datasheet and the single most important habit is this: read the columns, not just the headline number. A serious part lists each parameter as a minimum, a typical, and a maximum. The typical column is a friendly lie — it is what one average part did on one good day, and you must never design to it. The engineering lives in the min and max columns, because those are the only numbers the manufacturer actually guarantees. If a voltage reference reads 'typ 2.500 V, max 2.512 V, min 2.488 V', your design has to be happy anywhere in that ±12 mV band, because any part you buy is allowed to land there.
Selection is also more than electrical fit. The best part on paper is useless if it is out of stock, single-sourced from one fragile supplier, end-of-life, or wrong for your assembly line. Mature designers choose parts that are buyable, second-sourced, and in a package the factory can actually solder — the first quiet step of designing for manufacture. The list of every part and its alternates becomes the bill of materials, the document that turns your schematic into something a factory can build a thousand times.
Every number is a range
Start with the humblest part, the resistor — and the first surprise of real electronics: a '10 k' resistor is almost never exactly 10.000 k. Its tolerance is how far it is allowed to stray, marked by that last colour band you met back in the foundations rung — gold for ±5%, brown for ±1%. A 5% '10 k' is legally anything from 9.5 k to 10.5 k; the factory makes one huge batch, measures them, and sorts them into tolerance bins. Every passive part plays this game: ceramic capacitors are looser still (often ±10% or ±20%), and electrolytics can swing −20% to +80%.
Now feel what that does to a real circuit. Take the voltage divider you have used since the very first rung: 10 k on top, 5 k on the bottom, across a 9 V supply, giving 9 times 5 / (10 + 5) = 3.00 V. With 5% resistors, ask the cruel question — what is the highest the output can go? Make the top resistor as small as allowed (9.5 k) and the bottom as large as allowed (5.25 k): 9 times 5.25 / (9.5 + 5.25) = 3.20 V. The lowest does the opposite (top 10.5 k, bottom 4.75 k): 9 times 4.75 / (10.5 + 4.75) = 2.80 V. So your '3.00 V' node is really 2.80 V to 3.20 V — a ±6.7% spread squeezed out of merely ±5% parts. (And that is before the 9 V supply's own tolerance, which would widen it further.)
And tolerance is not a one-time number. A part also drifts with temperature (its tempco), ages over years, and shifts a little with humidity and self-heating. The number on the band is only the spread at the moment of manufacture, at room temperature. There is, though, a beautiful escape hatch: if two resistors are made together and share the same temperature coefficient, they drift together, so their ratio barely moves even as both wander. That is why a precision divider or a gain-setting pair is built from a single matched resistor network, not two loose parts from different reels — the absolute values still drift, but the ratio that actually sets your result holds steady.
Stacking tolerances and the worst-case corner
One toleranced part is manageable; the trouble is that a real circuit chains many of them, and their uncertainties pile up. This is tolerance stack-up: when several imperfect parts together determine one output, the result's spread is built from all of theirs at once. A divider's two resistors, a regulator's reference and its feedback pair, an amplifier's gain network — each contributes its own slop, and the question is no longer 'what does this do?' but 'what is the full range of what this could do across every legal combination of parts?'
The most conservative way to answer is worst-case design: imagine every single part landing simultaneously at the exact corner of its tolerance that hurts you most, and prove the circuit still meets spec even then. Walk it through on the LED current limiter from way back — pick R so that (5 - 2) / 0.01 = 300 ohm for a nominal 10 mA. But the supply is 5 V ±5%, the LED's forward drop is not a fixed 2 V but a datasheet range (say 1.8 V to 2.2 V), and the resistor is ±5%. To find the brightest worst case, line up every corner that raises current at once; to find the dimmest, flip them all.
* LED current limiter -- worst-case corner analysis
* I = (Vsupply - Vf_LED) / R nominal: (5 - 2) / 300 = 10 mA
LOW current corner NOMINAL HIGH current corner
Vsupply (5 V +-5%) 4.75 V 5.00 V 5.25 V
Vf LED (1.8..2.2 V) 2.20 V (high) 2.00 V 1.80 V (low)
R (300 +-5%) 315 ohm (high) 300 ohm 285 ohm (low)
--------------------------------------------------------------------------------
I = (V-Vf)/R (4.75-2.20)/315 (5-2)/300 (5.25-1.80)/285
= 2.55/315 = 3.00/300 = 3.45/285
= 8.1 mA = 10.0 mA = 12.1 mA
A '10 mA' LED really runs 8.1 mA .. 12.1 mA -- a ~+-20% spread from +-5% parts.Derating: ratings are limits, not targets
Tolerance is about a part's value being uncertain; derating is about not running a part anywhere near its breaking point. Every datasheet lists absolute maximum ratings — the voltage, current, power, and temperature beyond which the part is damaged. The rookie mistake is to treat those numbers as targets. They are cliffs. A capacitor rated 16 V will not magically survive on a 16 V rail with a transient spike on top; a resistor rated 1/4 W run at 1/4 W gets blisteringly hot and ages fast. The cure is design margin: deliberately run every part at a comfortable fraction of its rating, so tolerance, temperature, and surprises all have somewhere to go before anything breaks.
The rules of thumb are concrete. Run a resistor at most at half its rated power — if it must dissipate 0.2 W (P = V times I, or I^2 times R), do not reach for a 1/4 W part, fit a 1/2 W one. Pick an electrolytic capacitor's voltage rating well above the rail — a 25 V cap on a 12 V supply, not a nervous 16 V one. Keep a power MOSFET's drain-source voltage and its junction temperature far below their maxima, because a switching transient can ring well above the steady rail and silicon dies young when it runs hot — which is what the heat sink is for, dragging heat out so the junction stays cool. Margin is not waste; it is the price of a circuit that still works on the worst day.
- Compute the worst-case stress on the part — the highest voltage, current, or power it will ever see, using the worst-case corners from earlier, not the nominal values.
- Read the part's absolute maximum rating for that stress straight from the datasheet.
- Apply a derating factor: aim to use no more than about 50% of a power or current rating, and perhaps 60-80% of a voltage rating, with more margin where reliability matters.
- If the chosen part fails the check, step up to the next size, voltage, or power rating and recompute — do not just hope the spike never comes.
- Finally, confirm the part stays cool: estimate its temperature rise and check the junction or body temperature is comfortably under its limit across the whole operating range.
Corners over temperature and life — and honest traps
Tolerance and derating both sharpen into one idea: a robust design must hold across corners — the combinations of extremes it will actually meet. The big three are part tolerance, supply voltage, and temperature, and temperature is the sly one because it quietly moves nearly every parameter. A resistor's value drifts, a capacitor's value sags, a transistor's gain climbs. The diode you have leaned on all along is the clearest warning: that handy '0.7 V drop' is a model, not a constant — the real forward voltage follows an exponential law, so it shifts with current and falls by roughly 2 mV for every degree the junction warms. Design as if 0.7 V were carved in stone and a cold or hot day will move it out from under you.
The transistor hides a fiercer version of the same trap. A bipolar transistor's current gain, beta, is not a number you can trust: the datasheet might say it ranges from 100 to 400 for the same part number, and it shifts with temperature and current on top. A novice biases a transistor by counting on beta — and the circuit that works on one sample drifts hopelessly on the next. The mature answer, which you met in the BJT rung, is to design the bias so it barely depends on beta at all: a stiff voltage-divider bias plus an emitter resistor (emitter degeneration) sets the operating point from resistor ratios and the supply, quantities you actually control, leaving beta free to vary without dragging the Q-point with it. Good design does not predict the uncontrollable; it engineers around it.
Step back and the whole worldview comes into focus. Choosing components is never about the typical part on a good day; it is about the whole population of parts you will ever buy, across every temperature, supply, and year of service. Lean on the work others have done — a manufacturer's reference design is a vetted, worst-cased starting point, and a clean bill of materials with sane alternates is your hedge against a part vanishing. Designing for reliability means thinking about how each part can fail, not just how it works, so that a single weak component cannot quietly take the whole board down. With your components chosen and cornered, the next guide turns to the harder question of making blocks play together — how a sensor, an amplifier, a filter, and a converter share impedances, levels, and power without each one wrecking its neighbour.