ProjekinAja. Embedded Systems Stage 1 · Foundations Lesson 1.3
● Stage 1 · Lesson 1.3

Kirchhoff's Laws and Circuit Analysis by Hand

You will never run SPICE on every node — but you will always be able to sum voltages around a loop.

⏱ ~26 min read 🧮 Analysis you will do on napkins 🔋 Failure class: electrical → integration

🎯 By the end you'll understand

Kirchhoff's voltage and current laws as conservation statements you can apply without thinking. Series and parallel reduction, and the voltage divider — the most used and most misused circuit in embedded work. Thevenin equivalents as the tool that turns "what happens when I connect this?" into one multiplication. Why Beacon's 2 × 1 MΩ battery divider reads catastrophically low into an ADC, why the fix is a capacitor plus a wait rather than smaller resistors, and what the trade costs in sleep current. And why a ground trace is a resistor that KVL will not let you ignore.

KEY INSIGHTEvery divider is a source with an output resistance. Anything you attach loads it.

The firmware said the battery was at 1.2 V. The multimeter said 3.7 V. Both instruments were correct. They were measuring different circuits — because one of them was part of the circuit it measured.

Here is the bench scene, and it is one of the most common first-board bugs in battery products. The engineer needs to read the battery voltage. The cell sits at 3.7 V, the ADC reference is 3.3 V, so a divider is needed. Two 1 MΩ resistors — chosen large, correctly, so the divider does not drain the battery — produce 1.85 V at the midpoint. The multimeter confirms 1.85 V. The ADC reads the equivalent of 0.6 V. The engineer suspects the ADC, the reference, the firmware calibration, and finally the resistors, and replaces each in turn. None of it helps, because the divider is not broken. It is doing exactly what a 500 kΩ source does when you connect something to it: sagging. This lesson gives you the two laws and the one theorem that make that sag predictable in your head, before the board exists.

🔑 Hold onto this

KVL: the voltages around any closed loop sum to zero — energy is conserved. KCL: the currents into any node sum to zero — charge is conserved. From those two laws come series and parallel reduction, the voltage divider, and the Thevenin equivalent: any linear network seen from two terminals is a voltage source in series with a resistance. A divider's Thevenin resistance is R1 ∥ R2, and that resistance is what everything downstream sees. If the load is not much larger than it, the divider's output moves. The fix is almost never "smaller resistors" — that just moves the failure from accuracy to battery life.

CORE CONCEPTTwo conservation laws, and why they are never wrong

Kirchhoff's laws are not empirical rules that hold most of the time. They are restatements of conservation of energy and conservation of charge for lumped circuits — networks whose dimensions are small compared with the wavelength of the signals in them, which is every circuit in this course except the antenna. That is why you can trust them completely and why, when a measurement seems to violate one, the measurement is wrong or the circuit is not the one you drew.

Kirchhoff's voltage law (KVL)

Go around any closed loop in a circuit, adding up the voltage rises and subtracting the drops. You arrive back where you started at the same potential, so the sum is zero:

Σ (voltages around a loop) = 0

In the LED circuit of Lesson 1.2: +3.3 V (source) − 1.3 V (resistor) − 2.0 V (LED) = 0. The practical use is the other way round: if you know the source and all but one drop, KVL gives you the last one. It is also how you find bugs — if a loop's measured voltages do not sum to zero, there is an element in the loop you did not draw: a trace resistance, a diode, a bad joint.

Kirchhoff's current law (KCL)

At any node (a point where conductors join), the current flowing in equals the current flowing out. Charge cannot pile up at a junction:

Σ (currents into a node) = 0

A node feeding a divider's midpoint to an ADC has three connections: current arrives through R1, leaves through R2, and leaves into the ADC. KCL says IR1 = IR2 + IADC. When IADC is zero, the two resistors carry the same current and the divider formula holds. When IADC is not zero, R1 must carry more than R2 — which means more voltage across R1 (Ohm), which means less at the midpoint (KVL). That is the sag, in two sentences, and it is worth being able to say it that quickly.

KVL — around a loop 3.3 V + rise 130 Ω − 1.3 V − 2.0 V 3.3 − 1.3 − 2.0 = 0 a loop whose drops do not sum to the source contains an element you did not draw KCL — at a node R1 V_bat I₁ in R2 I₂ out ADC I_adc out I₁ = I₂ + I_adc if I_adc ≠ 0, R1 carries more than R2 → midpoint sags
The two laws in the two situations where you will use them most. KVL finds a missing drop or exposes an element you did not draw. KCL at the divider's midpoint is the entire explanation of the loading problem: any current the ADC takes must come through R1, and R1 charges for it in volts.

HOW IT WORKSSeries, parallel, and the divider you will draw a thousand times

Series: same current, voltages add

Resistors in series carry the same current (KCL — there is no node between them for current to leave), so their voltages add (KVL), and the equivalent resistance is the sum:

Rseries = R1 + R2 + …

Two 1 MΩ resistors in series across a 3.7 V cell: 2 MΩ, 1.85 µA, and by symmetry 1.85 V at the midpoint. That is Beacon's battery divider, and 1.85 µA is the number Lesson 1.2 flagged as 23 % of the sleep budget.

Parallel: same voltage, currents add

Resistors in parallel have the same voltage across them (they connect the same two nodes), so their currents add (KCL), and conductances add:

1 / Rparallel = 1/R1 + 1/R2 + …     (two resistors: R1·R2 / (R1 + R2))

Two 1 MΩ in parallel: 500 kΩ. Two useful limits: parallel resistance is always smaller than the smallest member, and a very large resistor in parallel with a small one barely changes the small one (10 kΩ ∥ 1 MΩ = 9.9 kΩ). That second limit is the "load must be much larger than the source" rule, in disguise.

The voltage divider

Two resistors in series across a source, output taken from the midpoint. The same current I = Vin / (R1 + R2) flows through both, so the output across R2 is:

Vout = Vin · R2 / (R1 + R2)

This formula is on every reference card and it is correct only when nothing is connected to the output. Every time it is used for an ADC input, a comparator, a MOSFET gate with a pull-down, a level shifter, or a sensor bias, something is connected — and the formula's hidden assumption (Iload = 0) gets tested. Most of the time the load is large enough that nobody notices. The times it is not are the subject of the next section.

💡 Three dividers you will meet this course, and their hidden loads

Battery sense (this lesson): the ADC's sampling capacitor and pin leakage. USB-C CC detection (Lesson 4.4): the 5.1 kΩ pull-down on CC forms a divider with the source's pull-up; the "load" is the comparator inside the charger — designed for, so it works. A resistive level shifter from a 5 V signal to a 3.3 V input (Lesson 2.1): the input's capacitance turns the divider's Thevenin resistance into an RC filter, and a fast edge becomes a slow one. Same physics each time; only the consequence differs.

DEFINITIONThevenin: every divider is a source with an output resistance

Thevenin's theorem says: any linear two-terminal network — resistors and sources, however tangled — behaves at those two terminals exactly like a single voltage source Vth in series with a single resistor Rth. Two measurements (or two calculations) determine them:

For the divider: Vth = Vin · R2/(R1+R2), the familiar formula. For Rth, short the source: R1 now goes from the output to ground, as does R2 — they are in parallel. So Rth = R1 ∥ R2. Check with the other method: Isc = Vin/R1 (shorting the output bypasses R2), so Rth = [Vin·R2/(R1+R2)] / [Vin/R1] = R1·R2/(R1+R2). Same answer.

the divider V_in = 3.7 V R1 = 1 MΩ out R2 = 1 MΩ 1.85 V open-circuit ≡ Thevenin equivalent V_th 1.85 V R_th R1 ∥ R2 = 500 kΩ identical at the terminals connect a load R_L 240 kΩ V_L = V_th · R_L / (R_th + R_L) = 1.85 × 240 / 740 = 0.60 V ×2 in firmware → the "1.2 V battery" of the opening scene
The divider, its Thevenin equivalent, and what happens when a 240 kΩ-class load is attached. The loaded output is just another divider: Rth on top, RL below. With Rth = 500 kΩ, a 100 kΩ load leaves one sixth of the voltage. Scaled back up by the firmware's ×2, that reads as a "1.2 V battery" — wrong by a factor that has nothing to do with the ADC's accuracy.

Once the divider is a Thevenin source, loading is a single multiplication: VL = Vth · RL / (Rth + RL). The error relative to the ideal is Rth / (Rth + RL). For a 1 % error you need RL ≈ 100 · Rth; for 0.1 %, a thousand times. With Rth = 500 kΩ, a 1 % reading needs a 50 MΩ load. That is not a number any ADC input will give you continuously — which is the whole problem.

What an ADC input actually looks like

A successive-approximation ADC — the type in the ESP32-C3 and almost every microcontroller — does not have a resistive input. It has a small sampling capacitor (picofarads) that is switched onto the pin for a brief acquisition window, charged to the input voltage, then disconnected and measured. Two consequences. First, during acquisition the source must charge that capacitor through Rth: the time constant is Rth · Cs, and if the acquisition window is shorter than several time constants the sample ends before the capacitor reaches the true voltage, and the reading is low. With Rth = 500 kΩ and a 10 pF-class sampling capacitor, τ = 5 µs, which is longer than many SAR acquisition windows. Second, if you sample repeatedly, the capacitor takes a gulp of charge each time, and averaged over time that looks like a resistor of roughly 1 / (fs · Cs) — which for hundreds of kilohertz and tens of picofarads lands around a hundred kilohms. That is the "100 kΩ-ish input impedance" the folklore describes; it is a description of a switched capacitor, not a resistor, and the distinction is exactly what makes the fix work.

📖 The fix: a capacitor plus timing, not a smaller divider

Put a capacitor Chold across R2 — Beacon uses 100 nF. Now the sampling capacitor's gulp comes from Chold, not through Rth: charge sharing between 100 nF and ~10 pF moves the voltage by one part in ten thousand per sample. The divider's high resistance no longer matters during the sample. It matters before: Chold must be charged to the true divider voltage through Rth, with time constant Rth · Chold = 500 kΩ × 100 nF = 50 ms. Settling to 99 % takes about 5τ = 250 ms. On a board where the divider is switched off in sleep and the capacitor starts empty every wake, that wait is a firmware requirement, not a suggestion. Smaller resistors would shorten the wait — and multiply the divider's current by the same factor. The capacitor moves the cost from battery life to time, and time is the cheaper currency here.

SEE ITLoaded divider — accuracy against sleep current

This bench models Beacon's battery sense: a 3.7 V cell, a divider you choose, a PMOS switch that connects the divider only for a measurement, an optional hold capacitor, and a SAR ADC modelled as a small sampling capacitor with an acquisition window, a per-sample interval, and pin leakage. The ADC parameters are illustrative — the ESP32-C3 datasheet does not publish a sampling capacitance or acquisition time, which is itself a lesson for 1.6 — but the physics is the physics. Start with the canon values (1 MΩ / 1 MΩ / 100 nF / 250 ms) and then try to "improve" them.

Interactive · Loaded divider bench

Lower the divider to fix the reading and watch the sleep budget go. Add capacitance to fix it and watch the wait grow. There is no free axis.

1 MΩ
1 MΩ
250 ms
Ideal midpoint
1.850 V
What the ADC reads
1.837 V
Reported battery
3.67 V
Divider's sleep-budget cost
0.01 µA

Voltage on the ADC node after the PMOS connects the divider, versus time, against the ideal midpoint. The reading is taken at the end of your chosen wait. Without a hold capacitor the node settles in microseconds but each sample then starves through Rth; with one, the node settles slowly but the samples are clean.

Things to try. Remove the hold capacitor with the 1 MΩ divider: the node settles instantly, and the reading collapses, because every 2 µs acquisition tries to charge 10 pF through 500 kΩ. Now "fix" it by dropping both resistors to 10 kΩ: the reading is perfect, and if the PMOS ever sticks on — as Rev A's did — the divider costs 185 µA, twenty-three times the entire sleep budget. Go back to 1 MΩ, add 100 nF, and shorten the wait to 25 ms: the node reaches only about 40 % of the true value, because 25 ms is half a time constant. At 250 ms it is within about 1 %, and most of that remainder is the settling tail itself (5τ still leaves 0.67 %); the pin's leakage through 500 kΩ accounts for only about 0.27 %. Then try 1 µF: the droop from 64 samples vanishes but the wait becomes 2.5 s, longer than the whole two-second motion-to-report budget. And finally set 1 µA of leakage: half a volt of DC drop across 500 kΩ that no capacitor fixes, which is why the divider's resistance has a ceiling as well as a floor.

BEACONBeacon's battery sense, with every number justified

The canon fixes the battery-sense circuit as 2 × 1 MΩ, a 100 nF capacitor across the lower resistor, and a small PMOS that connects the divider's top to the cell only during a measurement. Here is the reasoning behind each value, in the order a designer actually decides them.

1

Ratio: 1 : 1, because the ADC's usable range says so

A LiPo cell runs from about 4.2 V full to 3.0 V empty. The ESP32-C3's ADC, with its highest attenuation setting, measures usefully up to roughly 2.5 V (Lesson 1.6 covers the ADC's range and its nonlinearity near the top). Halving the cell voltage puts 4.2 V at 2.1 V and 3.0 V at 1.5 V — inside the range with margin at both ends. Equal resistors also make the firmware's scale factor exactly 2, which is one fewer thing to get wrong.

2

Magnitude: 1 MΩ, as insurance against the switch

If the PMOS switch worked perfectly the resistors could be 10 kΩ — a 185 µA divider connected for a millisecond a minute costs 0.003 µA on average. The canon chooses 1 MΩ anyway, because switches leak, gate drives float in deep sleep, and firmware forgets. With 1 MΩ the worst case — divider permanently connected — costs 1.85 µA, a 23 % hit on the sleep budget that hurts but does not kill. With 10 kΩ the same failure would end the product. Rev A's post-mortem records precisely this: an extra 1.8 µA, meaning the switch was not isolating, and the product still met its six months because of the resistor value. High resistance here is defense in depth, and the cost of that defense is the next two decisions.

3

Hold capacitor: 100 nF, sized between droop and settling

The capacitor must be large compared with the ADC's sampling capacitor so that 16 or 64 back-to-back samples do not droop it — 100 nF against ~10 pF is 10,000 : 1, so 64 samples droop by well under 1 %. It must be small enough to settle through Rth = 500 kΩ in a time the firmware can afford: τ = 50 ms, 5τ = 250 ms. A 1 µF part would droop less but need 2.5 s. A 10 nF part would settle in 25 ms but 64 samples would droop it by about 6 %. 100 nF is the compromise, and it doubles as a low-pass filter against the BLE burst's supply noise.

4

Timing: enable first, read last, hide the wait

250 ms is longer than Beacon's 45 ms active window, so the divider is enabled as the first act on wake and read as the last, with the sensor conversions and radio setup in between; the remainder of the wait is spent in light sleep (about 130 µA on the ESP32-C3 per its datasheet), which costs roughly 130 µA × 0.2 s / 60 s ≈ 0.4 µA of average current — inside the canon's 38 µA rounding. The 2 s motion-latency requirement (R4) absorbs 250 ms easily. This ordering is a firmware requirement that comes from a resistor value, and it is written into the state machine in Lesson 3.7.

🔑 What the firmware must know

The battery reading is v_adc × 2 only if the divider has settled and the ADC has been calibrated. Beacon's firmware therefore records three things per measurement: the wait actually achieved, the number of samples averaged, and the raw ADC value before scaling. When the pilot's data (Lesson 6.7) shows one unit reporting a battery 8 % lower than its neighbours, those three fields let you tell a leaky PMOS from a short wait from a bad ADC calibration without a site visit.

EXAMPLEKVL in the ground plane: the resistor you did not draw

KVL has a second job in embedded work that has nothing to do with dividers: it forces you to account for the voltage across conductors you were treating as ideal. A copper trace has resistance. A 0.2 mm-wide trace of 35 µm (1 oz) copper is roughly 2.4 mΩ per millimetre at room temperature — an illustrative figure from the standard copper resistivity; the IPC-2221 trace calculations in Lesson 5.6 make it exact. A 20 mm ground return of that width is about 50 mΩ. Nothing, until the radio fires.

During Beacon's BLE burst the module draws tens of milliamps with sharper peaks, and if that current returns to the battery through the same thin ground trace the ADC's ground reference uses, KVL says the ADC's "0 V" is now V = I · R = 0.05 A × 0.05 Ω = 2.5 mV above the battery's negative terminal — during the burst. If the firmware reads the battery ADC during a transmission, the reading is offset by a few millivolts, which at the flat part of a LiPo's discharge curve can be several percent of state of charge. The symptom in the field is "battery reads lower every time it transmits." Nobody drew that 50 mΩ resistor; KVL insists it is in the loop.

🧪 The engineer's reflex

When a measurement changes with something that "should not affect it" — the battery reading changing with radio activity, a sensor reading changing when the LED is on — write out the loop from the reference point of the measurement to the reference point of the thing that changed. If the two loops share a conductor, that conductor is a resistor carrying the other loop's current, and KVL gives you the error to within a factor of two before you touch a probe. Stage 5's ground-plane discipline (a solid plane under everything, sensitive returns not sharing paths with high-current ones) is this reflex, made into layout rules.

WEAKNESSWhere this bites

✗ Trusting the divider formula with something connected

The formula Vout = Vin · R2 / (R1 + R2) is an open-circuit statement. Every use of it in a real design connects something to the output, and the only question is whether that something is large compared with R1 ∥ R2. Multimeter inputs (10 MΩ) usually are; ADC sampling capacitors, comparator inputs with bias currents, MOSFET gates with pull-downs, and anything with a resistive input often are not. The reflex: compute Rth = R1 ∥ R2 every time you draw a divider and write it next to the schematic symbol, exactly as you write the power next to a resistor. Then ask what the load is. If you do not know, assume it is bad and add the capacitor.

✗ Fixing loading with smaller resistors

It works, which is why it is dangerous. Dropping Beacon's divider from 1 MΩ to 10 kΩ makes the ADC reading perfect with no capacitor and no wait — and turns a 1.85 µA worst case into 185 µA. On a mains-powered board that is a fine trade. On a battery product it converts an accuracy problem into a battery-life problem, and the battery-life problem does not show up until the switch that was supposed to isolate the divider leaks, floats, or is left on by a firmware path nobody tested. The capacitor-plus-timing fix keeps the resistance high and pays in a currency Beacon has plenty of: 250 ms once a minute.

✗ Believing the "input impedance" number

Forum answers quote an ADC "input impedance" as if it were a resistor: "the ESP32 ADC is about 100 kΩ, so use a divider below 10 kΩ." That advice is half right for the wrong reason. A SAR input is a switched capacitor; its apparent resistance depends on how often you sample and how long the acquisition window is, and a hold capacitor makes the number irrelevant. Designing to a fictional resistor produces dividers that waste current for no accuracy gain, or, worse, a divider that meets the fictional spec and still reads low because the acquisition window is shorter than Rth · Cs. Model the capacitor, not the folklore. And when a datasheet does not publish the sampling capacitance — the ESP32-C3's does not — measure it (Lesson 1.7) or design so it does not matter.

KEY INSIGHTHow this connects

  • Voltage, Current, Resistance, and Power (1.2) — Ohm's law supplies the individual drops; KVL and KCL are what let you combine them across a whole circuit. The 1.85 µA divider current flagged there is the reason this lesson's resistors are 1 MΩ.
  • Passive Components for Real (1.4) — the 100 nF hold capacitor is a real part with a dielectric and a DC-bias curve; at 1.85 V an X7R 0603 keeps most of its value, but the same reasoning that sized it here is what fails when a 10 µF X5R loses 60 % under bias.
  • Measurement (1.7) — the opening scene's "the multimeter said 3.7 V" is a loading question too: a 10 MΩ meter on a 500 kΩ source reads 5 % low. Every instrument is a load, and Thevenin tells you how much.
  • Beacon's battery sense on the schematic (2.2) — the circuit designed here is drawn, with the PMOS switch and its gate drive, on Beacon's Stage 2 schematic.

KVL, KCL, and Thevenin are the entire toolkit for hand analysis in this course. You will not need node-voltage matrices or SPICE for anything on Beacon's board. What you will need, constantly, is the reflex of drawing the loop, naming the node, and asking "what is the Thevenin resistance here, and what is connected to it?" That reflex is worth more than any simulator, because it runs before the schematic is finished.

GLOSSARYKey terms introduced

Kirchhoff's voltage law (KVL)
The algebraic sum of voltages around any closed loop is zero. A statement of energy conservation; used to find unknown drops and to expose elements not on the schematic.
Kirchhoff's current law (KCL)
The algebraic sum of currents into any node is zero. A statement of charge conservation; at a divider's midpoint it explains loading in one line.
Voltage divider
Two series resistors across a source with the output at the midpoint: Vout = Vin · R2/(R1+R2), valid only with no load. The most used and most misused circuit in embedded design.
Thevenin equivalent
Any linear two-terminal network reduced to a voltage source Vth (open-circuit voltage) in series with Rth (resistance with sources zeroed). For a divider, Rth = R1 ∥ R2.
Loading
The change in a source's output voltage when a load draws current through its Rth: VL = Vth · RL/(Rth + RL). Under 1 % error needs RL ≳ 100 · Rth.
Sampling capacitor
The small (pF-class) capacitor a SAR ADC switches onto its input during acquisition. It must charge through the source's Rth within the acquisition window, or the sample reads low.
Hold capacitor
A capacitor placed across a high-impedance divider's output to supply the ADC's sampling charge locally. Fixes loading at the cost of a settling time Rth · Chold that firmware must respect.
Time constant (τ)
R · C for a first-order RC network; the node reaches 63 % of its final value after τ, 99 % after ≈ 5τ. Beacon's divider: 500 kΩ × 100 nF = 50 ms.

ACTIVE RECALLCheck your understanding

Answer in your own words first, then reveal.

1A 4.7 kΩ / 10 kΩ divider from 5 V feeds a comparator whose input looks like 47 kΩ to ground. What does the comparator see, and by how much is it off from the open-circuit value?›

Open-circuit: Vth = 5 × 10 / 14.7 = 3.40 V. Rth = 4.7 ∥ 10 = 3.20 kΩ. Loaded: VL = 3.40 × 47 / (3.2 + 47) = 3.18 V — about 6.4 % low. Whether that matters depends on the comparator's threshold and the margin the design needs; the point is that you can know it before building anything. If 6 % is too much, either raise the load (a different comparator or a buffer) or lower the divider — the latter at a cost of 5 V / 14.7 kΩ = 340 µA, which is fine on a mains-powered board and unacceptable on Beacon.

2Beacon's 1 MΩ divider reads about 33 % of the correct value into a bare ADC input with a 2 µs acquisition window and a 10 pF sampling capacitor. Show where that number comes from, and explain why a 100 nF hold capacitor fixes it without changing the resistors.›

Rth = 1 MΩ ∥ 1 MΩ = 500 kΩ; τ = 500 kΩ × 10 pF = 5 µs. In a 2 µs window starting from an empty sampling capacitor the node reaches 1 − e−2/5 = 1 − 0.67 = 0.33 of the true voltage. The reading is not noisy or drifting; it is systematically one third, every time. A 100 nF capacitor across R2 holds 10,000 times the charge of the sampling capacitor, so each acquisition draws its charge from the hold capacitor and moves the node by about 0.01 %. The 500 kΩ no longer participates during the sample. It participates before: the hold capacitor must charge through it, τ = 50 ms, so the firmware waits ~250 ms after enabling the divider. The resistors are unchanged, the sleep current is unchanged, and the cost is time.

3A colleague argues that since the divider is switched by a PMOS anyway, 1 MΩ is pointless and 22 kΩ would let you drop the capacitor and the wait. What is the strongest counter-argument, and what is the honest case for their side?›

Counter-argument: the resistor value is insurance against the switch. A PMOS's gate must be driven to the cell voltage to turn it off; in deep sleep the driving GPIO's state is defined only if the firmware holds it, and any mistake, leak, or ESD-damaged gate leaves the divider connected. With 22 kΩ that failure costs 3.7 / 44 kΩ = 84 µA — ten times the sleep budget, the product dead in the field. With 1 MΩ it costs 1.85 µA, which Rev A actually suffered and survived. Defense in depth is worth 250 ms a minute. Honest case for 22 kΩ: the wait and the capacitor are real complexity; the wait must be hidden in light sleep and the firmware ordering must be right; and a well-designed switch with a verified gate drive is not exotic. On a mains-powered or short-life product the colleague is right. On a six-month battery product with 25 unattended units, the resistor value is cheaper than the site visit.

4The field data shows one Beacon unit's battery reading drops by 40 mV only during transmissions. Use KVL to name the likely cause, estimate the shared resistance, and say what layout change fixes it.›

The battery ADC's ground reference and the radio's return current share a conductor. During the burst, the current through that shared path develops a voltage (KVL: the ADC's "ground" is no longer at the cell's negative terminal), and the ADC reads the divider relative to a lifted reference. If the burst peak is around 100 mA (an illustrative figure — the ESP32-C3's BLE TX peak is higher than its 40 mA window average), a 40 mV offset implies R = 0.04 / 0.1 = 0.4 Ω of shared path — far too much for a plane, plausible for a thin trace, a via, or a connector pin shared by both currents. The fix is not in firmware: give the ADC and divider a return that does not carry the radio's current (a solid ground plane with the divider's ground referenced near the ADC, and the module's ground via directly into the plane), or, as a firmware mitigation, never sample the battery during a transmission. Stage 5 turns this into layout rules; Lesson 3.7's state machine orders the sample before the burst.

Next: a 10 µF capacitor is not 10 µF — passives with their fine print →