End-to-end radiation sensing
and $GM telemetry

System Overview and Operating Principle

An SBM-21 Geiger–Müller tube sits at ≈400 V, a bias the board manufactures from its own 3.3 V rail with an ESP32-driven boost stage and a one-stage voltage peak rectifier, closed-loop regulated through a 30 MΩ:R_fb divider into the ESP32-S3's ADC. When an ionizing particle crosses the tube, a Townsend avalanche, propagated along the whole anode wire by UV photons, dumps roughly a nanocoulomb through the tube; the cathode pickoff resistor turns that current into a positive 2–3 V pulse, which is AC-differentiated, clamped, and compared against 0.51 V by a TLV3201 to make one clean 3.3 V logic blip per event. The ESP32-S3 counts those blips in the PCNT hardware counter without interrupts or missed bursts, while a u-blox GNSS receiver supplies navigation data over I2C on XIAO D4/GPIO5 (SDA) and D5/GPIO6 (SCL). Once per second the firmware fuses count rate, an estimated dose rate, HV health, and the GNSS fix into one checksummed $GM sentence over USB-C, TTL UART, or BLE. The Pip-Boy 3000 Mk V (Espruino on an STM32) parses that sentence in a fixed buffer and draws a rad readout. This end result also explains why this document keeps repeating that a 21 mm tube with a ≈3 CPM background is a toy that teaches real physics, not a radiation-safety instrument.

The SBM-21 is a coaxial cylindrical counter: a thin anode wire of radius a ≈ 50 µm on the axis of a cathode of inner radius b ≈ 2.5 mm, filled with a noble gas plus a halogen quench fraction (Br₂/Cl₂) at ~100 torr. The field of that geometry is the whole story:

A charged particle crossing the gas leaves a track of primary ion pairs, on the order of tens to a few hundred pairs, essentially anywhere in the volume, almost all of it in the weak-field region. The electrons drift inward. Nothing multiplies until an electron falls inside the critical radius where the field exceeds the multiplication threshold E_c ≈ 10⁶ V/m:

Inside that skin, impact ionization runs away. The first Townsend coefficient, ionizations per electron per unit path, follows the semi-empirical form α/p = A exp(−Bp/E), so the multiplication per avalanche is M = exp(∫ α dr) over the last ~50 µm of fall. In a proportional counter the avalanche stays local and M is finite, so pulse height tracks primary ionization. The Geiger regime is different in kind, not degree: the avalanche is so dense that de-exciting gas atoms emit UV photons energetic enough to photoionize elsewhere. Those photons are not confined, they travel along the tube and seed new avalanches at other points on the wire, which seed more photons, until the entire anode length is sheathed in discharge. That is why a GM pulse's amplitude is independent of whether the particle made 5 primary pairs or 500: the tube is a trigger, not a spectrometer, and all energy information is destroyed by design.

The discharge terminates itself. Electrons are collected in nanoseconds, but the positive ions, ~1000× slower, remain as a space-charge sheath around the wire. That sheath screens the field below the multiplication threshold, and the avalanche starves. The tube is now blind: until the ion sheath drifts out to the cathode and the field recovers, a new particle produces no pulse (dead time), then undersized pulses (recovery time, the distinction between the two is whether the comparator threshold catches the runt pulse, which is why resolving time depends on your electronics, not just the tube). The transit sets the physics floor on the dead time:

The counter follows the non-paralyzable dead-time model. The measured rate can be corrected with N_true = N_obs/(1 − N_obsτ_d); at τ_d = 100 µs the correction reaches 1% at 6,000 CPM and 10% at 60,000 CPM. At background it is negligible. Sandbox 2 keeps the ion-clearing floor and the electronic R_aC recovery separate while plotting the loss.

Quenching: without help, ions arriving at the cathode would liberate secondary electrons and re-trigger the tube forever. Halogen molecules charge-exchange with the noble-gas ions (their ionization potential is lower), and the halogen ion dissipates its neutralization energy by dissociating, Br₂ → 2Br, rather than ejecting an electron. The atoms later recombine, so the quench agent is not consumed. Organic-quenched tubes (ethanol, etc.) crack their quencher permanently and die after ~10⁹ counts; halogen tubes self-heal, which is why a 1970s SBM-21 still works and why its plateau tolerates being run a little hot.

The SBM-21 was chosen for one reason, its provisional 21 mm envelope targets the available space, and it pays for that in sensitive volume. Computing it (all dimensions estimates for surplus tubes):

γ detection in any GM tube is mostly indirect: the photon interacts in the cathode wall and the ejected electron ionizes the gas. So sensitivity scales with cathode area and wall Z, and the tube's energy response is far from flat, over-reading soft γ by 2–5× relative to ¹³⁷Cs unless energy-compensated with a filter sleeve. A bare SBM-21 therefore measures exposure poorly: the ≈5 CPM/µSv·h⁻¹ factor is an order-of-magnitude anchor for ¹³⁷Cs geometry, and every µSv/h this document displays inherits that estimate. This is the first of the places where "hobby instrument, not a dosimeter" is a technical statement, not a disclaimer.

Mid-plateau (390–440 V here) the count rate changes by the plateau slope, a few percent per 100 V, so a supply that wanders ±10 V costs well under 1% in rate. Below the starting voltage (≈360 V) avalanches fail to become Geiger discharges and efficiency collapses; above the plateau knee, spontaneous avalanches from field emission and imperfect quenching make the tube count itself, and past ≈470 V it latches into continuous discharge, which shortens tube life in seconds-to-minutes, not hours. This is why the regulation loop in chapter 3 exists at all.

Surplus-tube dimensions, plateau limits, and sensitivity vary by suffix, lot, and source. The swap model therefore treats cross-tube values as comparative estimates; use the datasheet for the exact part in hand before changing the enclosure, bias, or calibration.

TP1 (anode, 1:100 HV probe only): a flat 400 V rail that dips ~50–100 V at each event and recovers with τ = R_a·C ≈ 25 µs. The dip amplitude is set by the discharge charge against the node capacitance (ΔV = Q/C ≈ 0.4 nC/5.3 pF ≈ 75 V, a lumped-C simplification that holds until you add probe capacitance, which is itself ~15 pF and visibly slows the recovery: the act of probing TP1 changes TP1). Open the scope dock and select TP1 while the master field is set hot.

The boost stage is an energy shuttle. While Q1 conducts, the inductor integrates current from the 3.3 V rail (v = L di/dt); when Q1 opens, that current has nowhere to go and the collapsing field drives the switch node as high as the load network allows, the flyback. Per cycle:

The C2/D1/D2 network is an AC-coupled peak rectifier. A switch waveform from 0 to 200 V produces approximately 199 V in the topology check; it does not produce 400 V. Q1 must withstand the full output voltage plus overshoot. The revised board uses a 600 V transistor. The converter still needs transistor switching, thermal, leakage and regulation measurements before assembly release.

Q1: Nexperia PBHV8560Z, 600 V, SOT223. Pin 1 is base, pin 2 and the tab are collector, and pin 3 is emitter. Replace the purchased MMBTA42; its 300 V rating is insufficient. RB1 is now 100 Ω, with a new 100 kΩ base-emitter pull-down RBE1. Nominal base drive is about 24 mA during a pulse; GPIO drive strength and turn-off storage require measurement. The ideal-switch feasibility model uses 5 kHz and a maximum 70 µs pulse, with cycle skipping near 400 V. Those are model inputs, not validated firmware settings.

D1/D2: purchased Vishay US1M-E3/61T, 1 kV. Reverse recovery, junction capacitance and hot leakage materially affect this low-current converter. The 1 mH SRR1260-102K remains fitted. At 3.3 V and 70 µs, the ideal ramp is 231 mA; the feasibility model including winding resistance reaches about 215 mA. Verify peak current, saturation margin and Q1 switching loss on the prototype.

Ripple at this load is not the classic sawtooth: between counts the rail droops only through the 30 MΩ divider (τ = 141 ms), and each Geiger discharge removes ΔV = Q/C_hv = 0.4 nC/4.7 nF ≈ 0.09 V. The supply is a trickle charger topping up a bucket, and the loop's real job is rejecting temperature and battery-voltage drift, not load transients.

TP5 (switch node, 1:100 only): flat near 0 V during t_on, then the flyback edge, a peak above 400 V with L·C_par ringing at a few MHz and decaying over ~1 µs, repeating at f_pwm. TP8: 400 V DC with the load-dependent sawtooth of the burst regulator. If TP5's peak flattens and the input current jumps, you are watching L1 saturate.

The divider is 30 MΩ over 68 kΩ: 400 V produces 0.905 V and draws 13.3 µA (5.32 mW). C4 is now placed beside the XIAO ADC pad. Its 10 nF value gives about 0.68 ms filtering with the divider source resistance. Use calibrated ADC attenuation that can also measure an overvoltage, rather than saturating near the normal operating point. The control loop and sampling timing require hardware validation.

Firmware-controlled high voltage needs explicit fault handling: start with PWM disabled; use bounded, finite pulse bursts; stop on stale or implausible ADC data, excessive voltage, missing output rise, or watchdog reset. A maximum duty cycle alone is not overvoltage protection. This PCB has no independent hardware overvoltage shutdown. The animations below are qualitative teaching models and must not be copied as control firmware or used to justify a component rating.

TP6: ≈0.905 V DC with small ripple. Without the 10 nF reservoir, each ADC acquisition charge-shares with the SAR sampling capacitor and yanks the high-impedance node down before it recovers through ~68 kΩ, a comb of droop spikes at the loop tick rate. With the reservoir fitted, the acquisition transients stay local and the ADC measures the divider instead of its own history.

During the discharge, the same current flows through the anode load R_a (from HV+) and the cathode resistor R_k (to ground). Picking off at the cathode is the trick that makes a single-supply readout clean: the node idles at exactly 0 V, and the event appears as a positive-going pulse of V = i(t)·R_k, R_k is a transimpedance, ~2–3 V peak for the ~0.4 nC discharge. Anode pickoff (Rev A) gives the same information riding on 400 V with awkward polarity, workable, but every downstream choice fights it.

C_c(R_k+R_s) is a high-pass, equivalently, a differentiator for signals slower than its corner:

Only the fast leading edge crosses; the slow tail is differentiated away and R_s restores the baseline to 0 V. Because the coupled waveform must integrate to zero, every pulse is followed by an equal-area undershoot, small at 9.4 µs, but grow τ and the undershoot of one pulse starts eating the amplitude of the next (pile-up); shrink τ and you shave the pulse amplitude itself. The parasitic capacitance on the sense node (trace + clamp + comparator input, ~10 pF) is the real limit: it forms a capacitive divider with C_c, which is why C_c can't just be made tiny "for speed."

The BAT54S clamp earns its place at the extremes: the first edge of a large event can punch past the rails, and the TLV3201's absolute-maximum input is V_cc+0.3 V. A Schottky's ~0.3 V forward drop clamps inside that window where a silicon diode's 0.7 V would not, the forward voltage is the specification here, not an incidental property.

R_a = 4.5 MΩ (three CRCW12061M50FKEA 1206 resistors in series, about 133 V each at 400 V versus a 200 V working rating) limits the avalanche current so the space-charge quench can win: too small and the supply refeeds the discharge into continuous glow; too large and the R_a·C recovery stretches past the ion-clearing time and becomes the dead time. R_k = 100 kΩ is a compromise between pulse amplitude and the recovery it adds on the cathode side. τ_hp = 9.4 µs remains below the 25 µs anode recovery, long enough to preserve the leading edge and short enough that consecutive resolvable pulses don't stack.

TP2 (1:10): the raw positive cathode pulse, 2–3 V, decaying over tens of µs. TP3 (1:10): the same event after C_c, a sharp ~2 V spike collapsing with τ = 9.4 µs and a shallow undershoot. If TP3 shows a comb of narrow spikes with no tube events, that is the switcher coupling into the sense node, chapter 5's threshold and chapter 7's layout are the two defenses.

The comparator is the analog/digital firewall: on one side, a continuous voltage carrying tube physics, switcher spikes, and thermal noise; on the other, edges that either exist or don't. The threshold divider sets the decision boundary, and the hysteresis network makes it two boundaries:

When the output snaps high, R_h lifts the sense node's baseline by 0.30 V, so the comparator holds until the shaped pulse has decayed well below where it tripped. Two consequences: the output cannot chatter on the pulse's noisy falling edge (one particle, one edge, the PCNT requirement), and the output width is stretched to ~20 µs, comfortably above any glitch-filter setting. The width is not a datasheet number, it falls out of τ_hp and the hysteresis, which is why sandboxes 3 and 5 both move it.

The TLV3201's push-pull rail-to-rail output drives the PCNT pin directly, no pull-up, and 40 ns propagation instead of the LM393's µs-class, open-drain rise through 10 k. At GM count rates the speed is irrelevant; the edge quality is not, because PCNT counts edges, and a slow RC rise through a noisy region can double-count without hysteresis. V_th at 0.51 V sits ~4× above the shaped switcher residue and ~4× below the pulse peak, the point of maximum separation between the two distributions, which the ROC below makes visible.

TP4 (1:10): idle low, one clean 3.3 V pulse of ~10 µs per event. Trigger on TP3 and overlay TP4 to see the trip point and the hysteresis-held release directly, that overlay is the waveform river below, automated.

The PCNT peripheral is a hardware edge counter clocked from the APB bus: the comparator's rising edge increments a 16-bit register with zero CPU involvement and zero per-event latency. The alternative, a GPIO interrupt, costs ~2–4 µs of ISR entry/exit per event and, worse, serializes events through the interrupt controller: two pulses inside the ISR window merge into one. At background rates either works; in a burst (bringing a check source close, or the tube's own double-pulsing when mis-biased) the interrupt path undercounts precisely when the reading matters. The PCNT glitch filter rejects edges shorter than N APB cycles (up to 1023 × 12.5 ns ≈ 12.8 µs), set it to ~1 µs: an order below the 10 µs comparator pulse, an order above coupled switcher spikes.

The count line lands on D3/GPIO4 because D2/GPIO3 is a strapping pin, sampled at reset to select boot behavior; a comparator that happens to hold it during a brownout reboot bricks the boot in a way that looks exactly like a firmware bug. And HV feedback must use an ADC1 channel (A1/GPIO2): ADC2 is arbitrated against the Wi-Fi radio's internal use and returns garbage or errors while the radio is active, a design that works on the bench and fails the moment telemetry is enabled, which is the most expensive kind of bug to find.

Move the count line to an interrupt: nothing visibly changes until the day it does, a burst undercounts, silently, with no error path. Skip the glitch filter: every switcher spike that survives layout becomes a count, and your "background" tracks duty cycle. Put feedback on ADC2: HV regulation dies when Wi-Fi transmits, the duty clamp catches it, and the rail sags, visible as hv_ok dropping in the $GM stream exactly during uploads.

Nothing, that is the point. After TP4 the signal exists as register state. The observable is the $GM sentence's CPM field, and the bring-up check is statistical (chapter 10), not oscillographic.

The GNSS receiver does all the satellite work and exposes its navigation stream through the MAX-M10S I2C-compatible DDC interface at 7-bit address 0x42. XIAO D4/GPIO5 is SDA and D5/GPIO6 is SCL; D6/GPIO43 and D7/GPIO44 remain available for the external host UART. Firmware reads the GNSS UTC and fix data over I2C and timestamps count bins from the ESP32's local timer. The threat runs the other way: the HV switcher is a deliberate generator of fast edges, and a GNSS front-end is a −130 dBm receiver.

Near-field coupling, not far-field radiation, is the mechanism at these sizes. The switch node swings ~400 V in an illustrative 100 ns edge, dV/dt ≈ 4 kV/µs, and couples capacitively (E-field): every pF between the node and a victim trace injects i = C·dV/dt ≈ 4 mA per pF. The input-current loop (C₈/C₉→L1→Q1→ground) carries a pulsed current of approximately 215 mA in the feasibility model and couples inductively (H-field) as a magnetic dipole whose near field falls as 1/r³, the loop's enclosed area is the knob you control in layout. The harmonics of a 5 kHz square edge extend into the tens of MHz, and while none of that reaches 1.5 GHz directly, it desensitizes the receiver through its LNA bias and the antenna's ground reference. Distance beats shielding at these geometries: the far corner of a 66 mm board buys ~(4×)³ ≈ 60× less H-field than adjacent placement, and the external u.FL antenna moves the aperture off the board entirely.

Compact, reversible tube mounting: PocketMagic specifies a 16 × 6 mm can, 5 × 2 mm pin and 21 mm overall length. The Littelfuse 01000058Z body clip is 7.11 mm long and 5.84 mm tall. The Keystone 6049 pin socket is 5.283 mm long and 5.156 mm tall. MC1 and MC2 have identical footprints: four plated holes in a cross, 5.08 mm between opposite holes, 1.4 mm finished drill and 2.2 mm solder lands. The clip uses the transverse pair; the socket uses the axial pair. Populate one contact per end.

To reverse the tube: swap the clip and socket between the two ends, rotate each 180°, and change the two JP1 links on the underside. Always use the numbered pads; the back view is mirrored. The cathode can connects to the pulse pickoff, not ground.

Disconnect power and discharge before changing parts or links. Remove both old links first; fit exactly one configuration and check electrode continuity before powering up. Sample fit is still required: the clip is specified for 6.3/6.35 mm fuses; the socket for a nominal 2.03 mm pin. Check grip on the 6 mm can, actual pin diameter, engagement, seating height and assembled insulation in both orientations. Dimensions and assembly notes · Contact BOM · Mounting data.

Swap the GNSS to the HV corner and fixes drop every time hv_ok is true. Route the sense node past the switcher and sandbox 4's false-count axis stops being hypothetical. Pour ground under the flyback node and the added ~pF slows the edge (good) while coupling it straight into the plane (bad), the reason the pour is pulled back there.

TP9 SDA and TP10 SCL (both 1:10): idle high; SDA falls while SCL is high for START, the 0x42 address shifts MSB-first, and the receiver holds SDA low for ACK on the ninth clock. At 400 kHz SCL is 2.5 µs per cycle. Confirm clean 3.3 V logic levels and measure the rising edges; RPU1/RPU2 now provide 4.7 kΩ pull-ups to 3.3 V. Their rise time with the assembled bus capacitance must meet the selected I2C timing mode. The RF problem is not directly visible on an ordinary probe trace; quantify it by comparing MAX-M10S C/N₀, time-to-first-fix, and fix retention with the HV converter disabled and enabled.

UART framing is voltage-level time-division: idle high; a start bit (low) arms the receiver's mid-bit sampler; eight data bits LSB-first; a stop bit (high) guarantees a re-armable edge. At 115200 baud one bit is 1/115200 = 8.68 µs, and both ends' oscillators must agree within ~2% over ten bits, trivial for crystals, the classic failure for RC-oscillator MCUs. The same bytes over USB-CDC ride a 12 Mbit/s packet protocol where framing is the host controller's problem; over BLE they ride notifications with ~30–100 ms latency. The sentence is NMEA-shaped on purpose: off-the-shelf GPS software will pass $GM through unharmed, and the raw $GNGGA can be forwarded alongside so the host sees a normal receiver plus one extra sentence type.

One fused line per second, produced by the ESP32 as the hub, the tube on a counter pin, the GNSS on I2C, one merged stream out over the dedicated host UART. The XOR checksum is weak as error detection (single-bit errors only, no burst protection) but it is free, it is what NMEA parsers already speak, and on a 15 cm board-to-board wire the error model is "connector fell off," not bit rot. The µSv/h field carries the calibration estimate with it, any consumer of this stream inherits the ±order-of-magnitude uncertainty of chapter 1, which is why hv_ok and raw CPM are transmitted too: CPM is a measurement, µSv/h is an interpretation.

Raise the report rate to 10 Hz and each frame carries 0.005 counts at background, pure Poisson noise. Drop the checksum and the Pip-Boy renders garbage the first time the connector half-seats. Use String.split(',') in the parser and the Espruino heap fragments. Chapter 9 is built around avoiding that failure.

TP7 (1:10): idle high, 8N1 frames at 8.68 µs/bit in once-per-second bursts ~5 ms long. FIG 8.1 is that trace, decoded.

// $GM emitter, call once per second from the fuse/report task.
// PIPBOY_MODE trims the frame for a memory-poor parser: integer cpm,
// dose in nSv/h (integer), single flags field. Same checksum scheme.
bool PIPBOY_MODE = false;

static uint8_t nmea_ck(const char *s) {          // XOR between '$' and '*'
  uint8_t ck = 0;
  while (*s) ck ^= (uint8_t)*s++;
  return ck;
}

void emit_reading(uint32_t cpm, float usvph, bool hv_ok,
                  double lat, double lon, int alt_m,
                  int fix, int sats, uint32_t utc_hhmmss) {
  char body[96];
  if (PIPBOY_MODE) {
    // $GP,<cpm>,<nsvph>,<flags>  flags bit0=hv_ok bit1=fix
    snprintf(body, sizeof body, "GP,%lu,%lu,%d",
             (unsigned long)cpm, (unsigned long)(usvph * 1000.0f + 0.5f),
             (hv_ok ? 1 : 0) | (fix ? 2 : 0));
  } else if (fix) {
    snprintf(body, sizeof body, "GM,%lu,%.2f,%d,%.4f,%.4f,%d,%d,%d,%06lu",
             (unsigned long)cpm, usvph, hv_ok, lat, lon,
             alt_m, fix, sats, (unsigned long)utc_hhmmss);
  } else {                                        // empty position fields, per spec
    snprintf(body, sizeof body, "GM,%lu,%.2f,%d,,,,0,0,%06lu",
             (unsigned long)cpm, usvph, hv_ok, (unsigned long)utc_hhmmss);
  }
  char line[112];
  snprintf(line, sizeof line, "$%s*%02X\r\n", body, nmea_ck(body));
  Serial.print(line);    // USB-C CDC (bench)
  Serial0.print(line);   // UART0 TTL, 115200 8N1 → Pip-Boy
}

Design decisions, because they are the memory discipline: the NMEA parser is a fixed Uint8Array(96) filled in place, no split(','), no substring garbage; a number parser walks bytes and accumulates arithmetic instead of allocating strings; history is a pre-sized Float32Array ring; every listener and interval handle is kept so teardown actually detaches. Pins and calibration live in USER/PIPGEIG.JSON; an absent or invalid RX pin boots to an explicit setup screen rather than pretending to know the hardware.

Loading readable device source…

{
  "id": "pipgeig",
  "name": "RobCo Field Dosimeter",
  "version": "1.00",
  "files": "APPS/PIPGEIG.JS,APPINFO/PIPGEIG.info",
  "src": "APPS/PIPGEIG.JS",
  "icon": "APPINFO/HOLO.IMG"
}

1. Disconnect USB and every other power source. Wait for the high-voltage rail to discharge, then use an HV-rated meter to verify TP8 is below 5 V before touching or installing the board.
2. Check MC1/MC2 contact seating and JP1 polarity links: default 1–3 / 2–4; reversed 1–2 / 3–4. The can must reach GM_CATHODE and the isolated pin GM_ANODE. Remove both old links before changing orientation; verify there is no anode-to-cathode short.
3. Insulate the complete PCB from the enclosure and all metalwork. Secure all three 2.2 mm M2 holes with suitable insulating hardware while preserving the shown washer and copper keepout envelopes.
4. Build J2 with pin 1 GND, pin 2 GC1 TX to verified host RX, and pin 3 host TX to GC1 RX. Cross TX/RX, provide common ground, and use 3.3 V logic only; never apply 5 V.
5. Route the cable clear of the approximately 400 V section, tube contacts, RF antenna and u.FL cable, the lower-edge J2 opening, and all mounting washers. Do not touch HV copper, tube contacts, or components while energized.
6. Treat Pip-Boy pin assignments, enclosure fit, calibration, and RX command support as unverified. RX requires matching GC1 firmware. This is engineering hardware, not a certified dosimeter.

1. Use a current-limited bench supply for first power-up. Confirm the ESP32 answers ?ID with PWM held at zero duty and no HV generated.
2. 3V3 present at the comparator and GNSS. Still no HV.
3. TP5 with the 1:100 probe: switching present, flyback peak < 260 V, ringing decays in ~1 µs. This is where Q1's 300 V margin is verified or the snubber gets added (chapter 2).
4. TP8, 1:100: rail regulates at the target ±2%. The right half of the board is now live whenever powered.
5. TP6, 1:1: 0.905 V with the 68 kΩ feedback resistor, hv_ok = 1.
6. TP2 → TP3 → TP4 (1:10): background gives a pulse every ~20 s, patience or a check source (thoriated lantern mantle, uranium glass) speeds this up.
7. CPM rises with the source near, falls when removed; PCNT total matches a hand count of TP4 edges over a minute.
8. TP9 SDA and TP10 SCL with sky view: address 0x42 is acknowledged, logic levels and rise time meet the selected I2C mode, and ?FIX eventually reports.
9. One valid $GM per second at TP7/USB, checksum verified by an independent parser, the demo above counts.

Calibration procedure, in the order that respects the statistics: (1) log background for ≥1 hour; at 3 CPM that is ~180 counts, ±7.5%; (2) with a reference source or a borrowed calibrated meter, log ≥300 counts at a fixed geometry and solve CPM_PER_USVPH = (CPM−CPM_bg)/rate; (3) quote dose only from the smoothed average, and only for the γ energy you calibrated at, the SBM-21's response is not flat and uncompensated (chapter 1). Without step 2 the µSv/h field is a relative indicator. This is the second place "not a dosimeter" is engineering, not liability language.