# GENERATORS — measurement.md

**Theme.** Operations derived from measurement itself: recipes that take a measurement problem you
already have and move it into a different part of the act of measuring — the standard it rests on,
the blank it subtracts, the threshold it crosses, the frame it stands in, the null it balances
against, the loop it traces.

### G001 — Find the load-bearing artefact
**Operation.** Draw the calibration chain for your quantity, from your reading back to the realisation of the unit. Mark every node that feeds more than one branch. At each such node ask two questions: is there a second, independent realisation of this node, and what property of the artefact or standard is assumed constant but never measured? Choose the node with the most descendants and no independent check, and measure the assumed property directly.
**Produces.** An experiment on the standard rather than on the sample; usually a systematic shared by a whole family of measurements, and therefore invisible in any single one of them.
**Worked example.** A laboratory's force scale traces every calibration to one set of stainless-steel mass pieces, whose mass is assumed constant. Weighing them against each other cannot detect a common change. Measure their buoyancy-corrected volume by hydrostatic weighing, and their surface layer by repeated weighings in dry and in humid gas. The output is a curve of adsorbed layer mass against humidity — a quantity that sets a shared floor under every force measurement in the building.
**Fails when.** The chain is short, or independent realisations already exist, so the node is not load-bearing and the extra work measures nothing new.

### G002 — Compute what the instrument cannot see
**Operation.** Write the measurement as an operator: which functional of the sample does the reading respond to? Enumerate its null space — the sample configurations that leave the reading unchanged. Sort them by whether your target physics lives there. If it does, modulate the sample along the null direction (translate it, rotate it, change the mode) and look for a response that appears only then.
**Produces.** Either a new channel, in which the blind direction becomes an observable, or a null test that separates the target from the dominant contaminant.
**Worked example.** A cavity ring-down cell reports loss at its mode frequencies. Its null space contains any absorber sitting exactly at a node of the standing wave, and any loss that is not mode-selective, such as scattering from the mirrors. Put the sample on a translation stage and record ring-down time against position. Gas absorption is flat in position; a film on the sample is not; mirror loss is flat in position but changes only when the mirror is touched. Three flat-or-not patterns separate three terms that one ring-down number adds together.
**Fails when.** The instrument is broadband and its null space is trivial, or the signal is below noise in every direction, so no modulation recovers it.

### G003 — Keep the phase, not just the amplitude
**Operation.** Whenever you modulate an input and read a response, record the complex response rather than its magnitude. Ask which sample properties set the amplitude and which set the phase. Then arrange the experiment so the quantity you want appears in the phase, where source power, collection efficiency and detector gain all cancel as multiplicative factors. Report both parts as two measurements with independent systematics.
**Produces.** A measurand immune to the multiplicative systematics that normally dominate, and often a second, independent determination of a quantity previously measured once.
**Worked example.** The thermal diffusivity of a coating is usually extracted from the amplitude of a thermal-wave signal, which requires knowing the absorbed laser power and the infrared detector's responsivity. The phase lag between a modulated heating laser and the surface temperature depends only on modulation frequency and diffusivity. Sweep frequency, fit the phase, and the power calibration drops out; the amplitude fit then becomes an independent check instead of the measurement itself.
**Fails when.** The phase is dominated by a parasitic lag in the electronics or detector that cannot be characterised, or the response is nonlinear so amplitude and phase do not separate.

### G004 — Make the blank the sample
**Operation.** List everything the pipeline subtracts: the dark frame, the empty cell, the solvent, the background spectrum, the no-sample run. Choose the subtraction assumed perfect. Then measure that blank as a function of time, temperature, fill history, and the sample that preceded it. Treat its variation as a signal with a mechanism, not as noise to be averaged down.
**Produces.** A contamination, carry-over or surface observable, plus a hard bound on how much of any real signal could be blank.
**Worked example.** A spectrophotometer records a cuvette blank once per session. Fill the same cuvette with pure solvent and read absorbance every minute for an hour, then empty and refill it repeatedly. The drift and the refill-to-refill steps are not instrument noise: they are the wall adsorbing and desorbing the analyte, and the step height after a refill gives a wall mass-transfer coefficient. That coefficient sets how long a low-concentration sample must sit before its reading means anything.
**Fails when.** The blank is genuinely empty, or blank and sample cannot be separated in time, so the variation carries no mechanism you can attribute.

### G005 — De-binarize every decision in the pipeline
**Operation.** Walk the signal path and mark every place where a continuous variable becomes yes or no: a discriminator, a threshold, a peak picker, a pass-fail specification, an accept flag. For each, record the continuous variable that fed it, even where the pipeline discards it. Then ask what distribution that variable has, and whether the threshold sits where the distribution's shape carries information about the sample rather than about the detector.
**Produces.** Detectors converted into spectrometers; a continuous channel recovered from an instrument that reports only counts.
**Worked example.** A proportional counter feeds a discriminator and reports counts. The pulse-height distribution below threshold is normally thrown away. Record it: its shape is the avalanche size distribution, whose mean and width depend on the gas gain and on where in the counter the primary ionisation occurred. The counter becomes a crude position-sensitive detector, and the gas gain becomes a measured quantity that can be tracked as the fill ages.
**Fails when.** The continuous variable is destroyed before it becomes accessible, or the distribution is set entirely by detector statistics and holds nothing about the sample.

### G006 — Turn the disagreement into the measurand
**Operation.** When two instruments, methods or laboratories report different values for nominally one quantity, resist deciding which is right. Define the difference, give it units, an error budget and controlled variables. Plot it against every parameter you can vary: rate, temperature, position, concentration, time since calibration. The functional form of the difference is a measurement of whatever distinguishes the two methods.
**Produces.** A new observable built from a known inconsistency, usually a systematic of both methods expressed as a physical property.
**Worked example.** Two thermometers in the same bath differ by a few millikelvin. Instead of calibrating one against the other, mount both on a translation stage and move them through a thermal gradient you have not characterised. The difference becomes a function of separation: its slope is the gradient at the scale of the sensor tips, and its intercept is the pair's intrinsic offset. Neither number is available from either thermometer alone.
**Fails when.** The two methods share the systematic, so the difference is blind to it, or the difference is smaller than the reproducibility of either method.

### G007 — Drive the detector into its nonlinearity
**Operation.** Identify the element that saturates: the amplifier, the photodiode, the ADC, the sensor's dynamic range. Measurements normally stop short of it. Instead, apply a step or a pulse that drives it well past saturation, then record the recovery at full bandwidth. Fit the recovery to the limiting mechanism and ask which material or device parameter sets its time constant.
**Produces.** A material or device parameter — lifetime, trap density, thermal time constant, compliance — measured by hardware that normally reports only intensity.
**Worked example.** A photodiode in a laser power monitor is kept in its linear range. Overdrive it with a short pulse and record the tail of the photocurrent as it returns. The decay contains the diffusion of carriers generated outside the depletion region; its shape against reverse bias gives a carrier lifetime and a diffusion length, both of which change with radiation damage. The power monitor doubles as a degradation sensor.
**Fails when.** Saturation destroys the information rather than encoding it, as in latch-up or a blown fuse, or the recovery is set by external circuitry rather than by the device.

### G008 — Ask which frame the sensor is standing in
**Operation.** For every probe, ask what the sample is doing at the probe's location: moving, accelerating, rotating, expanding, reacting. A sensor at rest in the laboratory measures a different quantity from one moving with the sample, and the conversion is usually a known physical relation. Move the probe, or use two probes in different frames, and treat the conversion factor as the measurement.
**Produces.** A quantity that was a correction becomes an observable, and probe geometry becomes a measuring principle.
**Worked example.** A thermometer in a fast gas stream reads stagnation temperature, not static temperature, because the gas decelerates at the probe. Two identical probes, one blunt and one in a streamlined housing that lets the flow pass, read different temperatures in the same stream. The difference depends on Mach number alone, so the pair measures flow speed without a Pitot tube, and the stagnation correction that labs apply by hand becomes a calibrated channel.
**Fails when.** The frame correction is smaller than the noise, or the sample has no well-defined frame because the region is turbulent or far from equilibrium.

### G009 — Extrapolate the probe to zero
**Operation.** Take any measurement in which probing changes the sample: beam damage, heating, depletion, charging, mechanical load. Measure the target at several probe intensities, doses or forces, always on fresh material, and fit against the probe level. The intercept is the unperturbed value; the slope is the sample's susceptibility to the probe. Report both, then ask which of the two your question actually needs.
**Produces.** Two numbers from one experiment — a perturbation-free value, and a cross-section describing how the probe couples to the sample.
**Worked example.** An electron microscope's beam changes the specimen while imaging it. Measure a lattice parameter, a composition ratio or a defect density at a series of beam currents and dwell times. The intercept is the pristine value; the slope is a damage cross-section per incident electron. That cross-section is a material property, it predicts the maximum dose any later experiment can use, and it is normally discarded as beam damage.
**Fails when.** The coupling is nonlinear so the extrapolation is not a line, or the damage changes mechanism with dose and the slope means different things at different doses.

### G010 — Put the instrument inside the circuit
**Operation.** Ask what the sensor does to the system it measures: its stiffness, thermal mass, input impedance, optical absorption. Write the measured value as a function of the instrument's own parameter. Then vary that parameter deliberately — swap cantilevers, add a series resistor, change the probe's heat capacity — and fit the measured value against it. The true sample value is the limit where the instrument's influence vanishes or diverges.
**Produces.** A de-embedded sample property plus a direct measurement of the instrument's loading coefficient, which is normally assumed rather than known.
**Worked example.** A contact modulus measured with a force sensor is pulled toward the sensor's own stiffness: the reading is a series combination of sample and instrument compliance. Measure the same spot with cantilevers of several known stiffnesses. Plot measured compliance against cantilever compliance; the intercept is the sample's compliance, and the slope shows how much of the indentation depth was instrument deflection. The same move applies to a thermometer with too much heat capacity and a voltmeter with too low an input impedance.
**Fails when.** Loading is negligible, so the correction sits below the noise, or several loading mechanisms act at once and cannot be varied independently.

### G011 — Null it
**Operation.** Convert the measurement from a deflection into a balance. Instead of reading how far an indicator moves, apply a known opposing influence until the indicator reads zero, and report the opposing influence. The indicator then only has to be sensitive near zero, not linear or calibrated across a range. Ask what known influence your laboratory can trade against the measurand.
**Produces.** A measurement traceable through the balancing quantity rather than through the sensor, and a null detector with a far smaller range and better linearity.
**Worked example.** A capacitance diaphragm gauge infers pressure from membrane deflection. Instead, apply an electrostatic pressure with the same electrodes and adjust the voltage until the membrane returns to its zero-pressure capacitance. Pressure is now read from voltage and electrode geometry; the membrane's stiffness and creep drop out of the result, and the gauge becomes a null detector whose only job is to hold a capacitance constant.
**Fails when.** No known opposing influence of the right size exists, or the null detector's zero is unstable, so the balance merely relocates the calibration problem.

### G012 — Let the sample report on itself
**Operation.** Ask which of the sample's own outputs depends on an internal state you cannot reach: its voltage, resistance, resonance, colour, acoustic emission. Find a known coefficient linking that output to the state. Then perturb the state deliberately — a current step, a light pulse, a pressure step — and read the sample's own response. The sample becomes the sensor and the instrument only records.
**Produces.** Internal-state measurements in places no sensor fits, plus a cross-check on external sensors separated from the sample by a thermal or electrical boundary.
**Worked example.** A shape-memory alloy wire is driven by current, and its resistance depends on both temperature and the martensite fraction. During a slow actuation cycle, record resistance and current. The fast component tracks temperature through the temperature coefficient, the slow component tracks phase fraction. The wire reports its own transformation progress — otherwise measured by diffraction — and the controller needs no extra sensor.
**Fails when.** The internal state and the readable output are not linked by a stable coefficient, or the sample responds more slowly than the state changes and reports an average instead.

### G013 — Swap the measurand for its exact conjugate
**Operation.** Look for an exact relation — a thermodynamic identity, a Maxwell relation, reciprocity, an Onsager relation — connecting the quantity you cannot measure to one you can. Write the relation, identify which derivative corresponds to the hard quantity, and design an experiment that takes that derivative along the easy axis. You are not approximating; you are choosing a different derivative.
**Produces.** Hard quantities — entropy, heat capacity at constant volume, chemical potential — replaced by easy ones with an exact link and entirely different systematics.
**Worked example.** Heat capacity at constant volume is hard to measure in a solid: it requires containing the sample without letting it expand. The identity relating the two heat capacities to the thermal expansion coefficient and the isothermal compressibility converts the problem. Measure expansion with a strain gauge and compressibility in a pressure vessel, combine with a conventional constant-pressure measurement, and the constant-volume value follows exactly, together with the stored elastic energy.
**Fails when.** The relation holds but the required derivative is harder to measure than the original, or the sample is out of equilibrium and the identity does not apply.

### G014 — Read the whole response matrix
**Operation.** A transducer responds to its intended quantity, and also to temperature, strain, humidity, light, magnetic field, orientation. Instead of treating these as cross-sensitivity to be cancelled, measure the full matrix: vary each environmental input independently and record the output. Then either invert the matrix to separate the inputs, or keep the extra channels deliberately as additional sensors.
**Produces.** Multi-parameter sensing from one device, and a quantitative ranking of which cross-sensitivity actually limits the intended measurement.
**Worked example.** A piezoresistive pressure sensor responds to pressure and temperature, and the two coefficients differ between the active resistor and its reference. Drive the sensor's temperature with a heater at constant pressure, then repeat at several pressures. The resulting surface gives the pressure sensitivity and the temperature coefficient, and combining the two elements' signals separates the inputs. The chip now reports pressure and temperature, and the compensation network becomes a calibrated thermometer.
**Fails when.** The cross-sensitivities move together with the measurand, making the matrix singular, or the extra channels are too noisy to resolve anything.

### G015 — Measure the loop, not the point
**Operation.** Take any quantity measured by sweeping an input up and back down: pressure, field, temperature, strain, concentration. Plot the two branches against each other. The width, area and shape of the loop are separate observables, each with a mechanism — loss, pinning, relaxation, a phase transition. Fit the loop rather than its midpoint, and check how its parameters scale with sweep rate.
**Produces.** A dissipation or memory quantity measured by the same apparatus that normally reports only a static value, and a way to separate elastic from anelastic response.
**Worked example.** A diaphragm gauge is calibrated by comparing applied pressure with indicated pressure. Sweep pressure up and down through the same range and plot indicated against applied. The loop's half-width is the diaphragm's mechanical hysteresis; its growth with sweep rate separates viscoelastic creep from plastic set; the centreline is the calibration curve the laboratory actually uses. A gauge that looked like a simple sensor now also reports its own anelastic loss, which predicts how far it will drift under a static load.
**Fails when.** The loop is dominated by the instrument's own hysteresis rather than the sample's, or the sweep is slow enough that the loop collapses onto a line.