Hand-considered hypotheses, 4901–5000
Written one at a time, not composed. This is the closing hundred of a planned set, restricted to production-line inspection and quality control as measurement science: in-line gauging, acceptance testing, gauge repeatability and reproducibility, defect detection thresholds, and traceability. Each entry carries its own logical-form label and no label is repeated, so the variety is structural rather than topical. None has been tested. No data are reported and no citation is made; where a magnitude is implied it is an estimate, not a measurement.
In-line gauging: geometry, fixturing, presentation
Misattribution
H4901. In-line laser triangulation usually blames its residual thickness variation on the strip, but at line speed the sensor integrates over a moving surface, so the reported profile is the true profile convolved with the sensor's exposure length. Running the same coil at two speeds separates the two: a real thickness feature keeps its amplitude, while an integration artefact changes shape with speed.
Hidden multiplier
H4902. Inside a caliper-style in-line gauge there is a hidden multiplier: the measurement axis is offset from the part's axis, so any angular error of the slide becomes a length error proportional to that offset. The multiplier is invisible in a calibration that presents the master at the same offset, and it appears only when the part is presented on the other side.
Angle-error bias
H4903. A gauge head fixed to the machine frame rather than to the part's normal reads a thickness multiplied by the cosine of the angle between them, a factor that is small at installation and grows as the frame's temperature tilts. Because frame and strip heat differently, the bias varies along a coil and imitates a thickness trend.
Clamp-defined measurand
H4904. For thin parts the in-line fixture's clamping force deforms the part more than the tolerance it is judged against, so the gauge reports the dimension under clamp rather than the free dimension. The free dimension is recoverable only when the fixture's stiffness and the part's local stiffness are both known, which makes the fixture part of the measurement model rather than part of the handling.
Calibration-material mismatch
H4905. Calibrated on a hardened master, a contact gauge reads compliant production parts small, because the contact flattens the part locally and the flattening depends on the part's elastic modulus. The bias is a material property leaking into a dimensional measurement, and it changes sign when a soft coating sits on a hard substrate rather than the other way round. The effect is largest on thin-walled parts, where local stiffness is lowest.
Tip-radius convolution
H4906. A scanning probe with a finite tip radius cannot enter a curvature tighter than its own, so a convex edge is reconstructed as an arc whose radius is the true radius plus the tip radius. The error is a constant offset rather than a scale factor, which makes it removable on a convex edge by subtracting the tip radius, while a concave feature tighter than the tip is not measured at all.
Support-condition mismatch
H4907. The same part measured on the line and on the acceptance bench differs because the bench supports are farther apart, letting gravity sag the part between them. The in-line gauge therefore reads a stiffer geometry than the bench, and the difference is a function of span rather than of either gauge's accuracy. Bench acceptance and line control then disagree permanently, and neither instrument is at fault.
Phase-dependent acceptance
H4908. In-line roundness is computed about a fitted centre, so a lobed part's diameter at two fixed gauge heads depends on the phase of the lobing relative to the heads. The same part can pass or fail according to how it was rotated when loaded, which turns the acceptance decision into a function of presentation rather than of geometry.
Contrast-slope bound
H4909. An optical edge finder that triggers at a fixed fraction of the intensity range places the edge where the surface's reflectivity crosses that fraction, not where the edge is. The bias grows with the edge's blur and with the distance between the trigger fraction and the half-intensity point, so it is largest exactly where the optics are least sharp and the surface is darkest.
Irrecoverable bias
H4910. For a shadow-casting gauge, diffraction pushes the apparent edge outward by an amount that grows with the square root of wavelength times standoff, and below some feature size the bias exceeds the tolerance. No post-processing recovers the true edge, because an intensity sensor never records the phase that would distinguish a diffracted edge from a smaller part.
Temporal aliasing
H4911. A gauge read out in step with the line's drive frequency can fold a structural vibration into a slow taper along the part, which at a single line speed is indistinguishable from a real taper. Changing the speed moves the alias frequency while leaving a real taper unchanged, so the speed change is the discriminating experiment.
Unfinished transient
H4912. A gauge powered down between shifts is still approaching thermal equilibrium when the first parts are measured, because its time constant is longer than the warm-up interval the line allows. The resulting bias is a deterministic curve rather than noise, so a warm-up check made at the start of the shift measures the transient and not the gauge.
Contact and tactile gauging
Mechanical filtering
H4913. A stylus tip mechanically filters the surface, and the filtered amplitude depends on the tip radius and the feature's spatial frequency, so two gauges with different tips disagree on the same surface by an amount predictable from their radii. The disagreement belongs to the pair rather than to either gauge, which means a surface-finish number is incomplete until the tip is stated.
Self-consistent bias
H4914. A contact gauge approaches from a fixed side, and its forward and return strokes differ by an amount proportional to the approach distance rather than to the displacement being measured. Calibration on a master approached in the same way cancels the bias from the calibration and leaves it in the part, so the gauge is most wrong where it looks best.
Size-independent dead band
H4915. A lever gauge has a dead band at reversal set by pivot friction, and the width of that band does not scale with the displacement being measured. A linearity check across the range therefore cannot see it, while a measurement taken near the tolerance limit is dominated by it. The dead band is a property of the mechanism, not of the range it is used across.
Degeneracy of two causes
H4916. A gauge's magnification change with temperature and its change with pivot wear produce the same slope error when only one master is used, so a single master cannot separate them. Two masters measured at two temperatures do separate them, because wear changes the slope near one end of the range while temperature scales the whole range.
Operator as stiffness probe
H4917. Two operators using the same hand gauge read different values because each applies a different contact force, and the difference is proportional to the part's compliance. The reproducibility term in that study is a measurement of the part's stiffness in disguise.
Disagreement as measurand
H4918. On a soft polymer a contact gauge and an optical gauge disagree systematically because the optical gauge sees the skin and the contact gauge sees the bulk. The disagreement is not an error to be reconciled but a measurement of the skin's thickness, and treating it as gauge error destroys the information.
Quantisation-induced sawtooth
H4919. A digital gauge whose resolution equals the last digit of the tolerance produces an acceptance probability that oscillates with the true value, because the decision boundary falls inside a rounding bin. The oscillation is invisible in a capability figure computed from the same rounded values, and it grows as resolution approaches the tolerance.
Wrong thermometer
H4920. An in-line gauge's reading shifts when the part's temperature differs from the master's, and the shift is proportional to the part's expansion coefficient times that difference rather than to the gauge's own temperature. A gauge with its thermometer on the frame is therefore compensating the wrong body.
Constraint-state transfer
H4921. A formed part springs back when the gauge's rollers release it, so a measurement taken under constraint describes a geometry that exists only inside the fixture. Acceptance limits should be written on the state the part will occupy in service, and a gauge that measures a different state needs an explicit transfer rather than a tighter tolerance.
Optical and non-contact sensing
Illumination-defined dimension
H4922. The measured width of a feature depends on the illumination's angular distribution, so a diffuse dome and a coaxial light give different widths for the same part. The difference scales with the feature's sidewall angle, which makes the illumination part of the dimensional definition rather than a viewing condition. Two plants using different lights are then measuring different dimensions from identical drawings.
Motion-induced noise
H4923. A laser displacement sensor's noise rises with surface speed not because the surface changed but because the speckle pattern decorrelates during the exposure. The noise therefore carries a signature set by the distance travelled in one exposure measured in wavelengths, and it can be mistaken for roughness when parts are measured only on the fly.
Periodic blindness
H4924. A thin-film interference gauge has bands of thickness in which its sensitivity falls to zero, fixed by the source's spectrum, and inside those bands the gauge reports a stable value whatever the film does. The blind bands are predictable from the spectrum, so a gauge specification that omits them is not a specification.
Spectral weighting
H4925. Two chromatic confocal sensors with different spectral ranges weight a rough surface's height distribution differently and therefore disagree by an amount that depends on the surface's slope distribution. That disagreement is itself a measurement of the distribution, and it vanishes on a smooth surface, which is why it survives every calibration performed on a mirror.
Optical-versus-geometric height
H4926. White-light interferometry on a part with a transparent layer places the fringe envelope at the optical height rather than the geometric height, and the offset is proportional to the layer's thickness times its excess group index. A coating thickness measured this way is an optical path, and converting it requires an index the measurement itself does not supply.
Directional smear
H4927. A vision inspection whose exposure time is comparable to the time a feature takes to cross a pixel smears the image along the motion direction. A defect therefore loses contrast when it is elongated across the motion and gains contrast when it is elongated along it, so the same defect can be found or missed according to its orientation on the part.
Field-position bias
H4928. A vision gauge whose scale varies across its field, from a slight perspective or a stage calibration, imposes an apparent taper that reverses when the part is rotated end-for-end. A real taper does not reverse in part coordinates, so measuring the same part in two orientations separates a field-scale error from a part error without an additional standard.
Foreshortening confound
H4929. A camera viewing an angled surface sees a foreshortened feature, and the sign of the apparent width error depends on which axis the part tilts about rather than on the tilt magnitude alone. A fixture that tilts one way therefore hides the error and a fixture that tilts the other way doubles it, which is why two lines with the same camera can disagree.
Centroid pulling
H4930. A laser triangulation spot's centroid is pulled toward a bright band on the surface, so a machined part with a reflective stripe yields a profile error that is a smoothed copy of its reflectance profile. The error survives calibration on a uniform master and appears only on parts that have structure.
Common-mode error
H4931. An unmodulated optical sensor includes ambient light in its reading, and plant ambient varies with skylight, welding, and nearby doors. The resulting error is correlated across every gauge on the line, which makes a process shift and a lighting change look identical in the data. Separating them requires a deliberate change of lighting with the process held constant.
Illuminant metamerism
H4932. A colorimetric gauge calibrated under one illuminant reports a different colour under another even when the part is unchanged, because the two illuminants project differently onto the sensor's channels. The shift is not a property of the part, and a colour tolerance stated without its illuminant is not a complete specification.
Depth-dependent detection
H4933. A vision system's blur grows away from its focus plane, so a defect at the edge of the depth of field has less contrast than the same defect at the centre. The effective detection threshold is therefore lower at the centre of the field, which biases where defects are found and makes escape depend on part placement.
Acceptance testing and decision rules
Space mismatch in guard banding
H4934. A guard band applied symmetrically to the measured value becomes asymmetric on the part when the measurement function is curved near the limit. Applying the band in measurement space therefore over-rejects on one side and under-protects on the other, and the error grows with the curvature.
Value-dependent test uncertainty
H4935. The conventional test-uncertainty margin treats the gauge's uncertainty as constant, but a gauge with a span error has an uncertainty that grows with distance from its calibration point. The effective margin is worst at the specification limit, which is exactly where the accept-or-reject decision is made.
Cost-weighted decision
H4936. The optimal acceptance threshold does not sit where false accepts and false rejects are equally likely; it sits where their rates weighted by their costs are equal. Because those costs are rarely known, thresholds are set by convention, and the convention silently embeds a cost ratio nobody chose.
Information loss by binarisation
H4937. Recording only pass or fail discards the information that would reveal a slow process shift, because a shift appears in the mean before it appears in the rate. A line that stores only decisions cannot distinguish a drifting process from a drifting gauge.
Directional decision rule
H4938. For an upper limit the safe decision is to accept only when the measurement plus its uncertainty lies below the limit, while for a wear limit the safe direction is the opposite. A single decision rule applied to both limits on one drawing must therefore be wrong for at least one of them.
Display-convention effect
H4939. A specification limit that lands exactly on a representable display value creates a discontinuity in the acceptance function, because the display's rounding mode decides which side of the limit that value belongs to. Changing the convention then changes the acceptance rate without changing a single part.
Order-dependent selection
H4940. When a part must pass several tests, the order changes the population that reaches the later tests, because an earlier test removes parts that would have failed later for a correlated reason. The apparent pass rate of the second test is therefore a property of the sequence rather than of the test.
Retest selection bias
H4941. Retesting failures until they pass selects the measurement noise rather than the parts, and the accepted population's mean is pushed away from the limit by an amount proportional to that noise. The bias is larger for the noisier gauge, so improving the gauge without changing the retest rule can make the bias worse.
Coverage-versus-count confusion
H4942. Inspecting every part is not the same as inspecting every surface, and the escape rate depends on defect size relative to scan pitch rather than on the fraction of parts examined. A line can report full inspection while a defect smaller than the pitch passes between scan lines unseen.
Absolute threshold on a relative signal
H4943. Fixed in absolute contrast, a defect threshold misses shallow defects on bright surfaces and false-calls on dark ones, because the informative signal is contrast relative to the local background. The threshold should scale with the background, and the scaling exponent is set by the sensor's noise model rather than by the defect's physics. A single global threshold is therefore a statement about the lighting as much as about the defects.
Saturation of sizing
H4944. An optical sizing system's reported defect size saturates once the defect is larger than the point-spread function, so a large defect and a much larger one can report the same size. The knee of that curve belongs to the optics, not to the defect, and acceptance limits written above the knee are unenforceable.
Objective mismatch
H4945. A detector tuned to maximise detections and a classifier tuned to minimise false calls optimise different objectives, and the operating point that maximises detections is generally not the one that minimises total cost. The gap between them is set by the shape of the defect-size distribution, which the detector never measures.
Orientation-dependent contrast
H4946. A defect's contrast depends on its orientation relative to the illumination, so a rotated defect of the same size can fall below a threshold that catches its neighbour. Escape is therefore anisotropic, and a detection rate quoted without the illumination geometry describes one orientation only.
Gauge repeatability and reproducibility
Study-population mismatch
H4947. The parts used in a repeatability and reproducibility study are the convenient ones, and they are usually more uniform than production. The part-to-part term then shrinks and the reproducibility term absorbs the difference, so the study reports a gauge worse than the one on the line.
Interaction absorbed into reproducibility
H4948. In a crossed study the operator term contains the operator-by-part interaction, and when that interaction is real the reproducibility estimate is inflated. The gauge is then blamed for a training effect that a different presentation technique would remove.
Feedback between replicates
H4949. In a nested study the within-part replicates are assumed independent, but an operator who can see the previous reading tends to repeat it. Repeatability is then underestimated, and the gauge looks better in the study than it does when the display is hidden.
Quantisation of the spread estimate
H4950. A gauge whose resolution is a large fraction of the process variation yields a spread estimate that depends on the bin width and on where the process mean sits relative to bin edges. The same gauge can appear capable or incapable according to a placement it does not control.
Time-scale blind spot
H4951. A short study measures repeatability and misses stability, while a long study confounds stability with part-to-part variation unless the same part is revisited across time. The two designs answer different questions, and the usual short study is silent about the drift that matters over a production week. The short study measures the gauge's best behaviour rather than its typical behaviour.
Single-point validation
H4952. A study run at one nominal size cannot detect a bias that grows away from that size, so the gauge is cleared over a range it was never tested on. Linearity has to be established with parts spanning the tolerance, not by replicating the master. The bias at the limit is the only part of the linearity that the decision actually uses.
Proxy attribution
H4953. The largest reproducibility term in many studies is not the operator but the fixture's seating, because the fixture is reseated between replicates and the operator is merely the agent doing it. The operator term is then a proxy for fixture repeatability, and retraining the operators changes nothing.
Relocation degeneracy
H4954. A part's surface has real variation across the measured feature, and a gauge that cannot relocate on the same spot folds that variation into repeatability. Repeatability and part uniformity remain degenerate until the measurement location is recorded.
Overlapping-average correlation
H4955. If the gauge averages a buffer of readings and the buffer overlaps between replicates, the replicates are correlated and repeatability is underestimated by a factor set by the overlap. The study's precision then depends on a firmware setting that is absent from the measurement record.
Service-in-the-loop
H4956. Recalibrating the gauge between replicates removes the bias component that production would actually see, so the study describes a gauge that is adjusted before every use. A gauge with a large but stable bias looks excellent under that protocol and poor under the one the line uses.
Assumption replacing replication
H4957. For a destructive measurement, repeatability must be replaced by a within-batch uniformity assumption, because two nominally identical specimens differ by more than the gauge's error. That assumption is rarely tested, and when it fails the study attributes specimen variation to the gauge.
Prevalence dependence
H4958. Agreement between two pass/fail gauges depends on how many defective parts the study contains, so the same pair agrees well in a clean batch and poorly in a marginal one. An agreement figure is therefore a property of the batch as much as of the gauges. Quoted without the defect prevalence, such a figure is not comparable between plants.
Imperfect reference ceiling
H4959. When the reference in an attribute study is a second inspector or a destructive test, the gauge's agreement with the reference is bounded above by the reference's own agreement with truth. A gauge can be perfect and still fail a study whose reference is not.
Static study, dynamic use
H4960. Run with the line stopped or slowed, a repeatability and reproducibility study omits the gauge's dynamic error from integration, latency, and motion blur. The capability demonstrated on the bench is therefore an upper bound on the capability the line will have at speed.
Defect detection and thresholds
Threshold hysteresis
H4961. A defect detector with hysteresis has two thresholds, and which one applies depends on the direction in which the scan crosses the defect. The same defect can be detected on a forward scan and missed on a reverse scan, which makes detection a function of scan direction rather than of severity.
Adaptive threshold instability
H4962. An adaptive threshold that normalises by the local background raises its sensitivity in a uniform region and lowers it beside a legitimate feature such as an edge or a weld. Defects close to features are therefore systematically harder to detect, and the width of that blind zone is set by the normalisation window.
Resolution floor
H4963. Below some defect size the detector's response is dominated by pixel or detector noise rather than by the defect, so detection probability falls steeply and the exact threshold stops mattering. The floor is set by sensor resolution rather than by the algorithm, which means tuning cannot move it.
Contrast-transfer function
H4964. A defect's image contrast is filtered by the optics and the sensor's modulation transfer, so two defects of equal physical depth but different widths produce different signals. Detection therefore favours wide defects, and a threshold calibrated on one width does not transfer to another. A depth limit and a width limit are not interchangeable.
False-call asymmetry
H4965. The cost of a false call is paid immediately by the operator who must disposition the part, while the cost of an escape is paid later and elsewhere. The threshold operators prefer is therefore systematically looser than the one the plant needs, and the gap persists even when both parties accept the same numbers, because their feedback arrives on different time scales.
Boundary effect
H4966. A defect that straddles the edge of the inspected field is measured over only part of its extent, so its reported size is smaller than its true size and it can pass a size limit. The field edge is therefore an escape path whose width equals the largest defect that can hide there.
Temporal integration
H4967. A detector that requires a defect to persist across several frames suppresses random noise but also suppresses real defects that move out of the field, and the two effects are indistinguishable at one line speed. Changing the speed separates them, because a real defect's residence time varies with speed while a noise event's does not.
Lighting uniformity
H4968. A defect threshold set on a uniformly lit target transfers poorly to a line where illumination falls off toward the field edges, because the same defect produces less contrast there. The falloff can be corrected, but only if the correction is applied before the threshold rather than after it. The falloff leaves a radial gradient in the false-call rate, which is easily mistaken for a defect distribution.
Polarisation dependence
H4969. A defect's visibility under polarised illumination depends on its surface texture, so a detector optimised for scratches on a machined surface can be blind to pits that scattering would reveal. The illumination's polarisation is therefore part of the detector's definition of a defect.
Defect definition
H4970. Two plants can agree on a defect's size and disagree on whether it is a defect, because the acceptance criterion is functional while the measurement is geometric. The disagreement is not a gauge problem, and no improvement in sizing resolution will resolve it. The fix is a functional test, not a better gauge.
Detection asymmetry
H4971. A dark defect on a bright background and a bright defect of equal contrast on a dark background are not equally detectable, because both sensor noise and illumination depend on the background level. Detector performance is therefore stated per polarity, and a single figure averages two different problems.
Calibration-set coverage
H4972. A detector calibrated on a set of known defects inherits their size and shape distribution, performing well on defects like those and poorly on any the set did not contain. The performance figure describes the calibration set more than it describes the line.
Traceability and the calibration chain
Traceability of the artefact, not the quantity
H4973. An in-line gauge calibrated against a master part is traceable to the master's geometry, but if the master's surface differs from the production surface the chain is traceable in name and not in measurand. The chain then certifies a length on a surface the line never measures.
Traceable to the wrong condition
H4974. Calibrating a gauge in place with a master at a different temperature from the line inserts the master's thermal state into the chain, so the calibration is traceable to a condition the process never occupies. The resulting bias is not a chain failure but a chain success at the wrong condition.
Hierarchy inversion
H4975. A working gauge can be more stable than the reference used to calibrate it, in which case every calibration adds scatter instead of removing bias. For such a gauge the optimal chain is inverted, with the working gauge as the reference and the laboratory standard checked against it. The inversion should be checked whenever a gauge's history shows better stability than its calibration reference does.
Two criteria for interval
H4976. A calibration interval set by observed stability is longer than the interval that minimises total cost, because the cost is dominated by the chance of accepting bad product between calibrations. The gap between the two intervals widens as the tolerance tightens.
Selected population of as-found data
H4977. As-found errors are recorded only for gauges that enter calibration, so the population of as-found data is selected by the calibration programme itself and cannot estimate the line's error rate. A line with many uncalibrated gauges will appear better than a line that calibrates everything.
Adjustment at the wrong temperature
H4978. An as-left adjustment that brings a gauge to nominal at the calibration temperature leaves a bias at the line temperature whenever gauge and part expand differently. The adjustment is optimal for the room it was made in, and the line inherits the difference.
Borrowed repeatability
H4979. An uncertainty budget for an in-line gauge often borrows the laboratory's repeatability, which excludes the fixture and the part. The in-place repeatability includes both, so the budget understates the uncertainty exactly where the decision is made.
Correlated chain
H4980. Two links in a calibration chain calibrated against the same reference have correlated errors, and combining them by root-sum-square understates the total. The correlation is invisible in either certificate because each reports only its own link. The correlation has to be established from calibration records rather than inferred from the certificates.
Versionless traceability
H4981. A gauge's firmware is part of its calibration, but the version is rarely recorded on the certificate, so an update that changes a filter or a rounding rule silently moves the gauge into an uncalibrated state. Traceability then points to a version of the instrument that no longer exists.
Orphaned value
H4982. A measurement stored without its time, temperature, part identity, and gauge identity is traceable to a gauge but not to a condition, which for acceptance is the same as being untraceable. The metadata is usually the first thing lost when systems are joined.
Inherited bias
H4983. Master parts are often recertified by the same laboratory that certified them, so a systematic error there is inherited by every gauge calibrated against the master and stays invisible in the chain's internal consistency. Only an independent path to the definition can expose it.
Consensus by weight
H4984. In a comparison among laboratories the consensus is dominated by the laboratories claiming the smallest uncertainties, so a laboratory that is right and precise can be flagged as an outlier because the consensus is wrong. The exercise measures agreement with the majority, not agreement with the definition.
Two traceabilities, one decision
H4985. A drawing's limit is traceable to a design definition while the gauge's calibration is traceable to a length definition, and the two are joined by an assumed measurement function. If the function in use is not the one the designer assumed, both ends remain traceable and the decision does not.
Thermal, environmental, and dynamic effects
Wrong thermal node
H4986. A painted part measured in-line has a skin temperature different from its core, and an optical gauge with shallow penetration sees the paint while a contact gauge sees the metal. The two gauges then disagree by the difference between two thermal states rather than by their accuracy.
Differential expansion
H4987. A gauge's scale calibrated at one temperature and used at another has an error equal to the difference between the gauge's expansion and the part's, not to the gauge's expansion alone. A fortunate material pair can make that difference nearly vanish, which is why one line is stable and its copy is not. A gauge transplant between lines therefore changes the error even when the gauge itself is unchanged.
Compensation at the wrong place
H4988. An interferometric gauge compensates the refractive index of air with a sensor at its head, but the beam traverses air along its whole path, which may be warmer or more humid than the head. The compensation is therefore exact at a point the beam does not occupy.
Moving thermal plume
H4989. A hot part moving through the line drags a thermal plume that changes the air's index along the beam, so the optical path varies with the part's position and speed. The effect grows with the part's excess temperature and with path length, and it is absent from every static calibration.
Film as part of the measurand
H4990. A wet part's optical measurement includes the coolant film, whose thickness depends on geometry and drainage time, so the gauge reports film plus part. The film contributes most where the part is flattest, which is usually where the dimensional tolerance is tightest.
Structural coupling
H4991. A gauge mounted on the line's frame sees the frame's vibration, whose spectrum peaks at the drive frequencies, so a gauge with enough bandwidth reports part variation that is really frame motion. The coupling can be identified by running the drives with no parts on the line.
Environmental noise floor
H4992. An ultrasonic gauge's noise floor rises near compressed-air exhausts and other broadband sources at its operating frequency, so the same gauge has different detection limits at different points on the same line. A specification measured in a quiet room does not transfer to the line's acoustic environment.
Gap-proportional humidity bias
H4993. A capacitive gauge's reading depends on the humidity of the air in its gap, and the dependence grows with gap because more air lies in the field. A gauge calibrated in dry air and used in humid air therefore carries a bias that is larger for large gaps, which is the opposite of what a scale-factor correction assumes. A one-point humidity correction at a fixed gap makes the error worse at other gaps.
Illuminant aging
H4994. An LED illuminator's spectrum shifts as it ages, so a vision gauge calibrated when the illuminator was new sees a different spectrum later and its colour-dependent thresholds move with it. A daily check on a white target will not show the shift, because the target's brightness can stay constant while its spectrum changes.
Data, software, and the line as a system
Phase-delay reversal
H4995. A low-pass filter applied to a profile delays features in the scan direction by an amount proportional to its phase response, and the delay reverses when the scan direction reverses. A bidirectional scan therefore shows a feature-position difference that belongs to the filter rather than to the part. Zero-phase filtering would remove the reversal, at the cost of no longer being causal.
Size-selective compression
H4996. Lossy compression removes high-frequency detail first, so a small defect's contrast is reduced more than a large one's and the effect is size-selective. It can be calibrated only against the defect-size distribution the inspection is meant to catch.
Curvature bias in interpolation
H4997. Sub-pixel edge interpolation assumes a straight edge, so on a curved edge it places the edge toward the inside of the curve by an amount proportional to curvature times pixel size. Radius measurements are therefore biased small on small radii, and the bias is invisible in a straight-edge calibration.
Commensurate averaging
H4998. Averaging several readings reduces random noise but not a periodic error locked to the line's cycle; the periodic error cancels only when the average spans an integer number of cycles. Choosing the averaging window from the noise level rather than from the line's period leaves a systematic error intact while appearing to improve precision.
Timestamp latency
H4999. A gauge timestamps data when it is received rather than when the measurement was made, so at line speed the part has moved in between and the recorded position is offset by latency times speed. The offset is invisible unless gauge and encoder are compared directly on a single part. Sorting by receive time rather than by measurement time corrupts every position-based analysis.
Bottleneck bound
H5000. A line's measurement capability is bounded by the slowest traceable reference in its chain, not by the gauge at the point of measurement, so improving the gauge cannot improve the line beyond that bound. The bound can be located by tracing which reference each acceptance decision actually depends on.
Status
100 entries, H4901–H5000. Together with the preceding ranges this closes the five-thousand-item plan. The forms are deliberately uneven in weight: some entries are bounds, some are degeneracy claims, some are decision-rule inversions, and a few are only worth writing down because they can be falsified cheaply.
If one of these is worth an experiment, it is H4975 — that a working gauge can be more stable than the standard used to calibrate it, so that calibration adds scatter instead of removing bias. It is checkable from any calibration history that already exists.
Incoprea Research · hand-considered hypotheses 4901–5000 · 3 November 2026 · ~/Desktop/Nov3/