Production, Change Control & Field Feedback

A scalar line pass conceals degraded RF behavior

How can production and field evidence preserve RF performance while enabling controlled cost, supplier, software and process changes?

A technician adjusts an antenna matching code until one power screen passes. The unit now looks better in its fixture. Yet a different channel, a simultaneous receiver mode or the radiation pattern on a metal bracket can get worse. Changing a reactance changes the impedances and current distribution seen by more than one measurement. If the line retains only the last passing number, the record has also lost the first-pass distribution and the intervention that changed the specimen.

Production must preserve a design intent that was established across configurations. It cannot recover that intent from a scalar threshold alone. An R2 conducted power reading can help control a transmitter characteristic, but it cannot establish antenna efficiency, S0 directional behavior, D0 service delivery or regional software correctness. A stable reading may even mean that a drifting fixture and a retuned reference unit now agree with each other.

This is the fictional teaching-node-v1 condition-monitoring node/gateway: generic 2.450 GHz QPSK, 20 kbit/s information, 10 ksymbol/s, RRC roll-off .35, 8-symbol filters and 80 ksample/s; 256-symbol bursts occupy 25.6 ms. It is not automatically Bluetooth, Wi-Fi or a permitted transmission. A and B below are separate conducted-power variants, C is a genealogy population and D is an installed cohort. Their rows cannot be joined.

Bring the 10.1 specification, 10.2 architecture, 10.3 physical review, 10.4 bring-up history and 10.5 closure restrictions. Path 08 owns the deeper measurement/debug methods; Path 09 owns actual market applicability. Here we own the production evidence and feedback plan.

Transfer design intent into critical characteristics

Which characteristic must remain controlled for the service requirement to stay believable?

R-RANGE-v2 requires at least 990 of 1000 unique complete 32-byte D0 payloads within 1 s, including timing uncertainty. SERVICE-FR-v1 fixes 60 s requests and one retry at +200 ms. The reference S0 geometry is 100 m unobstructed, antenna reference points 1.5 m high, vertical aligned polarization, azimuth 0°, stationary and dry at 25±2 °C. These conditions are part of the claim, not optional test notes.

Work backwards from service to critical characteristics: delivered RF power, receive behavior under interference, antenna/environment coupling, simultaneous-operation behavior and configuration integrity. Some can be screened rapidly on every unit. Others need periodic or sampled characterization, or a controlled assembly process that prevents a physical error. Each assignment needs evidence that its chosen control can detect the relevant variation; “covered” cannot mean simply that a test exists.

Keep interfaces explicit. R1 is the component RF port; R2 is the product antenna feed; S0 is a specified OTA geometry; R3 is the detector/decision boundary, not the receiver RF input. The inherited B-FEED is one matched 2 dB allocation: +10 dBm at R1 gives +8 dBm at R2, and a separate 3 dBi directional antenna term gives 11 dBm in that direction. Do not replace this budget with A’s or B’s higher-output bench variant. The implemented Tools catalogue can support calculations; a calculation still needs these plane and configuration labels.

Choose screens by what they can detect

What does saving test time cost in escape visibility?

Test time competes with settling time, connection repeatability and the opportunity to observe a mechanism. A single nominal power point is quick because it removes dimensions. Those removed dimensions may be exactly where a solder defect, a spur, a bad ground contact or a software error appears. The right reduction starts with an assigned failure question and a demonstrated relation between the short screen and the more complete characterization.

For example, a connectorized R2 measurement avoids uncertain OTA geometry, but access can disturb the antenna feed or require an adapter whose loss changes with wear. A shielded radiated check observes the assembled path, yet placement and coupling become part of its method. Neither approach is universally superior. Production need not repeat full DVT on every unit; its combination of screens, characterization and process controls must have a defensible detection argument.

Correlate the same quantity at the same plane

Can a perfect scatterplot relationship still give the wrong indication at every point?

Fixture p10-m06-A-paired-v1 supplies four paired R2 mean active-burst power indications in dBm under MAN-A-v1. Each lab/line pair shares specimen identity, waveform, temperature, power state and the named 50 Ω R2 termination. Both methods exclude idle time. An average over the entire burst-plus-idle interval would be a different statistic. A supported reference-plane transformation is required before subtracting readings from different ports.

The mean b is the arithmetic mean of logarithmic differences in dB, not mean RF power in watts. For a mean linear power, convert each dBm value to a linear unit before averaging and name that different statistic. Correlation needs at least two eligible pairs and nonzero variation in both series. Missing identity or an incompatible plane/configuration excludes a pair; row order never creates a match.

In A-SPREAD, the squared deviations sum to .04 dB²; divide by three for sample variance. Under independent sampling of paired differences, u(mean)=s_d/√n. Shared reference errors do not diminish merely because n increases. NIST TN 1297 provides the uncertainty framework, and NIST’s gauge-comparison guidance motivates examining matched differences. Zero observed spread in the baseline does not establish zero uncertainty, adequate resolution or stable transfer.

CORR-A-v1 defines y_corrected=y−0.5 dB within the supplied comparison domain. Its current CAL-A-v1 packet permits a labelled corrected display; it does not authorize production acceptance. Keep raw values and correction identity, quantify the remaining uncertainty and independently validate the transfer. Changed, overdue or unknown calibration/bias-model evidence blocks the correction. No extrapolation across frequency, power range, site, fixture or temperature is supplied.

Control references that can drift

What checks the checker when the line and its golden unit move together?

A calibrated reference standard carries a defined quantity, calibration evidence and uncertainty. A transfer artifact transports a comparison between setups. A golden DUT is a chosen product specimen with its own history. A fixture establishes access and repeatability. These roles may interact, but they are not interchangeable: a golden DUT is neither immutable nor proof that a production population is typical.

Thermal changes, cable movement and connector wear can change the measurement path. Retuning a golden DUT can also restore its displayed value while hiding a fixture problem. Keysight’s instrument guidance distinguishes random effects such as noise/reconnection from drift after calibration. NIST’s drift discussion explains why a short comparison cannot establish long-term stability. The evidence plan therefore records environment, connection, calibration, software, maintenance and artifact revision.

The local trigger treats 0.1 dB as below threshold, 0.2 dB as equality requiring review and 0.3 dB as above threshold. It identifies a changed relationship, not a defective DUT. If the line and golden check use the same drifting reference, apparent agreement can survive the error. The independent-source cross-check changes that common dependency; the separate DUT-by-fixture swap then helps localize the discrepancy. Neither supplied experiment alone proves a universal cause or replaces a full uncertainty review.

Separate specification, screen and process limits

Which boundary describes the product, and which boundary decides what the line accepts?

p10-m06-B-screen-v1 is a separate five-unit exercise for Q-PC-PWR-v1, R2 mean active-burst power. Its latent x values are exact supplied teaching premises, not uncertain laboratory truth. The internal specification is [9.5,10.5] dBm inclusive. The uncorrected line has y=x+0.2 dB, independently of A. Classification by x and acceptance by y are two different operations.

Unguarded’s conditional false-accept fraction is 1/2 nonconforming specimens; guarded’s conditional false-reject fraction is 1/3 conforming specimens. The corresponding fractions of all five specimens are each 1/5. Name the denominator. None estimates an actual factory escape probability: that needs a justified joint product/measurement model. The .2 dB guard band is a local teaching choice, not a confidence level or recommended general margin.

JCGM 106:2012 distinguishes acceptance intervals and decision risks. A real uncertainty-aware acceptance policy must be stated separately; it is not silently added to this known-latent example. Specification limits define a desired characteristic, screen limits act on indications, and process control limits describe a characterized process. Calibration correction adjusts a measurement model; RF tuning changes the specimen. These operations have different evidence and review consequences.

Preserve unit genealogy and first-pass history

Could you identify every unit that actually received the changed part or program?

A serial number becomes useful when it resolves to the configuration used for an observation. BOM revision alone may miss a second-source lot; firmware version alone may miss region and power tables. A passing result without fixture/software/calibration identity can become unusable when a line-method discrepancy appears. Genealogy connects these facts without rewriting the original observation.

C’s six manifests C-U01-v1 through C-U06-v1 are separate from A, B and D. U06’s crystal source is unknown. That is a missing fact in this exercise, not evidence for X1, X2 or a failure. A supplied genealogy packet creates a new manifest with a parent; it does not quietly fill the old record. Likewise, C-U03-v2 preserves v1 and changes its firmware to F2. A v1 review cannot automatically speak for that new state.

A result may remain historically correct while becoming inapplicable to a new claim. Retain the baseline artifact, record the changed evidence use, propose the action, then record a scoped review disposition after its evidence exists. This prevents both destructive overwriting and the opposite error of treating every historical report as indefinitely current.

Follow a change through evidence and units

Which exact unit-and-question uses need review after this change?

X2 can reopen frequency/spur and receiver/transmitter questions. E2 can alter antenna loading and the assembled transmitter path. F2 can change regional power or concurrent behavior. Metal mounting can change antenna currents and configuration instructions. These are reasons to review specified evidence uses, not declarations that every carrying unit is defective.

The count is 6 crystal uses + 9 enclosure uses − 2 shared U04 uses = 13 known pairs. Taking the cross product of all affected units and all questions would invent impacts. Known paths dominate a pair’s membership without erasing other unknown relationships. An unmodeled event requires impact discovery; an empty lookup cannot establish no impact.

RP-C-U03-RX-v1 supplies an illustrative review premise only for U03/PC-RX, C-U03-v1 and CH-X2-v1. It changes one of X2’s six known pairs to “review recorded,” leaving five known and three potential pairs pending. Membership remains the same. The wrong-version packet targets C-U03-v0; changing U03 to v2 also makes the v1 review inapplicable. No other unit, question or version inherits that disposition.

PC-DOC includes configuration and formal-evidence applicability. ISED RSP-100’s change provisions require attention to hardware/firmware modifications under the Canadian certification scope. As a separate Australian example, ACMA asks suppliers to document changed products, and its testing guidance calls for retesting after changes. This graph provides engineering questions for that review; it cannot decide a legal change class, waive a test or permit continued supply. The inherited EU-DE, CH, GB, US and CA questions stay with their Path 09 owners.

Turn field returns into reproducible hypotheses

Does the field cluster describe exposure, a reporting process or a reproducible RF mechanism?

p10-m06-D-field-v1 describes 100 fictional units installed before the common 30-day window, 2026-08-01 through 2026-08-30 inclusive. The complete packet supplies full follow-up with no withdrawals or censoring; it does not prove trouble-free service. A return means one unit with a reported intermittent service symptom returned for investigation; repeated contacts do not create extra returns. Exposure and return counts are supplied aggregate records, with no personal identifiers.

The firmware breakdown changes the descriptive comparison: within F2, the plastic stratum has more observed returns per installation. Firmware, batch, battery state, site, orientation, installation and reporting behavior may still be entangled. No fitted causal model is supplied. Incomplete follow-up, different service time and unreported failures can distort even a correctly calculated proportion. Gather only the technical evidence needed, with the appropriate access and retention plan; this lesson collects nothing.

At RF level, metal proximity can change impedance, radiation pattern or common-mode current. An RSSI report can also change because of orientation, cable state, receiver behavior or software interpretation. A controlled comparison must separate these possibilities. Keep the same specimen and physical conditions, change one documented mount configuration, repeat the sequence and revert. Observe R2, S0 and D0 together so a better match is not mistaken for improved service.

Close the learning loop and defend the portfolio

How does an investigation change what the next design must prove?

Suppose the controlled comparison supports a spacing/orientation mechanism under its declared conditions. Treat that as a conditional worked decision, not as a measurement supplied by the cohort. A note saying “avoid metal” would leave geometry, installation instructions and verification ambiguous. Instead, revise a requirement or configuration envelope, carry the old version forward as history and define the evidence needed to accept the new boundary.

The feedback reaches the model as well as the drawing. A new enclosure boundary challenges the rejected open-board thermal assumption from 10.2; a changed feed/cable state challenges a matched loss allocation; a mounting instruction changes the S0 representation claim. Reopen only the affected uses with a reason, but do not let a small graph hide an unmodeled interaction. A corrective action closes a defined question when the evidence supports it, not when the action is entered.