Architecture, Simulation & Component Decisions

A precise simulation answers the wrong question

Which models and experiments are sufficient to select architecture and components without mistaking simulation precision for product truth?

The gateway receives the monitoring node correctly while its backhaul is quiet. The matched-port allocation also clears its internal power target. Turn on the simultaneous backhaul stimulus and messages sometimes miss their application deadline. The schematic still looks correct. Re-running the same model with a finer mesh produces more decimal places, but no explanation of the missing messages.

That is not necessarily a solver defect. A calculation of transferred power between two matched ports never included receiver desensitization, supply disturbances or scheduling. Those mechanisms can change what reaches the detector at R3 or when a complete message reaches D0. The first architecture decision is therefore about the question: which observation would distinguish insufficient RF isolation from an incompatible control schedule?

This lesson assumes Paths 02–07 and 10.1’s framing snapshot. Use 05.7’s hard-constraint trade method, 03’s port conventions, 04’s component-model limits and 06.3’s integration evidence. Here the new work is deciding how much evidence is enough for a particular commitment.

Turn framing into architecture candidates

Can the attractive candidate satisfy the same problem as its competitors?

Translate the framed need into three proposed front ends before comparing preferences. C-INT keeps more functions within an integrated gateway arrangement. C-DISC offers separate filter, gain and conversion partitions. C-ALT is a possible substitute at C-INT’s interface. These are fictional architecture cards, with no vendor part numbers, guaranteed ratings or supply promises.

The common basis is the inherited service, not an isolated receiver sensitivity number. SERVICE-FR-v1 requires at least 990 of 1000 unique complete 32-byte payloads at gateway D0, within 1 s including the declared timing uncertainty. Requests occur every 60 s, with at most one retry at +200 ms. C-CONC adds the 5.180 GHz, +10 dBm R1 backhaul stimulus for 20 ms per 100 ms while receiving telemetry. Keep the 100 m, 1.5 m heights, vertical polarization, fixed azimuth and 25±2 °C geometry unchanged when comparing that claim.

Integration may reduce external interfaces but concentrates hidden coupling and control assumptions. A discrete topology may offer a useful filter or measurement point while adding loss, bias paths and LO isolation work. The manufacturer’s building-block discussion explains these coupled RF attributes; it does not guarantee either proposed architecture.

Reconcile one set of cross-budget inputs

Where was the feed loss already counted?

A power spreadsheet begins at the component port R1, an antenna report begins at the product feed R2, and a link budget begins with directional EIRP at S0. Their numbers can agree individually and still double-charge the same physical loss when combined. Name the transformation before moving the value. In the matched fixture, subtract the positive B-FEED once between R1 and R2, then add the accepted-power antenna gain.

At 3.0 dB feed loss, the pair becomes 7.0 dBm R2 and 10.0 dBm EIRP. At zero loss it is 10.0 and 13.0. CL-PWR uses an inclusive minimum: 2.5 dB loss gives exactly 7.5 dBm and supports the local model claim. Classification compares integer tenths, not a rounded label. Antenna mismatch is outside this fixture; adding realized gain later requires a new plane contract to prevent counting mismatch twice.

Choose the smallest sufficient model

What is the smallest model that could change this decision?

Choose fidelity after naming the disputed mechanism. A hand bound can reject a power allocation before a schematic exists. A behavioral model can test a timing or frequency-plan assumption. A circuit model may expose bias dependence or nonlinear distortion. An EM model becomes useful when current paths, fields and geometry are the uncertainty. A prototype is useful when its observable and configuration actually discriminate the alternatives.

Ansys’s pinned HFSS meshing guide describes refinement against changes at declared adaptation frequencies. That is a numerical check within a model. A stopping condition or small residual cannot establish that the enclosure, package, excitation or material values describe the product. Our thermal example is explicitly a fictional solver premise, not an HFSS run.

Write the model contract and falsifier

Which observation would make you stop trusting this model for this claim?

A model contract makes its use challengeable. It names the source revision, parameters, operating point, frequency/power/temperature domain, port normalization, excitation, package/stackup/enclosure assumptions and boundary. It also says which convergence check applies—and why solver or mesh is not applicable to direct arithmetic or a documentary review.

The falsifier connects that contract to a practical observation. If a calibrated R1/R2 measurement shows a mismatch or loss not represented by M-HAND-v1, a correct subtraction no longer establishes the physical power claim. If controlled backhaul off/on tests miss the same D0 criterion, a nominal single-radio packet cannot rescue the concurrency decision. Specify these rejection conditions before fitting parameters, so “good agreement” cannot become whatever the latest curve happens to show.

Select components with usable evidence

Does the candidate have the evidence you need, or only the number you hoped to see?

A typical number describes a representative characterization under stated conditions. A guaranteed limit is meaningful only with its declared operating conditions and qualification/test basis. A characterized curve adds information about a particular setup or population. An illustrative number is chosen for teaching. Keep those labels beside each value: placing them in the same column does not make their evidential roles equal.

As a separate real documentation example, ADL5602 Rev. A gives typical S-parameters at 5 V and 25 °C, with the fixture removed to the device pins. Its absolute-maximum discussion distinguishes stress limits from functional operation. The pin and land-pattern sections add bias and grounding obligations that a scalar gain cannot capture. This datasheet is not a part assignment to any fictional candidate.

AN-2591’s LFCSP discussion connects dissipated power, thermal resistance and boundary temperature to junction temperature. Copying only a package thermal number leaves out the rest of that relationship. The real ADL5602 product page displayed Recommended for New Designs on 2026-09-10; that dated observation is not a delivery commitment. No lifecycle status is invented for C-INT, C-DISC or C-ALT.

Partition the simulation without losing interfaces

What gets lost when one simulation hands its result to another?

Partition where the question changes, then reconcile the interface. A linear S-parameter network describes small-signal wave relationships at defined ports; it does not automatically carry large-signal compression or control-state history. A nonlinear circuit model needs the actual waveform, bias and load. A time-domain service model needs the RF impairment mapped into receiver decisions and then the same generated-message denominator.

A package-to-connector S-parameter dataset can already include a launch. Adding the launch model again produces a physically different network. Similarly, the same jitter source cannot be counted once in a total jitter value and again as a separate component. These are identity errors across budgets, even when every local equation is correct.

Correlate before locking the whole assembly

Will this prototype distinguish the architectures, or merely reproduce a result both already predict?

A cheap experiment earns priority only if its possible outcomes can change the decision. Repeating quiet single-radio gain may be easy, but both a coupling problem and a control-scheduling problem can produce the same gain in that state. It does not distinguish the disputed mechanism. Controlled simultaneous-mode evidence needs the aggressor state, wanted input, timing/control records and the identical service denominator.

Before acquisition, pin the board/sample IDs, calibration plane, instrument/method version, probe loading, thermal boundary, fitted parameters, uncertainty and acceptance rule. Repeat a reference condition after reconnecting or a long run to detect drift. Changing enclosure, board stackup or sample creates new evidence applicability work; fitting another free parameter after seeing the validation residual consumes its independence.

Freeze, preserve an option, or run an experiment

Which commitment is actually blocked by this missing fact?

Evaluate each claim independently, then apply the decision policy. Eligible support alone is supported; eligible contradiction alone is contradicted; both in the same scope are conflicting; neither is unsupported. Ineligible and inconclusive records remain visible without voting. A precise wrong-boundary model cannot outvote an applicable observation.

D-FRONT needs CL-PWR, CL-CONC and CL-THERM. D-SECOND needs CL-ALT. If a required hard claim is uncontested and contradicted, revise or reject the affected choice while retaining any other reconciliation actions. Otherwise resolve eligible conflicts, then obtain a compatible discriminator for unsupported claims. If no discriminator is defined, defer and define evidence. Complete support makes a decision ready for a bounded review; it does not supply that review.

1. Predict the canonical 8/11 dBm pair and inspect the rejected double subtraction.
2. Load Wrong domain and inspect why EV-THERM-v1 is ineligible.
3. Reset, add only RP-CONC and compare: only CL-PWR and CL-CONC are supported.
4. Add RP-THERM: D-FRONT is ready for model-interface review; D-SECOND is still blocked.
5. Add RP-CONFLICT: both concurrency conclusions remain visible and D-FRONT calls for reconciliation.
6. Change B-FEED to 3.0 dB: CL-PWR is contradicted, so revision takes precedence while concurrency still needs reconciliation. Inspect D-SECOND’s separate state.
7. Try a claim filter, remove retention, or remove the thermal plan; compare the causal change. Reset restores every defining fact and retains focus.

Defend the architecture and rollback plan

Can another engineer reconstruct—and reverse—your decision?

A useful review memo explains why the chosen next step follows from the evidence and how it could be wrong. It names the model that was rejected, the option that remains physically available and the observation that would reverse the commitment. This is more useful than a green architecture diagram with no explicit boundary.

The feed-only example reaches exactly CL-PWR and D-FRONT. It creates ALLOC-FEED-v2 and EV-BUDGET-v2 while retaining the old derivation. RETAIN-FEED-v2 explicitly maps unaffected concurrency, thermal and alternate uses to unchanged physical facts. Remove that premise and their applicability becomes pending. A real feed/component change may alter matching, heat or coupling and requires wider discovery; the tiny graph is not an exhaustive physics model.