# Independent review of the fault refinements

Date: 2026-11-11. Final independent disposition of this bounded extension. This record supplements `INDEPENDENT_REVIEW.md`. All 32 successful root fault cases, the retained failed no-load case, and the four-case conducting-ring comparison are now included. Completion of this review is not physical validation of a motor or finished magnet.

## Outer-boundary sensitivity resolved for one saved state

The saved 1.5 mm ring-free rotor's worst instantaneous fault-field state was re-solved with exactly the same rotor angle, phase ampere-turns, Br and near-rotor mesh. Outer air was enlarged from 3 to 6 to 12 mm while preserving 0.75 mm outer-air radial spacing.

| Outer air (mm) | Opposing H q99 (kA/m) | Phase-linkage vector norm (Wb) |
|---:|---:|---:|
| 3 | 978.07564 | 0.00594072 |
| 6 | 977.56503 | 0.00595703 |
| 12 | 977.05177 | 0.00595872 |

The 3-to-12 mm H change is −0.1047%; the final 6-to-12 mm change is −0.0525%. The linkage changes are +0.3030% and +0.0283%. Both meet the prespecified final-domain <=1% screen.

This narrows the numerical outer-boundary concern for that static current/angle state. The fault trajectory and cumulative material exposure were not recomputed in enlarged domains, and no 3D end geometry or material properties were validated.

Reproduction: `outer_domain_check.py`; inputs/criteria: `DOMAIN_CHECK_SCOPE.md`; results and hashes: `outer_domain_check.json`.

## Native five-construction refinement

Independently reproduced from the retained coarse/fine summaries:

| 1.5 mm construction at matched prefault torque | Coarse H envelope (kA/m) | Fine H envelope (kA/m) | Change relative to fine |
|---|---:|---:|---:|
| Radial / 1 mm SUS-class ring | 549.09794 | 549.17921 | 0.0148% |
| Radial / 1 mm generic-steel ring | 532.70110 | 532.42508 | 0.0518% |
| Radial / 1.5 mm SUS-class ring | 531.18366 | 531.73368 | 0.1034% |
| Halbach / no ring | 1082.35720 | 1006.87461 | **7.4967%** |
| Halbach / SUS-class ring | 1268.11499 | 1148.22295 | **10.4415%** |

Matched prefault torque means the four-position sampled mean rotor-torque magnitude, approximately 3.0445 N m, under a constant-q current choice. It does not mean identical instantaneous torque at every fault inception angle.

The two Halbach field percentiles fail the builder's frozen 3% refinement criterion. Their phase-current peaks change only about 1.1–1.3%. This is direct numerical evidence that aggregate electrical convergence does not establish local magnetic-field convergence.

Their precise required-knee thresholds are not accepted. Report both grid results as a numerical spread, not a confidence interval or guaranteed physical bound. Both currently exceed the selected nominal specimen comparison, but failure of numerical refinement does not establish physical infeasibility. An automatic n16 campaign is not required to rescue these cases. A further bounded refinement is decision-relevant if the separate 2 mm mass-surviving candidate approaches a material/process boundary or its ranking changes.

Artifact: `native_five_grid_review.json`.

The subsequently collected 2 mm ring-free case at 2,000 rpm also fails its own 3% field criterion: 1082.27598 to 1013.79079 kA/m, a 6.76% difference relative to fine. It must not inherit acceptance from the radial cases or from the separate outer-domain check. Its precise threshold remains unresolved; its material-mass credit remains conditional. The radial baseline's additional n8-to-n16 field change is only 0.0321% (549.17921 to 549.35558 kA/m), supporting that specific quantity within the declared model.

For segmented rotors, a physically useful next geometry specification includes inter-segment gaps/adhesive, actual corner radii and magnetization transition/orientation. Abrupt ideal block interfaces can concentrate fields; this is a plausible sensitivity, not a demonstrated cause of the observed convergence failure. A finer idealized calculation alone cannot establish the finished-segment requirement.

## Conducting-ring formulation reviewed before conductive outcomes

The axial investigator supplied the governing equations and code before the conducting-case outcomes. The implementation uses the material derivative of axial vector potential, a floating axial electric field, exact common-partition mass/transfer quadrature, and the existing stationary winding circuit.

For constant ring conductivity:

`Jz = sigma [e − (Anew(x)−Aold(R_minus_delta x))/dt]`.

With `w=dt e`, consistent mass M, material transfer T, b=M1 and S=1^T b:

`F(A)−C q + sigma/dt (M A−T Aold−b w)=0`,

`b^T A−1^T T Aold−S w=0`.

The second equation enforces zero net axial current for the connected ring. Eliminating w gives the proper conductivity-weighted floating-potential projection. The winding flux remains in the stationary stator frame and therefore uses unadvected old A. The Schur-complement signs and source/heat units are consistent.

The implementation's common four-subtriangle quadrature correctly addresses the different P1 partitions after an odd angular-column shift. Projection can dissipate unresolved piecewise-linear structure; it is not exact arbitrary-field interpolation.

Before outcomes, the reviewer requested:
- constant, integral and mass preservation;
- ordinary-triangle consistent-mass crosscheck;
- correctly directed material-motion checks;
- positive projected damping and passive decay without motion/excitation;
- zero/tiny conductivity agreement;
- local residual and net-current checks.

The original radial-layer motion test could not distinguish positive from negative rotation. An independent two-column non-axisymmetric manufactured test closed that gap before conductive outcomes: correct direction relative residual 4.78e−18; deliberately reversed direction residual 1.418. The checker and tested implementation hashes are recorded in `eddy_transport_direction_check.json`; executable code is alongside it.

The fault investigator's reported control results include mass relative error 3.65e−14, passive backward-Euler energy-balance error 9.63e−16, and tiny-conductivity field/current differences 2.30e−13/8.21e−15. These are numerical controls, not measurements of a physical ring.

## Conditions that prevent a physical fault claim

The selected conductivity and generic magnetic curve are an assumed combination, not a characterized or qualified material grade. The ring is represented by a homogeneous 2D axial current path, with zero-net-current closure but no finite-end return resistance. The initial condition has no pre-fault eddy currents. Magnets are nonconducting, rotor speed is prescribed, and permanent-magnet constitutive behavior has no irreversible history.

Consequently, even a numerically accepted conducting comparison is a sensitivity to a specific declared model. It is neither a conservative upper/lower bound on real local reverse field nor a validated 10 mm motor fault. The missing application inputs are ring conductivity/B-H at temperature, construction and electrical continuity, axial end connections, magnet conductivity/segmentation, pre-fault current/voltage history and rotor/propeller mechanical dynamics.

The ordinary-load diffusion-timescale audit justifies testing this omission; it cannot determine the sign or amount of the fault-field correction. Conductive comparison outcomes must be considered alongside coarse/fine differences before changing the integrated finding.

## Conducting-ring outcome disposition

The independent reviewer reconstructed phase-current maxima, material-envelope quantiles, ring energy and current-constraint errors from all four raw trace/envelope artifacts. All aggregates matched the investigator's summaries exactly to displayed floating-point precision.

| Angular grid | Zero-conductivity H envelope (kA/m) | Conducting H envelope (kA/m) | Paired change (kA/m) | Zero/conducting peak phase current (A) |
|---|---:|---:|---:|---:|
| n4 | 668.52766 | 667.45299 | −1.07468 | 17.87821 / 17.88824 |
| n8 | 669.97099 | 668.99050 | −0.98050 | 17.95408 / 17.96517 |

Both models pass their frozen 5% current/H refinement criteria: H changes 0.215–0.230% and current changes 0.423–0.428%. The paired H difference changes by only 0.09418 kA/m between meshes. However, the fine-grid effect of 0.98050 kA/m is smaller than either single-model grid shift, 1.44333 or 1.53751 kA/m. The original scope explicitly classifies an effect comparable with grid uncertainty as inconclusive. That rule is preserved; the apparent small decrease is not promoted into an accepted fault-line correction after seeing it.

**Disposition: a small conditional difference appears on both grids; a resolved general conductivity correction is inconclusive.** The computation establishes no large correction at this selected conductivity/material/scenario combination. It neither proves all rotor eddy effects negligible nor provides a conservative physical bound. It does not justify applying this change to other constructions, conductivities, axial closures or material states.

At n8 the computed peak ring heat is 3.69060 W and integrated ring energy is 6.92617 mJ over the two electrical periods. Peak heat changes about 10% under refinement and had no prespecified precision criterion; retain heat and energy as diagnostics, not accurate loss predictions or continuous-duty data. The largest normalized net-current imbalance is 2.92e−15, and all interpreted steps converged. Those checks verify the imposed numerical current closure, not the actual 3D return path.

Artifacts: `check_eddy_outputs.py`, `eddy_output_review.json`; original complete results under `../fault_eddy/`. No extra field campaign is needed to manufacture a decisive positive result from this small sensitivity.

## Matched material, winding and duration checks

The matched typical 45SH case uses Br=1.227 T, 322.35036 peak A-turn/coil and an equal-copper-heat rewind R20=0.774212607 ohm. At the separately assumed 100 C winding temperature the resistance is 1.017625051 ohm. The two-period, 2,000 rpm material-exposure result is 645.73023 kA/m. This coherently connects the selected hot remanence, matched torque current and rewind resistance in one declared scenario. The supplier knee remains a typical specimen observation and the modeled field remains physically unvalidated; their nominal difference is not a qualified finished-part margin. Winding temperature is a declared condition, not inferred from the 100 C magnet observation.

At Br=1.25 T and a 20 C winding, changing only to the equal-heat thin-steel rewind raises the fault screen from 669.97099 to 681.03685 kA/m. This demonstrates a coupled tradeoff within the model: adding copper reduces heat at the operating load while the lower resistance changes the subsequent short-circuit current and field. Adding 0.25 mH of assumed end inductance lowers the baseline 2,000 rpm screen from 587.13195 to 561.86062 kA/m but increases the approximate prefault voltage requirement. These are scenario comparisons, not independently validated construction improvements.

Extending the two coarse-grid control cases from two to seven electrical periods changes their cumulative H envelopes by less than 4e−9 kA/m: the baseline remains 549.09794 kA/m and thin generic-steel case remains 668.52766 kA/m. This establishes that those sampled trajectories had reached their largest material-exposure statistic within the shorter horizon. It does not supply an infinite-duration bound. Fault energy and angular impulse continue to accumulate, and constant-speed applicability becomes more demanding.

## Voltage and mechanical applicability arithmetic

The steady-dq approximation uses peak-amplitude currents and active-plus-assumed-end flux linkage:

`Vd = R Id − omega_e lambda_q`, `Vq = R Iq + omega_e lambda_d`.

The ideal sinusoidal SVPWM estimate is `Vdc = sqrt(3) hypot(Vd,Vq)`. An independent phase-domain reconstruction used the positive-sequence derivative matrix, `v_abc = R i_abc + omega_e D lambda_abc`, and `Vdc = sqrt(2) norm(v_abc)`. Across all 32 completed cases, the largest difference from the reported estimate is 5.68e−14 V.

The baseline estimates are approximately 33.80/53.73/90.72 V at 450/1,000/2,000 rpm. Adding 0.25 mH in the 2,000 rpm scenario changes the estimate to 103.47 V. The matched hot-material/hot-winding case estimates 95.48 V. These quantities expose a missing voltage/headroom condition. They are not solved phase-voltage waveforms or rigorous physical bus-voltage lower bounds: the instantaneous linkage snapshot does not provide position-harmonic derivatives, and inverter drops, modulation details and current-control limits are absent.

The mechanical diagnostic correctly integrates the rotor torque sign and uses the maximum absolute cumulative angular impulse, not just its final value:

`I_min = max_t |integral_0^t T_rotor dt| / (0.05 omega_initial)`.

It is a self-consistency estimate for the prescribed-speed trajectory with zero external torque. It is not a variable-speed fault solution or an inertia requirement for a real propulsion assembly. For example, the baseline 1,000 rpm coarse result rises from approximately 0.002014 to 0.007206 kg m2 when the reporting horizon increases from two to seven periods, despite unchanged H exposure. Propeller/load torque, reflected inertia through gearing and actual protection duration must be supplied before using fixed speed as a physical approximation.

The independent checker reproduced all 32 raw material-envelope aggregations and inertia values exactly to displayed precision, and all temperature-adjusted resistances matched their declared inputs. Artifacts: `check_final_fault_cases.py`, `FINAL_FAULT_CASE_AUDIT.json`.

## Final radial refinement and closure

The 1 mm radial magnet with the generic-steel ring at 2,000 rpm was also checked with doubled radial subdivisions while retaining the fine angular/time resolution. The H envelope changes from 669.97099 to 669.36593 kA/m (0.0904% relative to refined); the largest demagnetizing ampere-turn magnitude changes from 407.00026 to 406.71064 A-turn (0.0712%); peak phase current changes from 17.95408 to 17.94164 A (0.0693%). These all pass the frozen 3% refinement criterion. This supports the selected conditional numerical result, not the steel constitutive assumption, finite-end geometry or physical fault behavior.

The original failed no-load first step remains under `../fault/runs/sus15_2000_noload/`, including inputs, trace/log and source snapshot. Its near-zero-coil-force residual normalization was replaced by a PM-aware scale and absolute stopping floor; the corrected run is separate. The retained failure is not counted among the 32 successful cases. The documented scaling sensitivity and unchanged loaded-case reproduction support this numerical correction; no failed result was silently overwritten.

**Review closed for the bounded computational extension.** The direct equal-mean-torque calculations and copper arithmetic reject the thin radial/SUS construction under the stated equal-heat comparison, while the generic-steel opportunity remains a narrow conditional material-mass credit. The assumed doubled fault current has been replaced by circuit-coupled trajectories, with accepted radial refinements and explicit failures for precise segmented-rotor field thresholds. The physical current-to-field holdout remains inconclusive, as does a resolved conducting-ring correction under its frozen decision rule. The axial calculation is an idealized sensitivity, not a native motor correction.

No additional reviewer defect remains that requires another calculation to state these bounded findings. There is no verified lighter installed propulsion unit, qualified finished-magnet margin, or validated irreversible-demagnetization prediction. Further physical interpretation requires the declared geometry, material, control, thermal, retention and measurement inputs; computational completion does not remove those dependencies. Earlier failures, released evaluations and unaccepted numerical thresholds retain their original status.

