# Symmetric port-to-dipole plateau audit Date: 2026-08-30 Frozen checkpoint: `symmetric_dipole_mse_q1e4_v1`, update 260 This audit used an immutable copy of the live checkpoint. The production artifact and optimizer state were never changed. The live worker was paused only while the GPU ran each controlled probe and was then resumed from the same atomic checkpoint. ## Observed plateau The cavity is real and trustworthy, but its progress is too slow. At update 260 the conservative field/energy fit gave - Q = 601.877; - pole wavelength = 779.990 nm; - physical V = 0.202400 um3; - vector MSE = 0.0848952. Across the last 160 completed updates, Q increased by about 128 and its linear slope was only 0.783 per adjoint. At that rate a tenfold Q increase would take many thousands of additional expensive turns. The pole stayed close to 780 nm, so a branch-selection failure is not the cause. The normalized vector-MSE numerator is already dominated by the ringdown: 66.1% comes from 1.6--2.0 ps. Nevertheless, the gradient of a source-off-only MSE beginning at 1.6665 ps has cosine 0.9794 with the full-trace gradient. Temporal minibatching or merely increasing the late-time weight therefore does not supply a materially new search direction. ## Directional tests Every row below is a fresh forward FDTD replay from the same update-260 geometry. Nothing was accepted into the campaign. | perturbation | maximum density motion | Q | vector MSE | conclusion | |---|---:|---:|---:|---| | baseline | 0 | 601.877 | 0.084895 | trusted reference | | pending Adam | 0.00548 | 598.946 | 0.080969 | trace improves while Q falls | | fresh trace gradient | 0.001 | 599.477 | 0.080564 | resetting Adam alone does not fix Q | | source-off-only gradient | 0.001 | 600.602 | 0.081608 | late weighting also reduces Q | | mild 1.10 contrast sharpening | 0.0454 | 607.037 | 0.078472 | small Q gain, not a breakthrough | | beta=2 projection | 0.0537 | 624.881 | 0.700130 | modest Q gain but destroys the trace fit | | hard threshold at 0.5 | 0.499 | 216.662 | 0.883490 | immediate binarization destroys the pole | | pole-fit log-Q gradient | 0.003 | 607.120 | 0.115979 | Q rises in a direction rejected by trace MSE | The source-off and full-trace gradients are almost parallel, but the full-trace and differentiable `log(Q)` gradients have cosine **-0.9501**. This is the decisive result: at this geometry, improving the prescribed absolute time trace and improving the cavity lifetime are almost opposite local tasks. The time-trace optimizer can reduce loading-amplitude and carrier-phase errors without raising Q, which explains the plateau. Adam memory is also stale: its direction has cosine 0.1174 with the current trace gradient and 42% of pixel signs disagree. That produces visible oscillation, but it is secondary. A fresh trace-gradient replay still lowered Q, so clearing moments while retaining the same objective cannot deliver the desired Q gain. ## What pole shifting currently does—and does not do The live objective contains no shifted resonant metric. Its pole code fits once around 780 nm and repeats the fit around the first estimate with the recentering coordinate stopped from differentiation. This makes Q telemetry more robust; it does not make the fitted pole part of the optimized scalar. The Shaker--Johnson method instead evaluates a resonant response at the real part of the geometry-dependent eigenfrequency and differentiates the resulting composite objective, while constraining the pole to remain in a target frequency band. Its purpose is to remove the increasingly ill-conditioned fixed-frequency direction once Q is well above 100: The present Q near 600 is precisely in the regime where that handoff is appropriate. Calling the telemetry recentering “shifting” overstates what the current campaign implements. ## Minimal tenfold-Q successor Preserve update 260 as the discovery checkpoint and fork a pole-directed campaign; do not mutate the evidentiary trace-MSE history. The smallest useful successor is: 1. Fit the same dominant complex pole from the broadband ringdown. 2. Make the differentiated scalar `log(Q)` initially, or `log(Q/V)` if spatial localization must be protected from the first shifted step. 3. Include the geometry dependence of the fitted pole and constrain its real frequency to the physically illuminated source band around 780 nm. Merely stopping the tracking frequency is not the full shifted derivative. 4. Reset Adam moments because the new and old gradients are nearly opposite. The exact replay above establishes 0.003 maximum raw-density motion as a locally safe Q-ascent scale; an adaptive trust controller can grow it from there if forward replays are allowed. 5. Keep the existing 2 ps tape while the field/energy fits remain mutually consistent. Increase it only when fit residuals or horizon comparisons say the pole gradient is no longer resolved. 6. Delay projection. Once the shifted stage has gained roughly one decade in Q, introduce subpixel-smoothed projection gradually and let the shifted objective retune each sharpening step before eventual smooth-boundary optimization. For a time-trace-only formulation, the appropriate analogue is not a random temporal minibatch. Demodulate at the differentiable tracked pole and compare the source-off complex envelope or its one-step recurrence to the desired decay law. An amplitude/loading anchor is then required to prevent the trivial zero-field solution. Direct `log(Q/V)` is simpler and better aligned with the stated physical goal. ## Beta correction The artifact's `inferred_beta_*` arrays are not beta. They combine an amplitude projection with an uncalibrated Purcell normalization and reached values far above one. The scene has no guided-mode output-flux monitors and no closed total-loss flux measurement. The dashboard now suppresses this number entirely and reports physical beta as **not measured**. A future beta measurement needs guided TE modal power at both ports and total outgoing or decay power, with symmetry unfolding applied consistently: $$ \beta = \frac{P_{\mathrm{TE},+x}+P_{\mathrm{TE},-x}}{P_{\mathrm{total\ loss}}}. $$ The x mirror then implies equal directional contributions, but not their absolute value.