Detector-Noise Survival of Kerr Quadrature Memory

Status

The Ledger 13 coherent-quadrature memory lead survives the predeclared normalized detector-noise gate. At detector-noise standard deviation (10^{-8}), both (L_1(u_{t-3})) and (L_1(u_{t-4})) remain family-wise significant in the Kerr case across seeds 20260810, 20260811, and 20260812, and their Kerr-minus-disabled corrected-capacity differences remain positive in every seed. Their signed cross-seed lower envelopes are respectively 0.003779 and 0.014251.

This is a local normalized-model result, not detector calibration and not evidence of nonlinear computation. No degree-two or degree-three target replicates at the decision floor. The empirical gate for Conjecture 5 therefore remains closed.

The correct decision is to retain the Kerr-modified linear-memory lead and advance it to a bounded parameter-region siege. That next search must remain distinct from a claim of physical detector feasibility.

Predeclared falsification protocol

The protocol was frozen before the evidence-producing run:

Quantity Frozen value
Observation coherent real and imaginary bus-field quadratures
Detector-noise standard deviations (0,10^{-10},10^{-9},10^{-8},10^{-7})
Decision floor (10^{-8})
Seeds 20260810, 20260811, 20260812
Samples per seed 2400 after 128 warm-up symbols
Target family 55 Legendre-history targets, degree at most 3 and lag at most 4
Training fraction 0.7
Joint permutations 512
Family-wise quantile 0.99
Replication rule target passes the family-wise gate in every seed
Paired-advantage rule Kerr target replicates and Kerr-minus-disabled is positive in every seed

The (10^{-8}) decision floor was chosen because Ledger 13 measured the weakest Kerr quadrature feature standard deviation as approximately (3.4\times10^{-9}). The grid contains levels below and above that feature scale. Results at (10^{-7}) are retained as descriptive severity evidence; they did not move the predeclared decision boundary.

All detector-noise values are in normalized real-quadrature units. No dimensional LiTaO3 detector, local oscillator, bandwidth, ADC, shot-noise, or quantization model is inferred from them.

Observation-boundary implementation

noise_std remains dynamical state noise. The new detector_noise_std is independent configuration state and is applied only after the noiseless bus output

[ s_{\mathrm{out}}=s_{\mathrm{in}}-\sqrt{\kappa_{\mathrm{external}}},a ]

has been converted into the declared real observation interface. The noiseless observation is retained for diagnostics, while only the noisy observation is supplied to the readout.

Input, dynamical-noise, detector-noise, and permutation streams have distinct deterministic seed derivations. Kerr and Kerr-disabled cases with the same seed and detector-noise level receive the same detector standard-normal draws. Changing detector noise leaves input symbols, state power, thermal state, and noiseless observations unchanged.

Focused tests establish reproducibility, common random numbers, observation-boundary placement, state/detector RNG separation, and exact zero-detector-noise backward behavior.

Signal, SNR, and readout amplification

For feature (j), the signal scale (s_j) is the standard deviation of the noiseless feature on the training partition only. With declared detector-noise standard deviation (\sigma_d), the reported linear power SNR and decibel SNR are

[ \operatorname{SNR}j=\frac{s_j^2}{\sigma_d^2}, \qquad \operatorname{SNR}{j,\mathrm{dB}}=10\log_{10}\operatorname{SNR}_j. ]

Zero detector noise is represented explicitly as infinite SNR for nonconstant signals and as an undefined zero-over-zero case for constant signals; non-finite floating-point values are not serialized. The report also records the realized training-sample RMS of the injected noise. Across all nonzero cases, realized-to-declared RMS lies in [0.9666, 1.0325], with mean 0.9962.

If a readout fitted in standardized coordinates has weights (w_j) and observed training scales (r_j), its raw-equivalent norm and detector-noise gain are

[ \lVert w_{\mathrm{raw}}\rVert_2 =\sqrt{\sum_{j:r_j>0}\left(\frac{w_j}{r_j}\right)^2}, \qquad g_d=\sigma_d\lVert w_{\mathrm{raw}}\rVert_2. ]

The standardized norm (\lVert w\rVert_2) is also reported. A degenerate feature is set to zero during training; raw conversion would be marked undefined if it nevertheless acquired a nonzero standardized weight. No undefined conversion occurred in the frozen suite.

At the (10^{-8}) gate, across the six Kerr lag-three/four readouts:

Readout quantity Minimum Maximum Mean
Standardized weight norm 6.523 15.305 11.944
Raw-equivalent weight norm (1.244\times10^6) (4.939\times10^6) (2.775\times10^6)
Detector-noise gain 0.01244 0.04939 0.02775
Corrected target capacity 0.01203 0.05129 0.02793

The weights remain large in raw normalized coordinates. Passing the normalized noise test does not remove this engineering concern.

Noiseless rank and noise-aware observable dimension

Three notions are kept separate:

  1. observed standardized numerical rank, which becomes full when nonzero independent noise is standardized;
  2. noiseless standardized numerical rank at the configured relative singular tolerance (10^{-6}), which records ideal mathematical span;
  3. noise-aware observable dimension.

At zero detector noise, the noise-aware dimension is defined to equal noiseless numerical rank. At nonzero noise it counts noiseless raw principal standard deviations strictly greater than (\sigma_d), capped by noiseless numerical rank. The strict threshold boundary and zero-noise case are unit-tested. Every noiseless standardized singular value and every noiseless raw principal standard deviation is serialized so the count is auditable.

Across all three seeds, the dimensions are:

Detector-noise std Kerr noiseless rank Kerr noise-aware dimension Disabled noiseless rank Disabled noise-aware dimension
0 14 14 4 4
(10^{-10}) 14 14 4 4
(10^{-9}) 14 13 4 4
(10^{-8}) 14 12 4 4
(10^{-7}) 14 9 4 4

At the decision floor, the smallest Kerr per-feature signal scale lies between (3.01\times10^{-9}) and (3.40\times10^{-9}) across seeds, giving a worst per-feature SNR between -10.44 dB and -9.38 dB. The two below-floor features span -10.44 dB to -7.34 dB across the three seeds. Twelve of fourteen Kerr features individually exceed the noise standard deviation, and the principal-axis criterion also yields dimension 12. The strongest feature scale remains approximately 0.09. The disabled system has only four noiseless directions; its remaining features are constant before detector noise.

Paired capacity outcome

The Kerr-minus-disabled total and historical-capacity differences remain positive at every level and seed:

Detector-noise std Total difference range Total mean Historical difference range Historical mean
0 [0.015012, 0.020287] 0.017792 [0.015171, 0.020375] 0.017836
(10^{-10}) [0.036568, 0.065679] 0.054970 [0.035121, 0.063522] 0.052767
(10^{-9}) [0.028408, 0.057255] 0.041580 [0.027235, 0.055301] 0.039651
(10^{-8}) [0.022500, 0.039771] 0.033830 [0.022453, 0.037882] 0.032324
(10^{-7}) [0.018369, 0.035495] 0.028876 [0.018564, 0.034316] 0.027954

The two predeclared delayed targets have the following cross-seed outcomes:

Noise std Target Kerr significant seeds Disabled significant seeds Positive paired deltas Signed lower envelope Mean delta
0 (L_1(u_{t-3})) 3/3 3/3 3/3 0.006939 0.008130
0 (L_1(u_{t-4})) 3/3 3/3 3/3 0.010641 0.011823
(10^{-10}) (L_1(u_{t-3})) 3/3 3/3 3/3 0.009115 0.011521
(10^{-10}) (L_1(u_{t-4})) 3/3 1/3 3/3 0.021863 0.029051
(10^{-9}) (L_1(u_{t-3})) 3/3 3/3 3/3 0.006079 0.007287
(10^{-9}) (L_1(u_{t-4})) 3/3 1/3 3/3 0.015147 0.024363
(10^{-8}) (L_1(u_{t-3})) 3/3 3/3 3/3 0.003779 0.005831
(10^{-8}) (L_1(u_{t-4})) 3/3 1/3 3/3 0.014251 0.021342
(10^{-7}) (L_1(u_{t-3})) 3/3 3/3 3/3 0.003640 0.006318
(10^{-7}) (L_1(u_{t-4})) 3/3 1/3 3/3 0.014924 0.019287

At the (10^{-8}) gate, the only Kerr targets that replicate are the current input and the two declared delayed linear targets. No nonlinear target replicates. Detector noise therefore preserves the local Kerr-associated redistribution of linear memory; it does not uncover nonlinear delayed capacity. Lag three also exists significantly in the disabled control in 3/3 seeds, and lag four in 1/3 control seeds. The earned result is thus the replicated positive Kerr-minus-disabled advantage, not unique existence of delayed memory under Kerr.

Canonical SQLite evidence

Schema version 2 stores the result transactionally in packages/kerr-capacity/output/detector-noise-frozen/results.sqlite. The output directory remains an ignored reproducible artifact directory; the ledger and frozen configuration are Git-visible.

Normalized relations contain:

Relation Rows
Noise levels 5
Normalized observation-noise cases 30
Resource accounts 30
Target/readout records 1650
Singular-value records 420
Feature-scale records 420
Signal/SNR diagnostic records 420
Paired target differences 825
Cross-seed replication outcomes 275

SQLite reports integrity: ok and zero foreign-key violations. Derived JSON, CSV, and Markdown exports are produced from the same in-memory suite. Selected SHA-256 hashes from the evidence run are:

Artifact SHA-256
results.sqlite 4A1A403E23737CCFBA169B1F15198865BC37AB288102A5A8427F3D43749F6315
noise-suite.json E2FFD54D9ACDBA49EF066E3E2D4E60FE6B85822A48D098AFA8E7922D237B6D98
noise-replication.csv 454A6CF56A701F292D4EECE7D781559B08051B14AD1D1C25A267307EB40BC732

Deutsch–Popperian critique

The local conjecture under test was not that Kerr memory is generally robust. It was the narrower claim that the Ledger 13 lag-three/four quadrature advantage would survive a predeclared normalized independent observation-noise floor at one frozen parameter point. The experiment exposed that claim to failure and it survived.

That survival identifies new errors to attack rather than supplying confirmation:

  1. The noise has no detector physics. It is independent additive Gaussian noise in normalized quadrature coordinates. Correlation, phase dependence, local-oscillator limits, finite bandwidth, quantization, drift, and shot noise remain unmodeled.
  2. Raw readout amplification remains severe. Gate-target raw-equivalent norms are between (1.24\times10^6) and (4.94\times10^6). The measured noise gains remain finite in this model, but no actuator, estimator, or electronic dynamic-range constraint has been imposed.
  3. Any nonzero noise changes the standardized control design. Ten exactly constant Kerr-disabled features become pure-noise columns and are standardized to unit variance. Its observed standardized rank therefore jumps from 4 to 14 even though noiseless and noise-aware ranks correctly remain 4. The sharp capacity change between zero and (10^{-10}) is consequently an estimator/readout effect, not a physical enhancement. The paired design remains fair, but the magnitude of the noisy Kerr advantage is not monotone evidence of stronger physics.
  4. The observable-dimension rule is declared, not derived from an instrument. A one-standard- deviation principal-axis floor is an auditable robustness criterion, not a universal definition of observability.
  5. Replication is still local. Three seeds test sampling variation at one dynamical point. They do not establish a connected parameter region, fabrication tolerance, or comparison against a resource-matched linear cascade.
  6. Only one level was inferentially decisive. The (10^{-8}) gate was fixed in advance. The remaining levels map severity and must not be mined post hoc for a more favorable threshold.

The next severe test should therefore seek a connected region in normalized pump, detuning, Kerr strength, and symbol duration while retaining the same paired noise and replication rules. It should add a declared readout-gain constraint and a resource-matched linear or cascaded control. The siege should try to destroy the lead, not optimize a single successful point.

Verification

The implementation passed:

  • cargo fmt --check;
  • cargo check --all-targets;
  • cargo test --all-targets with 31 passing tests;
  • direct/pseudospectral full-RHS cross-check error (5.389158\times10^{-16}), below (10^{-9});
  • the release-mode frozen noise-suite command;
  • SQLite integrity and foreign-key checks;
  • git diff --check.

Decision

The predeclared detector-noise gate is passed. The Kerr-associated lag-three/four coherent linear memory survives normalized detector noise at (10^{-8}) across all three frozen seeds, while the noise-aware Kerr observation dimension falls from 14 to 12.

Advance the lead to a bounded categorical parameter siege, with readout-gain and matched-linear controls carried forward. Do not claim physical detector viability. Do not formulate Conjecture 5: the required replicated nonlinear delayed capacity remains zero.