Frozen SiN atom-slot beta campaign

Qualified success · stopped at a robust-projection plateau

This campaign is frozen at its last complete atomic transaction. It produced useful, fully binary single-etch designs and a trustworthy common-ledger beta near 81%, but no mask passed the complete erosion/nominal/dilation audit. The 90% robust-beta handoff was never claimed.

best provisional β81.331%η=0.45 · common outer ledger

best provisional Q207.0field/energy conservative fit

best provisional βFₚ0.0816not robustly promoted

escape work11 / 173adjoints / forward validations

robust entries0η=0.45 / 0.50 / 0.55

Device and fixed parameters

The 780.24 nm z-oriented atom sits at the end of a 1 µm-wide air-access slot, with no SiN within 500 nm. The free-form mask is two-dimensional and extruded through a 500 nm SiN film. A single 400 nm right-hand guide continues through the PML; x symmetry is therefore unavailable. Electric parity in y and z uses (0,+1,-1), reducing the 360 × 208 × 80 full grid by 4×. The design grid is 40 nm, the conic fabrication radius is 50 nm, and the output mode has n_eff=1.77778.

parameter	value
target wavelength	780.24 nm
film	SiN, 500 nm
atom clearance	500 nm radial
access slot	1.0 µm wide
output guide	400 nm wide, +x only
full / simulated cells	(np.int64(360), np.int64(208), np.int64(80)) / 4
trainable reduced-plane pixels	17,844
projection continuation	β=2 → 4 → 8 → 16 → 32 → 40 → 48 → 64 → 96
robust thresholds	η=0.45, 0.50, 0.55

What was actually tried

1. Cold ordinary beta optimization. Neutral density was filtered and projected without a cavity warm start. The objective maximized right-guide beta while keeping vacuum-normalized LDOS above 0.5 and penalizing material disconnected from the output. It ran 165 adjoints through exact binarization. Its raw terminal beta of 92.21% is historical only: it used the earlier channel accounting whose inconsistency motivated the stricter shifted replay.

2. Independent pole acquisition. Long ringdowns tracked field and stored- energy decay. A clean Q≈162 pole near 781.47 nm triggered the shifted formulation rather than pretending that the short ordinary tape remained adequate.

3. Common-ledger shifted beta. Guided TM, -x, residual +x, ±y, and ±z power were put on one outer surface ledger, including the corrected 1/8 pulse-mode normalization. The emission guard was removed as requested. Adam, momentum SGD, and L-BFGS were compared; L-BFGS won the compact equal-budget audit. The terminal v1 replay has beta=80.84%, Q=149.6, and 1.53% spectral inner/outer mismatch.

4. Transactional Pareto escape. Separate beta, Q, and guided-Purcell islands shared a nondominated archive. Every adjoint was followed by exact-binary 16/32/64/128-pixel line-search trials. Seven correlated restart screens tested 0.10/0.20/0.40/0.80 µm perturbations. Controls and L-BFGS curvature advanced together only after a trustworthy η=0.45 forward.

5. Robust promotion and intended Stage 2. Every selected candidate was audited at η=0.45, 0.50, and 0.55. Stage 2 would have maximized guided Purcell with an augmented-Lagrangian beta≥90% constraint. It never started: all candidates retained nominal/dilation pole or connectivity debt.

Histories and final Pareto set

The final archive contains 10 provisional masks. Its extrema are:

objective extreme	β	Q	Purcell	guided Purcell	wavelength
beta	81.331%	142.84	0.001170	0.000952	781.051 nm
Q	79.860%	206.98	0.000969	0.000774	782.499 nm
guided Purcell	78.794%	185.45	0.103500	0.081551	777.821 nm

The large provisional guided-Purcell jump is included because it was executed, but it is not promoted as a final device. It failed the nominal/dilated audit, as did every other archive entry. The qualified success is the demonstrated binary workflow, corrected bounded beta, automatic pole handoff, and preserved Pareto alternatives—not completion of the original >90% robust target.

Reproducibility and immutable evidence

artifact	SHA-256	bytes
benchmarks/artifacts/invdes_sin_slot_atom_beta_cold_v1.npz	5322c4611e01e0f0cffbdeb2e520639e25016fd2c055a0f12fb1bc0fb5b609f3	712,444
benchmarks/artifacts/invdes_sin_slot_atom_beta_shifted_v1.npz	a0cad55913fda10b20569c3f5b7ddb83f94673e8135d62da0567c6e8fe487f0b	798,027
benchmarks/artifacts/invdes_sin_slot_atom_beta_escape_v2.npz	d0bb8237ebe129e10e77471c67678ae4b414f13e651cd109a9e65f72fe8fa79a	1,525,253
benchmarks/artifacts/invdes_sin_slot_atom_beta_escape_v2.qualified_success_20260823.npz	d0bb8237ebe129e10e77471c67678ae4b414f13e651cd109a9e65f72fe8fa79a	1,525,253
benchmarks/artifacts/sin_slot_atom_beta_campaign_archive_v1.npz	6428a4398d6737a1dceb1b1b8edc453b66e63c88fc977fad64ad7dba5f6971d2	1,760,582

The evidence bundle replays the neutral start and all three Pareto extrema, then stores each material mask with its paired FDTD Ez field. The optimizer was stopped before this archive was generated; no pending proposal is shown as an accepted result.

Frozen: 2026-08-23T00:15:35.530769+00:00
Reason: pareto_archive_plateau_without_three_threshold_robust_promotion