# Persistent 780 nm production optimization

## Live decision

The full-vector fixed-frequency atom-coupling campaign is authorized and runs
without a preset iteration count. Its current state, retained-step count,
centered-emitter response, exact invariants, Q status, and spectral progress
are published on the live dashboard.

The optimizer uses `configs/fullvector_atom_780_production_refined.json`:

- wavelength `780 nm` and normalized frequency `1.0`;
- grid `40 x 70 x 28`;
- `235,200` physical vector unknowns reduced to `29,499` by the qualified
  component-correct x/y/z symmetry sector;
- immutable one-wavelength opening `|x| < 0.5 lambda0`;
- centered z-dipole LDOS as the primary objective;
- adaptive trust-region learning rate starting at `0.004`, bounded to
  `0.00025--0.064`, with up to six halvings during backtracking.

## Persistent iteration protocol

Each topology iteration is a fail-closed direct sequence:

1. carry the exact same-grid topology variables and density only;
2. discard all saved fields and solve a fresh 780 nm primal/adjoint pair;
3. compare the exact log-LDOS gain with the gradient's predicted gain and
   require an Armijo ratio of at least `0.25`, in addition to the direct
   response, field identity, localization, PML energy, protected air,
   reflection, extrusion, frequency, and wavelength gates;
4. save the retained topology as a new immutable step directory;
5. mark the retained checkpoint **provisional** unless it is a scheduled Q
   validation step.

The trust radius doubles after a first-try accepted step whose realized versus
predicted log-gain ratio is at least `0.75`. Any failed scientific or Armijo
gate halves the proposal until it passes or reaches the declared minimum. The
requested, attempted, accepted, and next learning rates, agreement ratio, and
failed gates are retained in every step record and shown on the dashboard.
This searches for the largest locally trustworthy step instead of leaving the
campaign pinned to the `0.001` launch-gate trial.

The exact `0.995--1.005` spectrum at 81 fixed frequencies and seven independent
complex probes runs at iterations `10, 20, 30, ...`. The partially completed
step-3 sweep from the original every-step protocol is also completed and kept
as an extra validated checkpoint. The dashboard always distinguishes direct
gate passage from Q validation and reports the iteration of the last valid Q
diagnostic.

If a scheduled exact diagnostic still finds analytic background, Q is reported as
`no in-window pole`. If a physical pole candidate appears, topology updates
pause before assigning Q. The nonlinear pole and same-basis mode must then be
reacquired and validated on that checkpoint.

## Stop and restart behavior

An operator stop preserves every completed step and any partial 81-point
spectrum. Restarting resumes the first incomplete step and reuses only exact
frequency samples belonging to its declared grid. A rejected topology is not
promoted. Any nonzero protected-air density or reflection/extrusion violation
stops the campaign.

All intermediate topology checkpoints are retained. If a scheduled sweep
finds a pole, the intervening provisional checkpoints can be backfilled to
locate its first appearance before any resonance claim is made.

The optimizer process and the static-site publisher run separately, so solver
progress remains checkpointed even if documentation publication is
temporarily unavailable.

The topology close-up outlines the exact mutable x/y mask with translucent
rectangles; the one-wavelength central gap and fixed interface slices are not
inside those overlays. A separate full-device view places the retained defect
between four explicit periodic mirror cells on each side. Dashed lead
interfaces mark where the displayed defect couples to the DtN self-energies;
arrows indicate that those mirrors continue semi-infinitely in the solved
model even though only representative periods can be drawn.

## Progress through 2026-08-18

Steps 4 through 15 passed every direct acceptance gate. Step 10's scheduled Q
diagnostic completed with holdout NRMSE `5.59e-13` and no physical
lower-half-plane pole in the target window. Adaptive steps 11--15 then tested
and accepted `0.004`, `0.008`, `0.016`, `0.032`, and `0.064` on their first
trials; their predicted/realized log-gain ratios remained between `0.888` and
`1.017`. At retained step 15, the centered-emitter LDOS is `1718.5305`, a
`1.149930x` (`+14.993%`) gain over the production baseline. Same-basis field
overlap is `0.997511`, gap energy is `0.789419`, and PML energy is `0.067070`.
The protected-air, reflection-symmetry, and z-extrusion errors remain exactly
zero. Steps 11--15 remain provisional until the scheduled step-20 Q
diagnostic.

The first attempt to start step 10 stopped because the supervisor read the
optimizer's JSON status file while that file was being rewritten. This was an
operator-monitoring failure, not a solver or scientific-gate failure; step 10
had not produced or promoted a topology. Status publication is now atomic and
the supervisor retries transient JSON reads. The campaign resumes from the
intact step-9 density-only checkpoint; the recovered step-10 validation and
subsequent adaptive steps are recorded above.

## Claim boundary

This campaign searches for improved emitter coupling and the creation of a
physical 780 nm resonance. Until a pole appears and passes nonlinear,
cross-basis, grid, and domain gates, the result is not a resonant-cavity Q,
Purcell factor, useful-port beta factor, tweezer-scattering budget,
fabrication-ready structure, or device claim.