3D air-strip nanocavity

Unattended multi-step fine-grid cooperativity campaign

Persistent expanded-x production command

The current production runner has no preset topology-step limit:

scripts/run_persistent_cooperativity.sh

It is normally launched in a persistent tmux session. The bootstrap adds two complete photonic-crystal mirror periods (0.75 lambda0) per side, changing the grid from 40 x 26 x 22 to 64 x 26 x 22 and design voxels from 960 to 2,112 while leaving the slab thickness, transverse grid, air strip, PML, and mirror phase unchanged. Newly explicit mirror cells are designable.

This is a grid/domain change. The script therefore transfers geometry only, solves a new 41-point wide, 449-point dense-local, and 28-point confirmatory seven-probe spectrum, runs combined_rational_validation.py, and repeats the seven-case transverse-PML gate before the first topology step. No pole, Q, spectrum, homogeneous LDOS, or PML evidence is inherited from the smaller domain. After qualification, restartable shifted-MMA and independent spectral promotion steps continue, with full PML campaigns scheduled by the explicit adaptive policy below until an operator interrupts the run or a scientific gate stops it.

Authoritative state is under:

outputs_atom_cooperativity_xlong_persistent/bootstrap_state.json
outputs_atom_cooperativity_xlong_persistent/campaign/campaign_state.json

To stop it deliberately, send Ctrl-C to its tmux session. The controller records status: stopped, retains all partial/restart data, and can be reissued exactly. It never converts a gate failure into a retry or an accepted step.

Proven campaign command

The full shifted-MMA topology-update, spectral-promotion, and checkpoint- specific PML cycle has completed under outputs_shifted_mma_demo/step_0001. To launch 20 additional accepted steps from that fully qualified seed on the original domain, run:

scripts/run_long_cooperativity.sh 20

For a run that survives terminal or SSH disconnection:

nohup scripts/run_long_cooperativity.sh 20 \
  > outputs_atom_cooperativity_shifted_long/controller.log 2>&1 < /dev/null &
echo $! > outputs_atom_cooperativity_shifted_long/controller.pid

Monitor machine-readable state and the detailed log with:

jq . outputs_atom_cooperativity_shifted_long/campaign_state.json
tail -f outputs_atom_cooperativity_shifted_long/controller.log

The live dashboard is the remote visual monitor. campaign_state.json and each adjacent metrics.json remain authoritative for scientific acceptance.

STEPS is the total number of accepted post-seed steps desired in the output campaign, not the number to add on each invocation. Increasing it resumes the existing campaign and skips every already verified stage. It is safe to reissue after interruption. A second controller for the same output root is rejected by an advisory process lock. To start a separate campaign directory:

scripts/run_long_cooperativity.sh 20 outputs_atom_cooperativity_shifted_trial2

The first shifted cycle measured 12:08 for topology, 6:39 for pole-centered promotion, and 51:44 for the unchanged two-case-parallel PML stage on this 16-CPU host. PML time is load-sensitive and still dominates. The topology process used 1.73 GiB peak RSS. Four provisional frequency workers are now enabled after a separate worker measured 0.854 GiB peak RSS; one 29-frequency gradient benchmark took 59.84 s versus 92.48 s serial, though a repeated parallel run under different load took 90.29 s. BLAS and worker counts are fixed explicitly. PML parallelism is unchanged.

Completion contract

scripts/run_multistep_cooperativity.py runs each production topology step without operator intervention. Success means all of the following are true:

  1. the input PML-anchor metrics report multipole_pml_gate_passed: true under sheet-local-v2, and input_pml_policy.json records the anchor fingerprint and evidence age;
  2. the exact topology and configuration fingerprints match the supplied wide, local, and confirmatory data;
  3. the centered z-projected LDOS enhancement improves strictly;
  4. shifted normalized LDOS at the authoritative pole's real part improves strictly;
  5. independently fitted shared-pole Q improves strictly;
  6. atom detuning stays within two pole half-widths, or monotonically restores feasibility when a transferred input starts outside the band;
  7. the trial pole remains within the configured trust region;
  8. the full wide spectrum, dense target-local spectrum, and a freshly solved staggered confirmatory grid pass every combined spectral gate;
  9. the protected air strip has density exactly zero and reflection symmetry remains at roundoff; and
  10. the complete checkpoint, metrics, histories, restart files, plots, timing, geometry, and candidate-frequency field are saved and published;
  11. the accepted topology is reacquired on a new pole-centered 129-point local grid and untouched 28-point confirmatory grid and passes the combined gate;
  12. pml_policy.json records the scheduling interval, early-trigger observables, last PML-anchor fingerprint, and whether the full campaign ran or was deferred;
  13. whenever scheduled or triggered, its exact configuration/raw fingerprint is bound into the unchanged seven-case transverse-PML campaign and a failed gate stops the run; and
  14. deferred checkpoints remain spectrally accepted, while only a checkpoint that actually passed the full campaign is called PML converged on the controlled grid.

Adaptive PML policy

adaptive-pml-v1 begins with a full campaign after every spectrally accepted step. A full result grows confidence only when its frequency shift, complex pole shift, and Q span are at least ten times inside the authoritative gate. The interval then progresses through 1, 2, 3, 5, and at most 8 topology steps.

A full campaign is forced early by any configuration or topology-vector shape change, one-step pole motion above 0.15 half-widths, relative Q motion above 5%, or cumulative raw-topology displacement above 0.04 from the last PML anchor. A finite campaign always validates its final checkpoint. A merely passing but non-tight result resets the interval to one; a failed gate remains a scientific stop condition. The last PML result is prior evidence for the bounded continuation, never evidence that an intermediate topology itself was PML tested.

The process exits nonzero and records status: blocked if any condition fails. Rejected trials and partial spectra remain on disk; a scheduler cannot confuse a completed Python process with an accepted scientific step.

Command

The legacy single-step entry point remains available:

scripts/run_fine_cooperativity_step.sh

An explicit output directory may be supplied as its only argument:

scripts/run_fine_cooperativity_step.sh outputs_atom_cooperativity_yz_step1_retry

The single-step runner refuses to overwrite a directory that already contains metrics.json. BLAS thread counts are fixed explicitly, and every trial spectrum has a configuration/raw/frequency fingerprint. Re-running after an interruption resumes only exact matching partial files.

Fixed scientific protocol

  • fixed seed: the checkpoint-bound promotion/PML bundle under outputs_shifted_mma_demo/step_0001/qualification; every later cycle uses the immediately preceding spectrally accepted checkpoint and records the exact most recent PML anchor and evidence age;
  • atom frequency: 1.0 (the configured 780 nm normalization);
  • wide full-cluster sweep: 41 points over 0.9885--1.026;
  • target-local trial sweep: at least 129 points over nominally plus/minus two tracked-pole halfwidths;
  • confirmatory sweep: 28 prescribed staggered points excluded from fitting;
  • rational order caps: 29 and 32;
  • topology step: three provisional real-axis shifted-cooperativity/Loewner gradient refreshes inside one symmetric constrained-MMA trust box; the production seed recommends radius 0.003, bounded to 0.00025--0.008, with adaptive growth/shrink and two factor-of-two outer backtracks;
  • local MMA constraints: predicted nondecrease of fixed-frequency log(F_atom) and log(Q), plus/minus-two-halfwidth atom-detuning band, physical density bounds, exact reflection symmetry, and one cumulative outer trust box;
  • maximum pole displacement: 0.25 input-pole halfwidth;
  • minimum LDOS improvement: strictly greater than 1e-7 relative; and
  • frequency workers: four on this 16-CPU host.

The wide sweep still reacquires the entire local resonance/lead-edge cluster. Only the meromorphic target-pole continuation and Fano background are fitted inside the sheet-consistent interval; crossing a periodic-lead band-edge branch point with a one-pole analytic background is not valid.

Terminal handling

scripts/run_atom_live.sh publishes a running state before the solve. On success, scripts/finalize_atom_run.py independently reads metrics.json, requires every requested step to be accepted, rebuilds the geometry and field from the exact accepted checkpoint, and publishes the complete state. On any nonzero exit it publishes a blocked state and launches no subsequent step.

The multi-step controller automates the formerly manual qualification between steps. Spectral evidence is always checkpoint-matched. PML evidence is either fresh and checkpoint-matched or explicitly identified as an older anchor under adaptive-pml-v1; it is never represented as a fresh result for a deferred topology.

The first completed run is outputs_atom_cooperativity_yz_step1. A prior pure-LDOS proposal is retained as outputs_atom_cooperativity_yz_step1_rejected_q_decrease; it passed spectral validation but reduced Q and therefore exposed the now-corrected fail-closed Q gate.

The first accelerated complete multi-step-controller cycle is outputs_shifted_mma_demo/step_0001. Its campaign state is complete, its latest topology is PML converged on this controlled grid, and its live geometry/field snapshot is published at https://www.christianmlange.com/nanocavity-optimizer/. Grid/domain convergence remains required before any physical Q claim.