Additional evidence reproduction¶

Run from a PhotonHub source checkout. Preserve every historical file under recorded-results/. The main reproduction script creates new baseline runs; none of the commands below installs a previous pass as a new result.

Reference extraction¶

The bundled v1 and v2 reference JSONs are immutable analysis inputs. To repeat extraction, obtain the cited CC BY primary PDF separately and set CHEN_PDF to its local filename. The digitizers verify the PDF SHA256 recorded in the JSON. PyMuPDF, NumPy and Shapely are required. Fresh replay outputs stay under artifacts.

test ! -e artifacts/chen2017_curved_coupler/fig4_v2_replay.json
.venv/bin/python benchmarks/replicate/chen2017_curved_coupler/digitize_v2.py "$CHEN_PDF" --output artifacts/chen2017_curved_coupler/fig4_v2_replay.json
test ! -e artifacts/chen2017_curved_coupler/fig2_replay.json
.venv/bin/python benchmarks/replicate/chen2017_curved_coupler/digitize_fig2.py "$CHEN_PDF" > artifacts/chen2017_curved_coupler/fig2_replay.json

Straight-pair mode study¶

This recomputes modes, including material sensitivity, without FDTD. The generator rejects an existing output. The original v2 record remains archived.

OMP_NUM_THREADS=2 OPENBLAS_NUM_THREADS=2 .venv/bin/python benchmarks/replicate/chen2017_curved_coupler/straight_pair_modes.py --output benchmarks/replicate/chen2017_curved_coupler/mode_controls/straight_pair_v2.json

Endpoint source control: baseline-hash limitation¶

The archived endpoint_plan.json pins the original 30 nm result and raw input. source_endpoints.py has no alternative-plan argument and no launch command. Do not copy that historical plan beside freshly rerun baseline results or override its checks. A supported new-plan/launch workflow is required to repeat this control against a newly generated baseline. The source and historical completed control are included so this limit is reviewable.

For a checkout retaining the original matching baseline and canonical plan, the following guarded command prepares a new source input only; it does not start FDTD or recreate the completed control:

test -f benchmarks/replicate/chen2017_curved_coupler/source_controls/endpoint_plan.json
test ! -e artifacts/chen2017_curved_coupler/source30_1520_replay_prepared
OMP_NUM_THREADS=2 OPENBLAS_NUM_THREADS=2 .venv/bin/python benchmarks/replicate/chen2017_curved_coupler/source_endpoints.py prepare --label source30_1520_replay_prepared --endpoint-nm 1520

Analysis requires a completed raw execution plus a new execution receipt compatible with that audited preparation. Only then use source_endpoints.py analyze with its fresh execution label and an unused source_controls output. Never substitute a historical result or receipt for that execution.

Versioned reassessment¶

This computes statistics without FDTD or mode solves. It requires both full cumulative bundles and the verified 1520 nm endpoint result; it cannot close the fresh-baseline endpoint limitation above. It preserves original v1 fields.

test -f benchmarks/replicate/chen2017_curved_coupler/source_controls/source30_1520_run2.json
test ! -e benchmarks/replicate/chen2017_curved_coupler/reference_reassessments/fig4_v2.json
.venv/bin/python benchmarks/replicate/chen2017_curved_coupler/reassess_reference.py --output benchmarks/replicate/chen2017_curved_coupler/reference_reassessments/fig4_v2.json
.venv/bin/python benchmarks/replicate/chen2017_curved_coupler/report.py

The padding material record contains exact compile/run commands and source hashes for its macOS arm64 setup probe. It depends on the matching prepared padding input and original native library; it is not rerun by the main recipe.

Conventional GPU full-observation analysis¶

The 25 nm result, qualified 40 nm backend check, and original analysis source snapshots are archived separately. The main recipe runs local CPU baselines; it does not recreate GPU executions. Repeating the 25 nm analysis requires the original complete GPU raw fields, manifests, submitted input, execution log, backend comparison and matching 30 nm baseline at the paths recorded in the archived result. These raw outputs and GPU binary are not included in the ZIP. Keep the archived analysis snapshots as provenance; do not install historical results as fresh live outputs. In a checkout retaining those original inputs:

test ! -e artifacts/chen2017_curved_coupler/mesh25_analysis_replay.json
OMP_NUM_THREADS=2 OPENBLAS_NUM_THREADS=2 .venv/bin/python benchmarks/replicate/chen2017_curved_coupler/analyze_gpu25.py --output artifacts/chen2017_curved_coupler/mesh25_analysis_replay.json

This command recomputes the 25 nm modal readout without FDTD and rejects an existing output. It makes no new execution claim. Finer-mesh analysis snapshots and their frozen dependency lists are archived too. They depend on their exact original raw-data layout and implementation versions. Their fixed historical output paths must not be overwritten; a fresh, versioned analysis output and its matching provenance are required to repeat them. The bundle does not provide a portable GPU launch workflow or install archived results as live evidence. No launch command is included.