Planar Capacitance Model Space (QCap, QCap-Lite, QCap-Fast) from registry v2 (earlier contents: 906bc45ae05c)
86e5ae4 verified Download registry.yaml from shanto268/PlanarCapacitanceModel: direct link, hf CLI and curl.
- Browser
- Download file 138 kB
-
https://huggingface.co/spaces/shanto268/PlanarCapacitanceModel/resolve/main/registry.yaml
- Command line
-
hf download hf://spaces/shanto268/PlanarCapacitanceModel/registry.yaml
-
curl -L -o registry.yaml https://huggingface.co/spaces/shanto268/PlanarCapacitanceModel/resolve/main/registry.yaml
138 kB
# Single source of truth for every released name, link and paper number of the tl-squadds capacitance models.
# Everything else (tlsq_infer model loader, Space apps, model cards, paper_ml/deploy_links.tex, the sync workflow)
# reads this file. Renaming a repo or Space = edit one line here, then run `python deploy/sync.py` (dry run) /
# `python deploy/sync.py --push`.
#
# Paths are relative to the repository root. Metrics: median |relative error| (%) / MAE (fF) / RMSE (fF) over all
# mutual entries, as in the paper (paper_ml/floats/tab_candidates.tex, paper_ml/tables/cost_rows.tex).
version: 2 # 2: corrected prior with substrate far field (2026-10-20, agents/farfield/ADOPTION.md); 1: before
package_version: "0.1.0"
hub:
org: SQuADDS
private: true # create every repo private; flip deliberately when the paper is public
space_hardware: cpu-basic # free tier: 2 vCPU, 16 GB RAM
flagship_space: shanto268/PlanarCapacitanceModel # org Gradio Spaces need a paid HF plan; move to SQuADDS/ by editing this line
# ("flagship" here = the default Space; the tiers are "coarse" / "fine"; since
# 2026-10-10 the fine tier runs on the "flagship prior", PRIOR_SCHEME.md Part II)
flagship_title: Planar Capacitance Model
paper_models_space: shanto268/2026PaperModels
paper_models_title: 2026 Paper Models
collection: "Zero-shot planar capacitance models (2026)"
collection_slug: null # filled in by sync.py --push after the collection is created
code_url: null # public source repository of the training code, when released
paper:
short: "Zero-shot capacitance of planar superconducting layouts"
draft_dir: paper_ml
links_tex: paper_ml/deploy_links.tex
arxiv: null # set when posted; cards and Spaces pick it up
training_stack:
substrate: Si (cryogenic)
eps_r: 11.45
thickness_um: 350
metal: zero-thickness sheet (Q3D thin conductor)
layers: {ground: [1, 0], conductor: [1, 10]}
training_designs: 40153
training_families: [interdigital (CapNInterdigital), paddle, concentric, star, multi-pad star]
eval_sets:
never_seen:
name: never-seen components (Xmon, TransmonPocket, shielded concentric)
designs: 926
entries: all mutual entries
ood60:
name: OOD-60 (procedural near / intermediate / alien shapes)
designs: 60
blend:
weights_gnn_side: {signal-signal: 0.6, signal-ground: 0.7, pad: 0.8} # log-space; rest = charge-corrected prior
note: "pad = any entry involving the third or later signal body (file order), as in the paper"
kinds: invariant # deployment: order-invariant kinds (deploy/MODEL_SELECTION.md amendment 3); "published" = paper file-order kinds
weights_gnn_side_invariant: {conductor-conductor: 0.6, conductor-ground: 0.9} # LOFO on the 5 training families (scripts/blend_lofo.py --kinds invariant)
note_invariant: "conductor-conductor = any pair of floating conductors, conductor-ground = pairs with ground; independent of polygon order"
self_capacitance:
row_sum_eps: 0.0143 # C_ii = sum_j |C_ij| / (1 - eps) (IMPROVE_IDEAS.md, training median)
substrates: # dropdown of the Spaces; the 11 materials of the stack DoE + the training substrate
Si (cryogenic, training): 11.45
sapphire (c-plane, isotropic eq.): 10.34
GaAs: 12.9
fused silica: 3.8
Si (room temperature): 11.7
SiC: 9.7
MgO: 9.8
InP: 12.4
Ge: 16.0
quartz (crystalline): 4.5
diamond: 5.7
LaAlO3: 24.0
artifacts:
charge_gbm: deploy/artifacts/charge_gbm.joblib # built once by `python -m tlsq_infer.charge fit`
charge_gbm_train_cells: data_local/charge_p1/train_cells.parquet
stack_error: deploy/artifacts/stack_error.json # built by `python deploy/stack_error.py fit`
# per-entry error model (ERROR_MODEL.md): NOT ENABLED in the release (user decision 2026-10-08: no intervals until the
# pre-declared v2 is evaluated on the fresh atlas + >7-conductor labels). To enable, set e.g.
# error_model: {all-data-v3s: error_model_fast.joblib, all-data-v3-rich: error_model_accurate.joblib,
# pair-local-fast: error_model_pair-local-fast.joblib}
# (files in deploy/artifacts/, built by `python scripts/error_model.py fit`; tests: deploy/tests/test_error_model.py)
error_model: {}
defaults:
flagship_model: all-data-lite1500 # fast tier = flagship-fast, DEFAULT since 2026-10-10 (user decision; deploy/MODEL_SELECTION.md Results 4)
coarse_model: all-data-v3s # coarse tier (default until 2026-10-10; deploy/MODEL_SELECTION.md, amendment 1 rule)
flagship_accurate_model: all-data-flagship # fine tier (key name kept). 2026-10-10, user-approved: all-data-v3-rich -> all-data-flagship
# (same weights, flagship prior; PRIOR_SCHEME.md section 8.1, deploy/MODEL_SELECTION.md Results 3)
large_layout_model: all-data-flagship # (PRIOR_SCHEME.md Part II): layouts on which the capped coarse prior would
# resolve < 0.75 panels across the smallest gap (outside the training regime; the grids of
# GRID_SCALING.md returned zero entries) are predicted with this flagship-prior model instead
# (predict(large_layouts="route"), reported in model_used / warnings). In-regime layouts unchanged.
paper_space_model: pair-local-fast # this work's headline (paper abstract)
tiers: # shown in the flagship UI; keys = tier names (alias accurate -> fine accepted by predict);
fast: {latency_s: 1, median_pct: 2.67, mae_ff: 0.768, rmse_ff: 1.874, set: "212 held-out designs (OOD-60, random TransmonPocket-3, 6 new Qiskit Metal classes)",
note: "flagship-fast (default): cheap micro-MoM prior (<= 1,500 merged panels; core-graded windows and screened-dipole far pairs beyond 8 conductors) and a network retrained on it. Less accurate than the coarse and fine tiers (large-N 40: 4.96 % vs 2.93 % / 1.63 %), several times faster, and the only tier that runs multi-qubit chips. On this free CPU Space (2026-10-10): 0.2-0.3 s for the Xmon and TransmonPocket examples, 8.5 s for a 200-finger comb, 14 s for a 66-qubit-grid layout, about 4 min for a 780-conductor chip (18 Mac cores: 15 s). PRIOR_SCHEME.md Part III section 15.3."}
# held-out = OOD-60 + pocket3 random-36 + Qiskit Metal new 116, pooled mutual entries
coarse: {latency_s: 4, median_pct: 1.75, mae_ff: 0.534, rmse_ff: 1.173, set: "212 held-out designs (OOD-60, random TransmonPocket-3, 6 new Qiskit Metal classes)",
note: "On this free CPU Space: 1-5 s for single components (measured on the examples, 2026-10-10). Layouts whose coarse mesh would fall below the training resolution (e.g. the TransmonPocket example, long combs) are predicted with the fine tier instead (reported under model used)."}
fine: {latency_s: 10, median_pct: 1.13, mae_ff: 0.374, rmse_ff: 0.836, set: "212 held-out designs (OOD-60, random TransmonPocket-3, 6 new Qiskit Metal classes)",
note: "Flagship prior: strip-merged Richardson (6 and 3 panels across the smallest gap on every layout), H-matrix solve for large systems. On this free CPU Space (2026-10-10): 1-2 s for the Xmon and interdigital examples, 20 s for the TransmonPocket, 80 s for the shielded concentric example; large layouts several minutes (200-finger comb 220 s). One fine-tier solve runs at a time; other requests are not blocked."}
models:
# ---------------------------------------------------------------- learned, released
pair-local:
title: "Pair-local network + earlier standard MoM prior (R=0, published v2 mesh)"
repo_id: SQuADDS/2026PaperModels-pair-local
kind: gnn
prior: standard
rounds: 0
status: available
checkpoints: [data_local/mega/ckpt/ALL_mom_r0_s0.pt, data_local/mega/ckpt/ALL_mom_r0_s1.pt, data_local/mega/ckpt/ALL_mom_r0_s2.pt]
parameters: 209410
summary: "Earlier prior (published v2 mesh). Former flagship; superseded by the coarse-prior models of this work."
paper:
appears_in: [tab:candidates, tab:costacc, fig:costacc, tab:splits, fig:zs-arch, fig:stack, tab:stack]
never_seen: {median_pct: 1.73, mae_ff: 0.83, rmse_ff: 1.62}
ood60: {median_pct: 1.81, mae_ff: 0.78, rmse_ff: 1.63}
in_family_id: {median_pct: 0.91, mae_ff: 0.063, rmse_ff: 0.181, note: grouped in-distribution split, signal-signal}
cost_s_per_design: {never_seen: 8.75, ood60: 0.60, note: one CPU core, all stages from the GDS}
message-passing:
title: "Message-passing network + earlier standard MoM prior (R=3, published v2 mesh)"
repo_id: SQuADDS/2026PaperModels-message-passing
kind: gnn
prior: standard
rounds: 3
status: available
checkpoints: [data_local/mega/ckpt/ALL_mom_s0.pt, data_local/mega/ckpt/ALL_mom_s1.pt, data_local/mega/ckpt/ALL_mom_s2.pt]
parameters: 752386
summary: "Earlier prior (published v2 mesh). Transfers slightly worse than pair-local."
paper:
appears_in: [tab:candidates, tab:costacc, fig:costacc, tab:sota, fig:zs-arch, fig:infamily]
never_seen: {median_pct: 2.08, mae_ff: 0.97, rmse_ff: 1.88}
ood60: {median_pct: 2.45, mae_ff: 0.98, rmse_ff: 1.94}
in_family_full: {median_pct: "0.21-1.05", mae_ff: "0.017-0.40", rmse_ff: "0.022-0.84", note: C_NS, range over the 5 families}
cost_s_per_design: {never_seen: 8.77, ood60: 0.61}
charge-corrected-prior:
title: "Charge-corrected earlier MoM prior (cell GBM, published v2 mesh)"
repo_id: SQuADDS/2026PaperModels-charge-corrected-prior
kind: charge_prior
prior: standard
status: available
checkpoints: [deploy/artifacts/charge_gbm.joblib]
summary: "Earlier prior (published v2 mesh): MoM cell charges corrected by a gradient-boosted model of local geometry; no capacitance labels."
paper:
appears_in: [tab:candidates, tab:costacc, fig:zs-cell, tab:splits]
never_seen: {median_pct: 2.18, mae_ff: 0.95, rmse_ff: 1.62}
ood60: {median_pct: 4.05, mae_ff: 1.30, rmse_ff: 2.33}
cost_s_per_design: {never_seen: 17.5, ood60: 1.06, note: "paper timing re-solves the MoM; tlsq_infer reuses the prior's solve"}
blend-pair-local:
title: "Blended model (earlier prior): pair-local network x charge-corrected earlier prior"
repo_id: SQuADDS/2026PaperModels-blend-pair-local
kind: blend
components: [pair-local, charge-corrected-prior]
status: available
summary: "Earlier prior (published v2 mesh); was the headline of the first draft. Deployed with order-invariant blend weights."
paper:
appears_in: [tab:candidates, tab:costacc, fig:costacc, fig:blend, fig:stack, tab:stack, fig:invdesign, tab:invdes]
never_seen: {median_pct: 1.64, mae_ff: 0.71, rmse_ff: 1.36}
ood60: {median_pct: 2.36, mae_ff: 0.87, rmse_ff: 1.75}
cost_s_per_design: {never_seen: 17.5, ood60: 1.06}
blend-message-passing:
title: "Blended model (earlier prior): message-passing network x charge-corrected earlier prior"
repo_id: SQuADDS/2026PaperModels-blend-message-passing
kind: blend
components: [message-passing, charge-corrected-prior]
status: available
summary: "Earlier prior (published v2 mesh); blend with the R=3 network."
paper:
appears_in: [tab:costacc, fig:costacc, fig:blend]
never_seen: {median_pct: 1.84, mae_ff: 0.82, rmse_ff: 1.56}
ood60: {median_pct: 2.37, mae_ff: 0.90, rmse_ff: 1.87}
cost_s_per_design: {never_seen: 17.5, ood60: 1.07}
geometry-only:
title: "Geometry-only network (no physics prior; earlier ablation)"
repo_id: SQuADDS/2026PaperModels-geometry-only
kind: gnn
prior: none
rounds: 3
status: available
checkpoints: [data_local/mega/ckpt/ALL_none_s0.pt]
summary: "Ablation. Accurate in family, fails on unseen geometry: do NOT use for new layouts. Released for reproducibility (seed 0 of 3)."
warning: "Not for unseen geometry (19 % median on never-seen components, single entries off by >10x)."
paper:
appears_in: [tab:candidates, tab:splits, tab:sota]
never_seen: {median_pct: 19.4, mae_ff: 6.1, rmse_ff: 9.8}
ood60: {median_pct: "4.7-15.0", mae_ff: "1.7-3.4", rmse_ff: "2.2-4.6", note: range of per-class medians}
in_family_full: {median_pct: "0.45-1.54", mae_ff: "0.011-0.90", rmse_ff: "0.015-1.32"}
# ---------------------------------------------------------------- training-free
mom-prior:
title: "Earlier standard MoM prior alone (training-free, published v2 mesh)"
repo_id: null
kind: prior
prior: standard
status: available
summary: "Untrained thin-film method of moments (default 8,000-cell cap). Physics baseline."
paper:
appears_in: [tab:candidates, tab:costacc, tab:prior, fig:prior-parity]
never_seen: {median_pct: 6.55, mae_ff: 3.03, rmse_ff: 4.62}
ood60: {median_pct: 8.09, mae_ff: 2.26, rmse_ff: 3.56}
fine-mom-prior:
title: "Earlier 16k MoM prior alone (training-free, published chip-anchored v2 mesh, 16k cells)"
repo_id: null
kind: prior
prior: fine
status: available
summary: "Untrained MoM, gap/6 with a default 16,000-cell cap. Physics baseline at ~3-5x the cost."
paper:
appears_in: [tab:costacc]
never_seen: {median_pct: 4.99, mae_ff: 2.38, rmse_ff: 3.60}
# ---------------------------------------------------------------- pending (trained on HPC; copy checkpoints, flip status)
pair-local-fine:
title: "Pair-local network + earlier 16k MoM prior (R=0, published chip-anchored v2 mesh, 16k cells)"
repo_id: SQuADDS/2026PaperModels-pair-local-fine
kind: gnn
prior: fine
rounds: 0
status: available
checkpoints: [data_local/mega/ckpt_fine/ALL_mom_r0_s0.pt, data_local/mega/ckpt_fine/ALL_mom_r0_s1.pt, data_local/mega/ckpt_fine/ALL_mom_r0_s2.pt]
summary: "Earlier 16k prior (published chip-anchored v2 mesh, 16k cells; was called the earlier fine prior). Trained on 35,283 designs (star families subsampled to 300 each)."
paper:
appears_in: [tab:candidates, tab:costacc, fig:costacc, tab:wonky-shape-main]
never_seen: {median_pct: 1.74, mae_ff: 0.98, rmse_ff: 1.87}
ood60: {median_pct: 0.95, mae_ff: 0.41, rmse_ff: 0.75}
cost_s_per_design: {never_seen: 38.0, ood60: 1.84}
message-passing-fine:
title: "Message-passing network + earlier 16k MoM prior (R=3, published chip-anchored v2 mesh, 16k cells)"
repo_id: SQuADDS/2026PaperModels-message-passing-fine
kind: gnn
prior: fine
rounds: 3
status: available
checkpoints: [data_local/mega/ckpt_fine/ALL_mom_s0.pt, data_local/mega/ckpt_fine/ALL_mom_s1.pt, data_local/mega/ckpt_fine/ALL_mom_s2.pt]
summary: "Earlier 16k prior (published chip-anchored v2 mesh, 16k cells; was called the earlier fine prior). Trained on 35,283 designs (star families subsampled to 300 each)."
paper:
appears_in: [tab:candidates, tab:costacc, fig:costacc]
never_seen: {median_pct: 1.74, mae_ff: 0.74, rmse_ff: 1.39}
ood60: {median_pct: 1.11, mae_ff: 0.47, rmse_ff: 0.91}
in_family_full: {median_pct: "0.18-1.04", mae_ff: "0.015-0.30", rmse_ff: "0.020-0.46"}
cost_s_per_design: {never_seen: 38.0, ood60: 1.85}
all-data:
title: "All-data pair-local network + earlier 16k MoM prior (R=0, published chip-anchored v2 mesh)"
repo_id: SQuADDS/2026PaperModels-all-data
kind: gnn
prior: fine
rounds: 0
status: available
checkpoints: [data_local/mega/alldata/fine_r0_s0.pt, data_local/mega/alldata/fine_r0_s1.pt, data_local/mega/alldata/fine_r0_s2.pt]
summary: "Trained on every labelled design (no held-out family). Not zero-shot on any paper test set. Candidate E of deploy/MODEL_SELECTION.md."
validation: {median_pct: 2.25, mae_ff: 0.778, rmse_ff: 2.27, set: "212 deploy held-out designs (deploy/MODEL_SELECTION.md)"}
paper:
appears_in: [sec:res-opt (All-data model paragraph)]
all-data-base:
title: "All-data pair-local network + earlier standard MoM prior (R=0)"
repo_id: SQuADDS/2026PaperModels-all-data-base
kind: gnn
prior: standard
rounds: 0
status: available
checkpoints: [data_local/mega/alldata/base_r0_s0.pt, data_local/mega/alldata/base_r0_s1.pt, data_local/mega/alldata/base_r0_s2.pt]
summary: "Trained on every labelled design, earlier standard 8k-cell MoM prior. Safety fallback of the coarse tier if its v3 prior cannot be solved (not needed on any test layout since the non-positive-definite fix). Candidate D of deploy/MODEL_SELECTION.md."
validation: {median_pct: 2.77, mae_ff: 0.96, rmse_ff: 2.657, set: "212 deploy held-out designs (deploy/MODEL_SELECTION.md)"}
all-data-v3s:
title: "All-data pair-local network + coarse prior (v3, one gap/3 solve, R=0)"
repo_id: SQuADDS/2026PaperModels-all-data-v3s
kind: gnn
prior: v3
rounds: 0
status: available
fallback: all-data-base
checkpoints: [data_local/mega/alldata/v3s_r0_s0.pt, data_local/mega/alldata/v3s_r0_s1.pt, data_local/mega/alldata/v3s_r0_s2.pt]
summary: "Deployed coarse tier (default; fastest). Trained on every labelled design on the conductor-anchored v3 MoM prior (one gap/3 solve). Candidate V3S-PLall of deploy/MODEL_SELECTION.md."
validation: {median_pct: 1.75, mae_ff: 0.534, rmse_ff: 1.173, set: "212 deploy held-out designs (deploy/MODEL_SELECTION.md)"}
paper:
appears_in: [tab:candidates, tab:wonky-shape-main]
ood60: {median_pct: 1.06, mae_ff: 0.45, rmse_ff: 0.90, note: "OOD-60 not in training; never-seen families are"}
heldout:
ood60: {median_pct: 1.06, mae_ff: 0.449, rmse_ff: 0.898}
pocket3_random36: {median_pct: 1.90, mae_ff: 0.319, rmse_ff: 0.563}
qiskit_metal_new116: {median_pct: 1.84, mae_ff: 0.654, rmse_ff: 1.416}
pocket3_biased20: {median_pct: 3.10, mae_ff: 0.503, rmse_ff: 0.836}
cost_s_per_design: {median: 3.0, p90: 8.0, max: 12.8, note: "end to end from the GDS, 2 threads, HPC main-partition node"}
all-data-v3-rich:
title: "All-data pair-local network + fine prior (v3-rich, gap/3 + gap/6 Richardson, R=0)"
repo_id: SQuADDS/2026PaperModels-all-data-v3-rich
kind: gnn
prior: v3-rich
rounds: 0
status: available
fallback: all-data
checkpoints: [data_local/mega/alldata/v3_r0_s0.pt, data_local/mega/alldata/v3_r0_s1.pt, data_local/mega/alldata/v3_r0_s2.pt]
summary: "Fine tier until 2026-10-10 (now all-data-flagship: same weights, flagship prior). Trained on every labelled design on the Richardson-extrapolated v3 MoM prior (gap/3 + gap/6). Candidate V3R-PLall of deploy/MODEL_SELECTION.md."
validation: {median_pct: 1.31, mae_ff: 0.403, rmse_ff: 0.877, set: "212 deploy held-out designs (deploy/MODEL_SELECTION.md)"}
paper:
appears_in: [tab:candidates, tab:wonky-shape-main]
ood60: {median_pct: 0.66, mae_ff: 0.23, rmse_ff: 0.48, note: "OOD-60 not in training; never-seen families are"}
heldout:
ood60: {median_pct: 0.66, mae_ff: 0.230, rmse_ff: 0.481}
pocket3_random36: {median_pct: 1.75, mae_ff: 0.268, rmse_ff: 0.506}
qiskit_metal_new116: {median_pct: 1.40, mae_ff: 0.505, rmse_ff: 1.064}
pocket3_biased20: {median_pct: 3.15, mae_ff: 0.536, rmse_ff: 1.017}
cost_s_per_design: {median: 12.8, p90: 33.3, max: 50.4, note: "end to end from the GDS, 2 threads, HPC node; stress layouts up to 293 s (400-finger comb)"}
all-data-lite500:
title: "All-data pair-local network retrained on the flagship-fast lite500 prior (R=0)"
kind: gnn
prior: lite500
rounds: 0
status: candidate
checkpoints: [data_local/mega/alldata/lite500_r0_s0.pt, data_local/mega/alldata/lite500_r0_s1.pt, data_local/mega/alldata/lite500_r0_s2.pt]
summary: "flagship-fast candidate (PRIOR_SCHEME.md Part III): micro-MoM at <= 500 merged panels (one window up to 8 conductors; core-graded windows + screened-dipole far pairs beyond); network retrained on every labelled design on this prior (HPC 10-10). Not deployable until the user decides."
all-data-lite1500:
title: "flagship-fast: all-data pair-local network retrained on the lite-1500 micro-MoM prior (R=0)"
display_name: flagship-fast
repo_id: SQuADDS/2026PaperModels-all-data-lite1500
kind: gnn
prior: lite1500
rounds: 0
status: available
checkpoints: [data_local/mega/alldata/lite1500_r0_s0.pt, data_local/mega/alldata/lite1500_r0_s1.pt, data_local/mega/alldata/lite1500_r0_s2.pt]
summary: "Default released tier since 2026-10-10 (flagship-fast). Cheap physics prior: the flagship mesh family at <= 1,500 merged panels in a canonical frame (one window up to 8 conductors; beyond, core-graded windows of <= 8 conductors with closed-form / screened-dipole far pairs), network retrained on every labelled design on this prior (PRIOR_SCHEME.md Part III). Faster but less accurate than the coarse and fine tiers; the only tier that handles multi-qubit chips."
validation: {median_pct: 2.67, mae_ff: 0.768, rmse_ff: 1.874, set: "212 deploy held-out designs (deploy/MODEL_SELECTION.md)"}
heldout:
largen40: {median_pct: 4.96, mae_ff: 0.406, rmse_ff: 1.303}
pocket3_biased20: {median_pct: 5.75, mae_ff: 0.641, rmse_ff: 1.076}
cost_s_per_design: {median: 0.23, p90: 0.68, note: "end to end from the GDS, 2 threads, Mac M5 Pro, 212 held-out designs; 780-conductor grid ~15 s on 18 Mac cores"}
all-data-lite1500td16:
title: "flagship-fast candidate: all-data pair-local network on the lite-1500 prior with approximate pair tokens td16 (R=0)"
kind: gnn
prior: lite1500
tokens: td16
rounds: 0
status: candidate
checkpoints: [data_local/mega/alldata/lite1500td16_r0_s0.pt, data_local/mega/alldata/lite1500td16_r0_s1.pt, data_local/mega/alldata/lite1500td16_r0_s2.pt]
summary: "Token-variant candidate (PRIOR_SCHEME.md 14.3): far boundary samples block-decimated (td16); network retrained on these tokens. Not deployable until the user decides."
all-data-lite1500td64:
title: "flagship-fast candidate: all-data pair-local network on the lite-1500 prior with approximate pair tokens td64 (R=0)"
kind: gnn
prior: lite1500
tokens: td64
rounds: 0
status: candidate
checkpoints: [data_local/mega/alldata/lite1500td64_r0_s0.pt, data_local/mega/alldata/lite1500td64_r0_s1.pt, data_local/mega/alldata/lite1500td64_r0_s2.pt]
summary: "Token-variant candidate (PRIOR_SCHEME.md 14.3): far boundary samples block-decimated (td64); network retrained on these tokens. Not deployable until the user decides."
all-data-lite1500tc600:
title: "flagship-fast candidate: all-data pair-local network on the lite-1500 prior with approximate pair tokens tc600 (R=0)"
kind: gnn
prior: lite1500
tokens: tc600
rounds: 0
status: candidate
checkpoints: [data_local/mega/alldata/lite1500tc600_r0_s0.pt, data_local/mega/alldata/lite1500tc600_r0_s1.pt, data_local/mega/alldata/lite1500tc600_r0_s2.pt]
summary: "Token-variant candidate (PRIOR_SCHEME.md 14.3): far boundary samples block-decimated (tc600); network retrained on these tokens. Not deployable until the user decides."
all-data-formula:
title: "All-data pair-local network retrained on the flagship-fast formula prior (R=0)"
kind: gnn
prior: formula
rounds: 0
status: candidate
checkpoints: [data_local/mega/alldata/formula_r0_s0.pt, data_local/mega/alldata/formula_r0_s1.pt, data_local/mega/alldata/formula_r0_s2.pt]
summary: "flagship-fast candidate (PRIOR_SCHEME.md Part III): closed-form edge-coupling prior (no mesh); network retrained on every labelled design on this prior (HPC 10-10). Not deployable until the user decides."
# ---------------------------------------------------------------- flagship prior (PRIOR_SCHEME.md Part II, 2026-10-09): same
# network as all-data-v3-rich, fed with the strip-merged, resolution-controlled Richardson prior (H-matrix solver for large
# layouts). all-data-flagship = released fine tier since 2026-10-10 (user-approved; PRIOR_SCHEME.md section 8.1).
all-data-flagship:
title: "All-data pair-local network + fine prior (flagship: strip-merged Richardson gap/6 + gap/3, resolution-controlled, H-matrix solve; R=0)"
repo_id: SQuADDS/2026PaperModels-all-data-flagship
kind: gnn
prior: flagship
rounds: 0
status: available
checkpoints: [data_local/mega/alldata/v3_r0_s0.pt, data_local/mega/alldata/v3_r0_s1.pt, data_local/mega/alldata/v3_r0_s2.pt]
summary: "Deployed fine tier (most accurate) since 2026-10-10. The all-data-v3-rich network (unchanged weights, trained on every labelled design) with the flagship prior: 6 + 3 panels across the smallest gap on every layout, Richardson-extrapolated, dense Cholesky below 12k panels and an H-matrix / PCG solve above. Scale-invariant: the same resolution from one capacitor to a multi-qubit grid. Also the model that the coarse tier routes large layouts to."
validation: {median_pct: 1.13, mae_ff: 0.374, rmse_ff: 0.836, set: "212 deploy held-out designs (deploy/MODEL_SELECTION.md; PRIOR_SCHEME.md section 8.1)"}
paper:
appears_in: [tab:flagship, fig:flagship, tab:flagship-cost, sec:res-opt (released models)]
ood60: {median_pct: 0.60, mae_ff: 0.204, rmse_ff: 0.367, note: "OOD-60 not in training"}
heldout:
never_seen926_in_training: {median_pct: 0.50, mae_ff: 0.235, rmse_ff: 0.420, note: "the all-data network trained on these families: not zero-shot"}
ood60: {median_pct: 0.60, mae_ff: 0.204, rmse_ff: 0.367}
pocket3_random36: {median_pct: 1.29, mae_ff: 0.234, rmse_ff: 0.488}
qiskit_metal_new116: {median_pct: 1.28, mae_ff: 0.478, rmse_ff: 1.021}
pocket3_biased20: {median_pct: 2.51, mae_ff: 0.477, rmse_ff: 0.988}
cost_s_per_design: {median: 0.87, p90: 11.6, max: 14.0, note: "212 held-out designs, end to end from the GDS, 2 threads, Mac M5 Pro (PRIOR_SCHEME.md 8.1); 1,158 held-out designs incl. never-seen: median 1.51, p90 27.6, max 78.4 s; peak RSS <= 8.3 GB (grid_qubit_9, 220k panels)"}
all-data-flagship-coarse:
title: "All-data pair-local network + flagship-coarse prior (strip-merged Richardson gap/3 + 2 gap/3, resolution-controlled; R=0)"
repo_id: SQuADDS/2026PaperModels-all-data-v3-rich
kind: gnn
prior: flagship-coarse
rounds: 0
status: candidate
pending_reason: "not released: fails the PRIOR_SCHEME.md section 7 accuracy criterion on the in-family test split (section 8.1); evaluation candidate only"
checkpoints: [data_local/mega/alldata/v3_r0_s0.pt, data_local/mega/alldata/v3_r0_s1.pt, data_local/mega/alldata/v3_r0_s2.pt]
summary: "The fine-tier network (unchanged weights) with the cheaper flagship pair (3 + 1.5 panels across the smallest gap, Richardson-extrapolated)."
validation: {median_pct: 1.22, mae_ff: 0.418, rmse_ff: 0.908, set: "212 deploy held-out designs (PRIOR_SCHEME.md section 8.1)"}
# ---------------------------------------------------------------- this work: conductor-anchored v3 priors (paper_ml, 2026-10-08)
# coarse prior = v3s (one gap/3 solve; was "fast"), fine prior = v3-rich (gap/3 + gap/6, Richardson; was "accurate").
# Model ids and repo names keep the old words (*-fast, *-accurate): they are published links. Metrics: paper tab:candidates
# (median % / MAE fF / RMSE fF, all mutual entries); in_family_test = held-out test split of the 5 families.
pair-local-fast:
title: "Pair-local network + coarse prior (v3, R=0) - paper headline"
repo_id: SQuADDS/2026PaperModels-pair-local-fast
kind: gnn
prior: v3
rounds: 0
status: available
fallback: pair-local
checkpoints: [data_local/mega/v3s/ALL_mom_r0_s0.pt, data_local/mega/v3s/ALL_mom_r0_s1.pt, data_local/mega/v3s/ALL_mom_r0_s2.pt]
summary: "This work's headline: zero-shot on never-seen components 1.23 % median. Five training families, coarse conductor-anchored prior (one gap/3 solve)."
paper:
appears_in: [abstract, tab:candidates, tab:costacc, fig:costacc, tab:infamily-v3, tab:wonky-shape-main, fig:newtopo-all]
never_seen: {median_pct: 1.23, mae_ff: 0.57, rmse_ff: 1.02}
ood60: {median_pct: 1.33, mae_ff: 0.47, rmse_ff: 0.92}
in_family_test: {median_pct: 0.30, mae_ff: 0.047, rmse_ff: 0.144}
deploy_heldout: {median_pct: 2.63, mae_ff: 0.616, rmse_ff: 1.193, set: "212 deploy held-out designs (deploy/MODEL_SELECTION.md)"}
message-passing-fast:
title: "Message-passing network + coarse prior (v3, R=3)"
repo_id: SQuADDS/2026PaperModels-message-passing-fast
kind: gnn
prior: v3
rounds: 3
status: available
fallback: message-passing
checkpoints: [data_local/mega/v3s/ALL_mom_s0.pt, data_local/mega/v3s/ALL_mom_s1.pt, data_local/mega/v3s/ALL_mom_s2.pt]
summary: "Five training families, coarse prior, three message-passing rounds."
paper:
appears_in: [tab:candidates, tab:costacc, fig:costacc, tab:infamily-v3, fig:newtopo-all]
never_seen: {median_pct: 1.67, mae_ff: 0.72, rmse_ff: 1.32}
ood60: {median_pct: 1.33, mae_ff: 0.64, rmse_ff: 1.48}
in_family_test: {median_pct: 0.21, mae_ff: 0.032, rmse_ff: 0.078}
deploy_heldout: {median_pct: 2.70, mae_ff: 0.638, rmse_ff: 1.259, set: "212 deploy held-out designs (deploy/MODEL_SELECTION.md)"}
pair-local-accurate:
title: "Pair-local network + fine prior (v3-rich, R=0)"
repo_id: SQuADDS/2026PaperModels-pair-local-accurate
kind: gnn
prior: v3-rich
rounds: 0
status: available
fallback: pair-local-fine
checkpoints: [data_local/mega/v3/ALL_mom_r0_s0.pt, data_local/mega/v3/ALL_mom_r0_s1.pt, data_local/mega/v3/ALL_mom_r0_s2.pt]
summary: "Five training families, fine (Richardson-extrapolated, gap/3 + gap/6) prior."
paper:
appears_in: [tab:candidates, tab:costacc, fig:costacc, tab:infamily-v3, tab:wonky-shape-main, fig:newtopo-all]
never_seen: {median_pct: 1.57, mae_ff: 0.92, rmse_ff: 1.57}
ood60: {median_pct: 1.02, mae_ff: 0.35, rmse_ff: 0.66}
in_family_test: {median_pct: 0.23, mae_ff: 0.040, rmse_ff: 0.140}
deploy_heldout: {median_pct: 1.91, mae_ff: 0.484, rmse_ff: 1.022, set: "212 deploy held-out designs (deploy/MODEL_SELECTION.md)"}
message-passing-accurate:
title: "Message-passing network + fine prior (v3-rich, R=3)"
repo_id: SQuADDS/2026PaperModels-message-passing-accurate
kind: gnn
prior: v3-rich
rounds: 3
status: available
fallback: message-passing-fine
checkpoints: [data_local/mega/v3/ALL_mom_s0.pt, data_local/mega/v3/ALL_mom_s1.pt, data_local/mega/v3/ALL_mom_s2.pt]
summary: "Five training families, fine prior, three message-passing rounds. Best in family (0.18 % on the held-out test split)."
paper:
appears_in: [tab:candidates, tab:costacc, fig:costacc, tab:infamily-v3, fig:newtopo-all]
never_seen: {median_pct: 1.34, mae_ff: 0.60, rmse_ff: 1.07}
ood60: {median_pct: 1.22, mae_ff: 0.43, rmse_ff: 0.82}
in_family_test: {median_pct: 0.18, mae_ff: 0.031, rmse_ff: 0.100}
deploy_heldout: {median_pct: 2.11, mae_ff: 0.695, rmse_ff: 2.074, set: "212 deploy held-out designs (deploy/MODEL_SELECTION.md)"}
charge-corrected-prior-fast:
title: "Charge-corrected coarse prior (v3c cell GBM)"
repo_id: SQuADDS/2026PaperModels-charge-corrected-prior-fast
kind: charge_prior
prior: v3
status: available
checkpoints: [deploy/artifacts/charge_gbm_v3c.joblib]
summary: "Coarse (gap/3 v3) prior with its panel charges corrected by a gradient-boosted model of local geometry fit on the training families' surface charge; no capacitance labels."
paper:
appears_in: [tab:candidates]
never_seen: {median_pct: 2.67, mae_ff: 0.96, rmse_ff: 1.50}
ood60: {median_pct: 1.22, mae_ff: 0.43, rmse_ff: 0.75}
in_family_test: {median_pct: 0.87, mae_ff: 0.24, rmse_ff: 0.76}
blend-pair-local-fast:
title: "Blended model: pair-local coarse-prior network x charge-corrected coarse prior"
repo_id: SQuADDS/2026PaperModels-blend-pair-local-fast
kind: blend
components: [pair-local-fast, charge-corrected-prior-fast]
status: available
summary: "Log-space blend. Paper numbers use the published file-order weights (0.6 / 0.7 / 0.8); the package uses the order-invariant weights (0.6 / 0.9), see deploy_heldout."
paper:
appears_in: [tab:candidates]
never_seen: {median_pct: 1.28, mae_ff: 0.52, rmse_ff: 0.90, note: published file-order weights}
ood60: {median_pct: 0.98, mae_ff: 0.36, rmse_ff: 0.74, note: published file-order weights}
in_family_test: {median_pct: 0.35, mae_ff: 0.082, rmse_ff: 0.264}
deploy_heldout: {median_pct: 2.33, mae_ff: 0.547, rmse_ff: 1.097, ood60: "1.09 / 0.405 / 0.799", set: "212 deploy held-out designs, order-invariant weights"}
blend-pair-local-accurate:
title: "Blended model: pair-local fine-prior network x charge-corrected coarse prior"
repo_id: SQuADDS/2026PaperModels-blend-pair-local-accurate
kind: blend
components: [pair-local-accurate, charge-corrected-prior-fast]
status: available
summary: "Log-space blend. Paper numbers use the published file-order weights (0.6 / 0.7 / 0.8); the package uses the order-invariant weights (0.6 / 0.9), see deploy_heldout."
paper:
appears_in: [tab:candidates]
never_seen: {median_pct: 1.75, mae_ff: 0.87, rmse_ff: 1.48, note: published file-order weights}
ood60: {median_pct: 0.96, mae_ff: 0.31, rmse_ff: 0.54, note: published file-order weights}
in_family_test: {median_pct: 0.32, mae_ff: 0.081, rmse_ff: 0.266}
deploy_heldout: {median_pct: 2.02, mae_ff: 0.485, rmse_ff: 1.048, ood60: "1.11 / 0.329 / 0.594", set: "212 deploy held-out designs, order-invariant weights"}
fast-prior:
title: "Coarse prior alone (v3, one gap/3 solve, training-free)"
repo_id: null
kind: prior
prior: v3
status: available
summary: "Untrained thin-film MoM on the conductor-anchored v3 mesh, one gap/3 solve (default 8,000-panel cap)."
paper:
appears_in: [tab:candidates, tab:prior]
never_seen: {median_pct: 5.68, mae_ff: 2.67, rmse_ff: 3.91}
ood60: {median_pct: 6.35, mae_ff: 1.66, rmse_ff: 2.56}
in_family_test: {median_pct: 4.38, mae_ff: 0.70, rmse_ff: 1.73}
accurate-prior:
title: "Fine prior alone (v3-rich, gap/3 + gap/6 Richardson, training-free)"
repo_id: null
kind: prior
prior: v3-rich
status: available
summary: "Untrained thin-film MoM, gap/3 and gap/6 solves on the v3 mesh with Richardson extrapolation."
paper:
appears_in: [tab:candidates, tab:prior]
never_seen: {median_pct: 2.71, mae_ff: 1.42, rmse_ff: 2.34}
ood60: {median_pct: 5.05, mae_ff: 1.30, rmse_ff: 2.08}
in_family_test: {median_pct: 2.71, mae_ff: 0.53, rmse_ff: 1.54}
# ---------------------------------------------------------------- baselines (paper comparison; not deployable on a GDS)
parameter-mlp:
title: "Parameter MLP (published surrogate, MLQubitDesign)"
repo_id: null
kind: baseline
status: not-released
deployable: false
summary: "Per-family MLP on Qiskit Metal design parameters; input undefined on a new layout."
paper:
appears_in: [tab:candidates, tab:sota, fig:infamily, tab:fewshot]
in_family_full: {median_pct: "1.78-2.62", mae_ff: "0.041-1.32", rmse_ff: "0.055-1.87", note: C_NS}
parameter-gbt:
title: "Parameter GBT (gradient-boosted trees on design parameters)"
repo_id: null
kind: baseline
status: not-released
deployable: false
summary: "Per-family GBT on design parameters; input undefined on a new layout."
paper:
appears_in: [tab:candidates, fig:infamily]
in_family_full: {median_pct: "0.35-3.36", mae_ff: "0.010-1.93", rmse_ff: "0.015-2.92"}
neural-operator:
title: "Fourier neural operator on conductor masks (no prior)"
repo_id: null
kind: baseline
status: not-released
deployable: false
summary: "Image model; most accurate in distribution, fails on held-out families."
paper:
appears_in: [tab:candidates]
in_family_id: {median_pct: 0.24, mae_ff: 0.082, rmse_ff: 0.40}
edge-nn-mom:
title: "Edge nearest-neighbour + MoM (training-free correction)"
repo_id: null
kind: baseline
status: not-released
deployable: false
summary: "Nearest-neighbour correction of the MoM prior in embedding space; needs the training index."
paper:
appears_in: [tab:candidates, tab:lofo]
never_seen: {median_pct: 4.02, mae_ff: 1.5, rmse_ff: 2.4}
ood60: {median_pct: "3.1-6.0", mae_ff: "1.2-1.6", rmse_ff: "1.5-2.2", note: range of per-class medians}
# ---------------------------------------------------------------------------------------------------------------------------
# paper_v3 final recipes (tlsq_infer.final; parity with scripts/tm_final.py: deploy/tests/test_final_parity.py) and the new
# Space (PLAN_HF_SPACE_V2.md; user decisions 2026-10-14: one PRIVATE Space with a model toggle, built as a new copy, the
# 2026-10-10 Spaces above never overwritten; geometry = training geometry, no ground crop by default). Separate from
# `models` (the 2026-10-10 tiers) so the existing apps and tests are unchanged. Accuracy = the paper's test-once numbers of
# the REPORTED model (data_local/train_matrix/eval/final/summary.parquet; median % / MAE fF / RMSE fF over mutual entries).
v2_hub:
space: shanto268/PlanarCapModel # NEW private repo (user, 2026-10-14); never the 2026-10-10 Spaces above
title: Planar Capacitance (paper models)
private: true # stays private until the paper is out (user, 2026-10-14)
final_defaults:
model: E3
trust_rule: true # APPLIED by default to E3 and E3-D (user decision 2026-10-14); trust_rule=False switches it off
trust_rule_note: "Trust rule (applied by default to E3 and E3-D). The E3 model is three independently trained networks. When they agree, their answer is used for pairs of conductors closer than 100 um, and the physics-only C0 value for all other pairs. When they disagree (median spread of log C over the close pairs above tau = 1.254e-3, a threshold fixed in the paper before its blind test), the whole matrix falls back to the physics-only C0 recipe. Every E3 result shows the spread, the threshold and whether the fallback happened. On the paper's blind 9-qubit grid the networks disagreed (spread 2.28e-3) and the rule chose C0: 0.453 % (MAE 0.247 / RMSE 0.473 fF) on the 135 close pairs, against 1.153 % (0.235 / 0.319 fF) for E3 alone. The test results listed for E3 are those of E3 without the rule. Switch the rule off with the 'apply the trust rule' box (UI), trust_rule=False (/predict_gds) or predict_final(..., trust_rule=False) in the package."
final_models:
E3:
title: "Model (E3): extrapolated physics + per-kind offsets + bounded learned correction"
recipe: E3
pair: [1.5, 3.0]
role: reported model (paper_v3)
ledger_id: em_1_p1.5
source: data_local/e13/em/runs/1 # info_p1.5.json + s{0,1,2}_p1.5.pt (training checkpoints); bundles carry an export
summary: "log-Richardson R(1.5, 3) of the slab lite micro-MoM, per-entry-kind offsets and a bounded (+-3 %) SceneGNN correction; mean log C of 3 seeds. Trained on exp6-12 train + exp13-15 + random 200."
accuracy:
in_family_test_exp6_11: {set: "in-family test exp6-11 (3,819 designs; test-once)", median_pct: 0.110, mae_ff: 0.017, rmse_ff: 0.093}
in_family_test_exp12: {set: "in-family test exp12 (190)", median_pct: 0.089, mae_ff: 0.009, rmse_ff: 0.016}
ood60: {set: "OOD-60 freeform (60)", median_pct: 0.393, mae_ff: 0.195, rmse_ff: 0.437}
qm116: {set: "qm116 unseen qubits / couplers (116)", median_pct: 0.761, mae_ff: 0.227, rmse_ff: 0.591}
p3rand36: {set: "p3rand36 (36)", median_pct: 0.690, mae_ff: 0.099, rmse_ff: 0.180}
largen: {set: "large-N 8-12 conductors (40)", median_pct: 0.972, mae_ff: 0.058, rmse_ff: 0.158}
atlas: {set: "quantum-metal atlas (348)", median_pct: 1.240, mae_ff: 7.102, rmse_ff: 11.734}
grid_qubit_5_near: {set: "grid_qubit_5, pairs within 100 um (67 entries)", median_pct: 1.135, mae_ff: 0.247, rmse_ff: 0.319}
E1:
title: "Fast (E1): extrapolated physics + per-kind offsets (no network)"
recipe: E1
pair: [0.75, 1.5]
role: reported model (paper_v3), fast mode
ledger_id: em_1_E1_p0.75
source: data_local/e13/em/1/info_E1_p0.75.json
summary: "log-Richardson R(0.75, 1.5) of the slab lite micro-MoM plus the model's per-entry-kind offsets c_kind; no network, about 3x cheaper solves."
accuracy:
in_family_test_exp6_11: {set: "in-family test exp6-11 (3,819 designs; test-once)", median_pct: 0.656, mae_ff: 0.105, rmse_ff: 0.366}
in_family_test_exp12: {set: "in-family test exp12 (190)", median_pct: 0.964, mae_ff: 0.058, rmse_ff: 0.110}
ood60: {set: "OOD-60 freeform (60)", median_pct: 1.175, mae_ff: 0.503, rmse_ff: 1.159}
qm116: {set: "qm116 unseen qubits / couplers (116)", median_pct: 1.788, mae_ff: 0.568, rmse_ff: 1.135}
p3rand36: {set: "p3rand36 (36)", median_pct: 1.681, mae_ff: 0.380, rmse_ff: 0.652}
largen: {set: "large-N 8-12 conductors (40)", median_pct: 1.894, mae_ff: 0.154, rmse_ff: 0.326}
atlas: {set: "quantum-metal atlas (348)", median_pct: 1.386, mae_ff: 9.423, rmse_ff: 16.172}
grid_qubit_5_near: {set: "grid_qubit_5, pairs within 100 um (67 entries)", median_pct: 0.544, mae_ff: 0.360, rmse_ff: 0.822}
C0:
title: "Physics only (C0): extrapolated micro-MoM, no fitted parameter"
recipe: C0
pair: [1.5, 3.0]
role: reported model (paper_v3), physics only
ledger_id: C0_p1.5
summary: "log-Richardson R(1.5, 3) = 2 log MoM(3) - log MoM(1.5) of the slab lite micro-MoM; zero fitted parameters."
accuracy:
in_family_test_exp6_11: {set: "in-family test exp6-11 (3,819 designs; test-once)", median_pct: 1.259, mae_ff: 0.172, rmse_ff: 0.312}
in_family_test_exp12: {set: "in-family test exp12 (190)", median_pct: 1.374, mae_ff: 0.114, rmse_ff: 0.226}
ood60: {set: "OOD-60 freeform (60)", median_pct: 1.065, mae_ff: 0.303, rmse_ff: 0.572}
qm116: {set: "qm116 unseen qubits / couplers (116)", median_pct: 1.260, mae_ff: 0.255, rmse_ff: 0.591}
p3rand36: {set: "p3rand36 (36)", median_pct: 0.836, mae_ff: 0.111, rmse_ff: 0.181}
largen: {set: "large-N 8-12 conductors (40)", median_pct: 0.955, mae_ff: 0.081, rmse_ff: 0.215}
atlas: {set: "quantum-metal atlas (348)", median_pct: 0.609, mae_ff: 3.262, rmse_ff: 5.882}
newfam926: {set: "never-seen 926", median_pct: 0.597, mae_ff: 0.231, rmse_ff: 0.390}
grid_qubit_5_near: {set: "grid_qubit_5, pairs within 100 um (67 entries)", median_pct: 0.735, mae_ff: 0.238, rmse_ff: 0.460}
E3-D:
title: "Deployment model (E3 recipe, D weights): trained on every labelled design"
recipe: E3
pair: [1.5, 3.0]
role: deployment model (never a paper claim, never test-scored)
ledger_id: null
source: data_local/e13/em/runs/D
summary: "E3 recipe retrained on 42,967 designs (exp6-12 train + val + test, exp13-15 (all 1,754), random 1,000, OOD-60). Its scores on those sets would be in-sample; held-out scoring (atlas, qm116, p3rand36, large-N, grids) not yet run."
warning: "Deployment model D: trained on the in-family test split, exp13-15 and OOD-60, so no test number of the paper applies to it; its held-out scoring is pending. The paper's independently tested model is 'E3'."
scoring_note: &d_scoring "Scoring note for D (2026-10-14, documentation only, no set scored): IN-SAMPLE for D (trained on them, any score would not be a test): in-family test exp6-11 and exp12 (D trained on train + val + test), never-seen 926 (exp13-15, all 1,754 trained), OOD-60 (unusual), random 1,000. HELD OUT from D (no rows of these in its training data): quantum-metal atlas (348), qm116 (116), p3rand36 (36), large-N (40), grid_qubit_1 / 5 / 9, multi-stack, margin. D has not been scored on any of them; numbers shown are the paper's reported model, never D's."
accuracy: {}
E1-D:
title: "Deployment fast (E1 recipe, D constants): per-kind offsets fitted on every labelled design"
recipe: E1
pair: [0.75, 1.5]
role: deployment model, fast mode (never a paper claim, never test-scored)
ledger_id: null
source: data_local/e13/em/D/info_E1_p0.75.json # scripts/em_matrix.py e1 --models D --pair 0.75 (2026-10-14; train_keys_md5 = D's E3 rows)
summary: "log-Richardson R(0.75, 1.5) of the slab lite micro-MoM plus per-entry-kind offsets c_kind fitted (median log residual, the E1 rule of scripts/em_matrix.py) on D's 42,967 training designs; no network."
warning: "Deployment fast model (D constants): fitted on the in-family test split, exp13-15 and OOD-60, so no test number of the paper applies to it; its held-out scoring is pending. The paper's independently tested fast model is 'E1'."
scoring_note: *d_scoring
accuracy: {}
# ---------------------------------------------------------------- GNN Space (deploy/GNN_SPACE.md; user 2026-10-16)
# Built by `python deploy/sync.py --gnn` (dry run) / `--gnn --push`; gated by deploy/tests/test_gnn_space.py. New models: add an
# entry below, keep the old one with status: retired (never delete), point gnn_hub.default_model at the new id.
gnn_hub:
space: shanto268/PlanarCapGNN # NEW private repo (user, 2026-10-16); never the Spaces above
status: retired # paused 2026-10-20 (HfApi.pause_space; private, files kept, Space sha 4193e5a7), superseded by the flagship + paper Spaces; restore = HfApi().restart_space
title: Planar Capacitance (GNN preview)
private: true # Spaces stay private (CLAUDE.md rule 11)
space_hardware: cpu-basic # free tier: 2 vCPU, 16 GB RAM
default_model: GNN
compare_models: [E3, C0] # final_models entries served next to the GNN (tlsq_infer.final, as in the v2 Space)
finalization: agents/geomfast/FINALIZATION_V1.json # frozen checkpoint SHA256s; sync.py refuses checkpoints not listed there
parity_reference: data_local/geomfast/raw_parity.json # the 14 TRAIN layouts of the raw-GDS parity gate
frozen_stage: /Users/shanto/Documents/ChatGPT/transfer_learning/geomfast_stage # frozen inference code (parity reference; local)
fastprior_status: agents/geomfast/FASTPRIOR_STATUS.md # cpu_fast allowed only with a "PASSED for ... cpu_fast" line there (2026-10-16: cpu_fast FAILED, 9.6e-6)
comparison_note: "Paired against E3 on twelve held-out cohorts (paper_v4; equivalence margin 0.25 percentage points of layout-mean median error or 0.05 fF of MAE, set after the results were seen): equivalent on eight, materially better on four (two-net test, component atlas, the three added families against the five-family E3 that never saw them while the GNN trained on 828 other designs of them, sealed synthetic layouts on the median only), materially worse on none; equal end-to-end latency (the two prior solves dominate both)."
gnn_models:
GNN:
title: "Accurate (physics + network, ~1 s): corrected prior v2m + 3-seed GNN ensemble"
status: available
tier: accurate
arm: full_gnn
role: "accurate tier, default (since 2026-10-17; arm Bm of agents/priorv2): the frozen P2d full_gnn weights (FINALIZATION_V1, unchanged) + prior v2m solved by the compiled exact solver"
ensemble: arithmetic mean of the three full Maxwell matrices
checkpoints:
- {path: data_local/geomfast/runs/p2d_full_gnn_s0/model.pt, sha256: fcb5db5d8237436d546aca0e7463a2f1d4f2e1b18757b2a35eff646c03f3bc5c}
- {path: data_local/geomfast/runs/p2d_full_gnn_s1/model.pt, sha256: 9aef64adfaf99cad35da891a2329d43eac46b2768cd133b5313085891a4fed92}
- {path: data_local/geomfast/runs/p2d_full_gnn_s2/model.pt, sha256: 357c16539967404b0a7eea514ef3ff3dc763c764da192b219d919cb01f697669}
prior:
method: local_gap # micro_mom_loc (slab, local-gap mesh); base = 2 log C(3) - log C(1.5)
ppg: [1.5, 3.0]
alpha: 0.49685219816506904 # unchanged (TRAIN-fitted)
prior_impl: frozen
version: v2m_fc_dipfill_margin # agents/priorv2/FINALIZATION_V2.json "prior"
patches: [priorv2m] # VALUE-CHANGING: geomgnn/priorv2.py = scripts/priorv2/priorv2.py, installed(policy='fc', dip=True, margin=True)
module_sha256: 1a0bdcb71500c7f11e63238afe531c18caabd5c9f6f7aaa73d9b7dcd0ad61f8c
mode: exact_compiled # geomgnn.fastprior_pkg solver mode (fv + hmat_vec + mixed + numba compiled H-matrix); TLSQ_GNN_PRIOR_MODE overrides
speed_package_sha256: 4fff680f35cb8b098de6c9b1b6fc55780064362ffc829bb47bf69a77454ce38a # scripts/fastprior_pkg/fastprior_pkg.py, vendored verbatim
speed_approved: "user 2026-10-17 ('faster at equal accuracy is a win'): fastprior_pkg 'Ready to deploy' (agents/fastprior_pkg/RESULTS.md); exact_compiled VAL tie on all 13 families; deploy-path parity gate in deploy/tests/test_gnn_space.py (<= 1e-6 vs the v2m exact path)"
certified_opt_in: "mode certified (+ Rayleigh-Ritz windows with an error certificate) is gated on VAL (tie) but not enabled here; set TLSQ_GNN_PRIOR_MODE=certified (Space variable) to opt in"
value_changing_approved: "user 2026-10-17 ('do whatever to make the models faster and better'); prior v2 PASSED its pre-declared protocol, arm Bm (frozen nets + v2m) recommended for deployment (agents/priorv2/RESULTS.md verdict 4, FINALIZATION_V2.json)"
summary: "The paper_v4 GNN (same frozen weights) on prior v2m: the frozen prior with three fixes of its windowed path (layouts with more than 8 conductors): rescue panels for metals the window mesh leaves without a panel (fc; also removes the rejection tail), the dipole fill for same-pocket far pairs (a physics bug of the frozen fill) and a wider window ground margin. Predictions for N <= 9 are bit-identical to the v1-prior model."
accuracy_note: "Validation split (VAL, development; NOT a test set), arm Bm of agents/priorv2 (data_local/priorv2/val_summary.csv): layout-mean median relative error, MAE and RMSE over mutual entries, 0.1 fF floor; priors computed on the GH200 CPU. v1 prior on the same VAL: all 0.337 % (0.0750 / 0.0997 fF), N > 9 3.468 % (0.2360 / 0.7004 fF). Test numbers (paper_v4, v2m): accuracy_test; the prior-v1 test scoring is kept as accuracy_test_prior_v1."
accuracy:
all_val: {set: "VAL, all layouts", layouts: 4608, median_pct: 0.2287, mae_ff: 0.06737, rmse_ff: 0.0792}
val_n_le_9: {set: "VAL, N <= 9 (bit-identical to the v1-prior model)", layouts: 4428, median_pct: 0.2096, mae_ff: 0.06846, rmse_ff: 0.07529}
val_n_gt_9: {set: "VAL, N > 9", layouts: 180, median_pct: 0.6979, mae_ff: 0.04063, rmse_ff: 0.1752}
val_b3_n_gt_9: {set: "VAL, campaign B3, N > 9", layouts: 115, median_pct: 0.8698, mae_ff: 0.04342, rmse_ff: 0.1724}
val_synth_n_gt_9: {set: "VAL, synthetic, N > 9", layouts: 65, median_pct: 0.3938, mae_ff: 0.0357, rmse_ff: 0.1801}
val_fc_tail: {set: "VAL, rejection-tail layouts (fc fires)", layouts: 16, median_pct: 0.6802, mae_ff: 0.03806, rmse_ff: 0.1842}
val_fill: {set: "VAL, same-pocket dipole fill fires", layouts: 54, median_pct: 1.092, mae_ff: 0.06035, rmse_ff: 0.2571}
val_margin: {set: "VAL, window margin widening fires", layouts: 40, median_pct: 1.207, mae_ff: 0.07865, rmse_ff: 0.3569}
val_twonet: {set: "VAL, two-net", layouts: 98, median_pct: 0.9075, mae_ff: 2.105, rmse_ff: 2.105}
accuracy_signal_self:
all_val: {set: "VAL, all layouts, signal self terms", layouts: 4608, median_pct: 0.2873, mae_ff: 0.1205, rmse_ff: 0.141}
val_n_gt_9: {set: "VAL, N > 9, signal self terms", layouts: 180, median_pct: 1.108, mae_ff: 0.2903, rmse_ff: 0.5737}
accuracy_test_note: "TEST (paper_v4 Table tab:ladder-test and SM Table tab:ladder-cohort, column 'model' = corrected prior v2m, compiled solver; read once): layout-mean median relative error %, MAE and RMSE fF over mutual entries, 0.1 fF floor. Zero-shot macro = mean of the nine component cohorts' layout means (1,878 layouts)."
accuracy_test:
infamily_test: {set: "in-family test (exp6-12 test split)", layouts: 4009, median_pct: 0.177, mae_ff: 0.0203, rmse_ff: 0.0250}
zeroshot_macro: {set: "zero-shot macro (9 component cohorts below)", layouts: 1878, median_pct: 0.587, mae_ff: 1.092, rmse_ff: 1.152}
newfam926: {set: "added families: Xmon, TransmonPocket, shielded concentric (926)", layouts: 926, median_pct: 0.264, mae_ff: 0.1164, rmse_ff: 0.1657}
twonet_test: {set: "two-net test", layouts: 147, median_pct: 0.594, mae_ff: 2.387, rmse_ff: 2.387}
random_sealed: {set: "random layouts, sealed", layouts: 98, median_pct: 0.559, mae_ff: 0.0903, rmse_ff: 0.1216}
synth_sealed: {set: "procedurally generated layouts, sealed", layouts: 107, median_pct: 0.651, mae_ff: 0.0462, rmse_ff: 0.1583}
p3rand36: {set: "held-out TransmonPocket + 3 couplers", layouts: 36, median_pct: 0.326, mae_ff: 0.0686, rmse_ff: 0.1152}
unusual: {set: "freeform geometries", layouts: 60, median_pct: 0.570, mae_ff: 0.2197, rmse_ff: 0.2873}
qm116: {set: "unseen qubits and couplers", layouts: 116, median_pct: 0.672, mae_ff: 0.2103, rmse_ff: 0.3522}
atlas: {set: "quantum-metal component atlas", layouts: 348, median_pct: 0.993, mae_ff: 6.621, rmse_ff: 6.631}
largen: {set: "large N, 8-12 conductors", layouts: 40, median_pct: 0.657, mae_ff: 0.0662, rmse_ff: 0.1522}
grid_qubit_1: {set: "one-qubit lattice, N = 15", layouts: 1, median_pct: 0.370, mae_ff: 0.0538, rmse_ff: 0.1769}
grid_qubit_5: {set: "five-qubit lattice, N = 35", layouts: 1, median_pct: 1.009, mae_ff: 0.0356, rmse_ff: 0.1449}
grid_qubit_9: {set: "nine-qubit lattice, N = 67", layouts: 1, median_pct: 0.960, mae_ff: 0.0221, rmse_ff: 0.1042}
accuracy_lofo_note: "Held-out component family (leave-one-family-out over the five original families; VAL only, no test cohort read): the shipped recipe (P2d full_gnn on prior v2m, 3-seed mean) retrained without the family, scored on that family's held-out VAL layouts (agents/shipped_lofo/RESULTS_P1.md, row d, protocol PROTOCOL_P1.md frozen 2026-10-18). Layout-mean median relative error %, MAE and RMSE fF over mutual entries, 0.1 fF floor. Better than the physics prior alone C^R = R(1.5, 3) on every family, median, MAE and RMSE (pooled 1.554 %, 0.166 / 0.205 fF). The served weights saw these families; this measures a new component type."
accuracy_lofo:
lofo_pooled: {set: "held-out family, pooled VAL", layouts: 4071, median_pct: 0.408, mae_ff: 0.0494, rmse_ff: 0.0591}
lofo_macro: {set: "held-out family, macro of 5", layouts: 4071, median_pct: 0.390, mae_ff: 0.1065, rmse_ff: 0.1383}
lofo_interdigital: {set: "interdigital held out", layouts: 1931, median_pct: 0.413, mae_ff: 0.0397, rmse_ff: 0.0435}
lofo_paddle: {set: "paddle held out", layouts: 1592, median_pct: 0.417, mae_ff: 0.0274, rmse_ff: 0.0312}
lofo_concentric: {set: "concentric held out", layouts: 170, median_pct: 0.544, mae_ff: 0.3824, rmse_ff: 0.4906}
lofo_star: {set: "star held out", layouts: 202, median_pct: 0.200, mae_ff: 0.0395, rmse_ff: 0.0492}
lofo_multipad: {set: "multipad star held out", layouts: 176, median_pct: 0.376, mae_ff: 0.0435, rmse_ff: 0.0771}
latency_test_note: "paper_v4 Table tab:ladder-test: end to end per layout from the layout file, median / 90th percentile, every arm in the same pass on two CPUs of an x86 cluster node (AMD EPYC 7513 / 7542, Xeon Silver 4116)."
latency_test:
infamily_test: {set: "in-family test", layouts: 4009, median_s: 0.845, p90_s: 5.96}
zeroshot: {set: "zero-shot cohorts, pooled", layouts: 1878, median_s: 1.76, p90_s: 18.1}
all_test: {set: "sum over all test layouts", layouts: 5902, sum_s: 20385}
accuracy_test_prior_v1_note: "paper_v4 Table 2 test cohorts, scored once with the SAME weights on prior v1 (the retired default; kept for the record). Layouts with N <= 9 give bit-identical predictions under v2m; the v2m test numbers are accuracy_test."
accuracy_test_prior_v1:
infamily_test: {set: "in-family test (exp6-12 test split)", layouts: 4009, median_pct: 0.1775, mae_ff: 0.02032, rmse_ff: 0.02502}
unusual: {set: "freeform shapes (OOD-60)", layouts: 60, median_pct: 0.5702, mae_ff: 0.2197, rmse_ff: 0.2873}
newfam926: {set: "added families: Xmon, TransmonPocket, shielded concentric (926 test designs; the GNN trained on 828 others of these families)", layouts: 926, median_pct: 0.2642, mae_ff: 0.1164, rmse_ff: 0.1657}
qm116: {set: "unseen qubits and couplers (qm116)", layouts: 116, median_pct: 0.6722, mae_ff: 0.2103, rmse_ff: 0.3522}
p3rand36: {set: "held-out TransmonPockets (p3rand36)", layouts: 36, median_pct: 0.3262, mae_ff: 0.06864, rmse_ff: 0.1152}
largen: {set: "large N, 8-12 conductors", layouts: 40, median_pct: 0.6575, mae_ff: 0.06625, rmse_ff: 0.1522}
atlas: {set: "quantum-metal component atlas", layouts: 348, median_pct: 0.9929, mae_ff: 6.621, rmse_ff: 6.631}
twonet_test: {set: "two-net test", layouts: 147, median_pct: 0.5945, mae_ff: 2.391, rmse_ff: 2.391}
random_sealed: {set: "random layouts, sealed", layouts: 110, median_pct: 0.5672, mae_ff: 0.09022, rmse_ff: 0.1225}
synth_sealed: {set: "synthetic layouts, sealed (2-68 conductors)", layouts: 107, median_pct: 0.6995, mae_ff: 0.1108, rmse_ff: 0.4076}
grid_qubit_1: {set: "grid unit, N = 15 (one layout)", layouts: 1, median_pct: 0.3703, mae_ff: 0.05383, rmse_ff: 0.1769}
grid_qubit_5: {set: "grid cluster, N = 35 (one layout)", layouts: 1, median_pct: 1.009, mae_ff: 0.0356, rmse_ff: 0.1449}
latency_note: "prior v1 numbers (paper_v4 Table 6). prior v2m: same cost on N <= 9; on the rejection tail the prior is 18-68x faster (agents/priorv2/RESULTS.md: VAL tail 3,609 -> 52.6 s on the GH200 CPU), the margin widening adds ~2 %."
latency:
development: {set: "development: 14 TRAIN layouts, one per source, N <= 18 (job 6492904)", layouts: 14, total_median_s: 1.258, total_p95_s: 53.25, geometry_median_s: 0.1202, prior_median_s: 0.95, network_median_s: 0.0116, paired_speedup_vs_E3: 1.14, speedup_ci95: [1.04, 1.26], machine: "GH200 node c02-13 (aarch64), 2 threads, 3 interleaved invocations per layout, median over layouts"}
external: {set: "external: 9 label-free layouts, N = 2-68 (job 6493337)", layouts: 9, total_median_s: 13.08, total_p95_s: 67.88, geometry_median_s: 1.796, prior_median_s: 10.6, network_median_s: 0.0428, paired_speedup_vs_E3: 0.955, speedup_ci95: [0.924, 0.981], machine: "GH200 node c02-13 (aarch64), 2 threads, 3 interleaved invocations per layout, median over layouts"}
GNN-prior-v1:
title: "GNN, prior v1 (default until 2026-10-17): full geometric embedding + two-resolution MoM prior, 3-seed ensemble"
status: retired # 2026-10-17: replaced as default by GNN (same weights, prior v2m); kept, never deleted
arm: full_gnn
role: retired default (paper_v4 primary; frozen 2026-10-16, FINALIZATION_V1); prior v1 + rejtail_fv
ensemble: arithmetic mean of the three full Maxwell matrices
checkpoints:
- {path: data_local/geomfast/runs/p2d_full_gnn_s0/model.pt, sha256: fcb5db5d8237436d546aca0e7463a2f1d4f2e1b18757b2a35eff646c03f3bc5c}
- {path: data_local/geomfast/runs/p2d_full_gnn_s1/model.pt, sha256: 9aef64adfaf99cad35da891a2329d43eac46b2768cd133b5313085891a4fed92}
- {path: data_local/geomfast/runs/p2d_full_gnn_s2/model.pt, sha256: 357c16539967404b0a7eea514ef3ff3dc763c764da192b219d919cb01f697669}
prior:
method: local_gap # micro_mom_loc (slab, local-gap mesh); base = 2 log C(3) - log C(1.5)
ppg: [1.5, 3.0]
alpha: 0.49685219816506904 # gap-share alpha fitted on TRAIN only
prior_impl: frozen # frozen | cpu_fast (geomgnn.fastmom; cpu_fast FAILED parity 2026-10-16: keep frozen)
patches: [rejtail_fv] # value-preserving rejection-tail fix (geomgnn/rejtail_fv.py; agents/rejtail/RESULTS.md; user 2026-10-17); bit-identical matrices
summary: "Message-passing network (2 rounds, 89,858 parameters per seed) over the full conductor vectors (shape scalars, Fourier contour, width spectrum), pair vectors with local context tokens and the global vector; corrects the log-Richardson extrapolation of the slab micro-MoM at 1.5 and 3 panels per gap by at most +-0.2 in log and predicts every self term through a learned positive row-sum closure, so the output is the complete Maxwell matrix (symmetric, Maxwell-signed, positive-definite signal block by construction). Trained on 35,354 Q3D layouts (all training sources, 2-68 conductors), Si 11.45 / 350 um."
accuracy: # paper_v4 Table 2 (data_local/geomfast/consolidated/layout_macro_summary.csv): layout-mean, 0.1 fF floor
infamily_test: {set: "in-family test (exp6-12 test split)", layouts: 4009, median_pct: 0.1775, mae_ff: 0.02032, rmse_ff: 0.02502}
unusual: {set: "freeform shapes (OOD-60)", layouts: 60, median_pct: 0.5702, mae_ff: 0.2197, rmse_ff: 0.2873}
newfam926: {set: "added families: Xmon, TransmonPocket, shielded concentric (926 test designs; the GNN trained on 828 others of these families)", layouts: 926, median_pct: 0.2642, mae_ff: 0.1164, rmse_ff: 0.1657}
qm116: {set: "unseen qubits and couplers (qm116)", layouts: 116, median_pct: 0.6722, mae_ff: 0.2103, rmse_ff: 0.3522}
p3rand36: {set: "held-out TransmonPockets (p3rand36)", layouts: 36, median_pct: 0.3262, mae_ff: 0.06864, rmse_ff: 0.1152}
largen: {set: "large N, 8-12 conductors", layouts: 40, median_pct: 0.6575, mae_ff: 0.06625, rmse_ff: 0.1522}
atlas: {set: "quantum-metal component atlas", layouts: 348, median_pct: 0.9929, mae_ff: 6.621, rmse_ff: 6.631}
twonet_test: {set: "two-net test", layouts: 147, median_pct: 0.5945, mae_ff: 2.391, rmse_ff: 2.391}
random_sealed: {set: "random layouts, sealed", layouts: 110, median_pct: 0.5672, mae_ff: 0.09022, rmse_ff: 0.1225}
synth_sealed: {set: "synthetic layouts, sealed (2-68 conductors)", layouts: 107, median_pct: 0.6995, mae_ff: 0.1108, rmse_ff: 0.4076}
grid_qubit_1: {set: "grid unit, N = 15 (one layout)", layouts: 1, median_pct: 0.3703, mae_ff: 0.05383, rmse_ff: 0.1769}
grid_qubit_5: {set: "grid cluster, N = 35 (one layout)", layouts: 1, median_pct: 1.009, mae_ff: 0.0356, rmse_ff: 0.1449}
latency: # paper_v4 Table 6 (data_local/geomfast/consolidated/timing_summary.csv)
development: {set: "development: 14 TRAIN layouts, one per source, N <= 18 (job 6492904)", layouts: 14, total_median_s: 1.258, total_p95_s: 53.25, geometry_median_s: 0.1202, prior_median_s: 0.95, network_median_s: 0.0116, paired_speedup_vs_E3: 1.14, speedup_ci95: [1.04, 1.26], machine: "GH200 node c02-13 (aarch64), 2 threads, 3 interleaved invocations per layout, median over layouts"}
external: {set: "external: 9 label-free layouts, N = 2-68 (job 6493337)", layouts: 9, total_median_s: 13.08, total_p95_s: 67.88, geometry_median_s: 1.796, prior_median_s: 10.6, network_median_s: 0.0428, paired_speedup_vs_E3: 0.955, speedup_ci95: [0.924, 0.981], machine: "GH200 node c02-13 (aarch64), 2 threads, 3 interleaved invocations per layout, median over layouts"}
GNN-LP:
title: "Fast (learned prior, ~0.1 s; trained component families only): learned prior surrogate + the same 3-seed GNN ensemble"
status: available
tier: fast
arm: full_gnn
role: "fast tier (2026-10-18; agents/learnedprior final arm): the learned prior surrogate t100_norm in place of the two field solves, feeding the frozen P2d full_gnn weights (FINALIZATION_V1, unchanged); no field solve"
ensemble: arithmetic mean of the three full Maxwell matrices
checkpoints:
- {path: data_local/geomfast/runs/p2d_full_gnn_s0/model.pt, sha256: fcb5db5d8237436d546aca0e7463a2f1d4f2e1b18757b2a35eff646c03f3bc5c}
- {path: data_local/geomfast/runs/p2d_full_gnn_s1/model.pt, sha256: 9aef64adfaf99cad35da891a2329d43eac46b2768cd133b5313085891a4fed92}
- {path: data_local/geomfast/runs/p2d_full_gnn_s2/model.pt, sha256: 357c16539967404b0a7eea514ef3ff3dc763c764da192b219d919cb01f697669}
prior: null # no field solve; the learned prior below predicts both priors
learned_prior:
name: surrogate_t100_norm
checkpoint: {path: data_local/learnedprior/runs/t100_norm/model.pt, sha256: 622fc374cc989bcc5569373787996169b8cde1254efdd118baa7252da0d788ce} # = agents/v4test/HASHES.json surrogate_t100_norm
floor_ff: 1.7161304599611071e-09 # smallest positive prior mutual entry of the surrogate's training data (learnedprior/p2d_prepared TRAIN, both levels): the predicted priors are floored here before the networks' log (agents/embedspeed; grid_full underflowed to 0)
code: "geomgnn/lp_surrogate.py = scripts/learnedprior/lp_surrogate.py (sha256 55804c93..., verbatim) + schema_p2d.json (17eba5f7...); geomgnn/learned_prior.py = lp_bench.py Learned.from_gds after the features"
ppg: [1.5, 3.0] # predicts log P at 1.5 and 3 panels per gap (the priors the networks read)
training: "tri-GNN (1.02M parameters, width 128, 4 rounds) on the exact frozen prior of 35,354 geomfast TRAIN layouts + 384k generated layouts (v1 235k + targeted 149k), 60k steps, H200 job 6497375 (agents/learnedprior/STATUS.md, PROTOCOL.md)"
summary: "The accurate tier's networks (same frozen weights) fed by a learned surrogate of the two field-solve priors instead of the solves: about 0.1 s per layout, no field solve. Its accuracy holds only for the 13 component families its learned prior and networks were trained on: interdigital capacitors, paddle capacitors, concentric, star, multipad star, Xmon, TransmonPocket, shielded concentric, two-net layouts, random layouts, procedurally generated layouts (2-7 and 8 or more conductors) and the synthetic multi-conductor campaign layouts (up to 18 conductors). On a component family it never saw, the surrogate's prior error rises 3-17x and the fast tier is worse than the physics prior alone (accuracy_lofo). For a new component type use the accurate tier (GNN)."
warning: "Use Fast only for the 13 component families it was trained on: interdigital capacitors, paddle capacitors, concentric, star, multipad star, Xmon, TransmonPocket, shielded concentric, two-net layouts, random layouts, procedurally generated layouts (2-7 and 8 or more conductors) and synthetic multi-conductor campaign layouts (up to 18 conductors). For any other component type use Accurate (GNN). On a component family held out of its training (learned prior and networks both retrained without it; five families, 4,071 held-out validation layouts) Fast is at 3.11 % (MAE 0.533 / RMSE 0.633 fF), worse than the physics prior alone, 1.55 % (0.166 / 0.205 fF), while Accurate held out the same way is at 0.41 % (0.049 / 0.059 fF); worst on concentric, 11.7 % (6.43 / 7.36 fF). In family (test) Fast is at 0.630 % (0.113 / 0.141 fF) vs Accurate 0.177 % (0.0203 / 0.0250 fF); on the never-seen test components (zero-shot macro) 3.58 % (4.61 / 5.06 fF) vs 0.587 % (1.09 / 1.15 fF). Use Accurate for any number you rely on."
accuracy_note: "Validation split (VAL, development; NOT a test set; agents/learnedprior ladder_val4608, data_local/learnedprior/e2e_phase2_t100/ladder_val4608.json): layout-mean median relative error, MAE and RMSE over mutual entries, 0.1 fF floor. Same VAL, exact prior v1 + the same networks: 0.337 % (0.0750 / 0.0997 fF); accurate tier (v2m): 0.229 % (0.0674 / 0.0792 fF); solver-free network: 1.486 % (0.630 / 0.728 fF)."
accuracy:
all_val: {set: "VAL, all layouts", layouts: 4608, median_pct: 0.8457, mae_ff: 0.4910, rmse_ff: 0.5486}
val_interdigital: {set: "VAL, interdigital", layouts: 1931, median_pct: 0.5782, mae_ff: 0.06004, rmse_ff: 0.06874}
val_paddle: {set: "VAL, paddle", layouts: 1592, median_pct: 0.4224, mae_ff: 0.03180, rmse_ff: 0.03893}
val_star: {set: "VAL, star", layouts: 202, median_pct: 0.7109, mae_ff: 0.1384, rmse_ff: 0.1741}
val_multipad: {set: "VAL, multipad", layouts: 176, median_pct: 1.314, mae_ff: 0.1447, rmse_ff: 0.2437}
val_concentric: {set: "VAL, concentric", layouts: 170, median_pct: 1.778, mae_ff: 1.160, rmse_ff: 1.449}
val_campaign_b3: {set: "VAL, campaign B3 (N up to 18)", layouts: 149, median_pct: 3.926, mae_ff: 0.2707, rmse_ff: 0.8017}
val_twonet: {set: "VAL, two-net", layouts: 98, median_pct: 2.811, mae_ff: 17.14, rmse_ff: 17.14}
val_random: {set: "VAL, random", layouts: 94, median_pct: 2.292, mae_ff: 0.3121, rmse_ff: 0.3924}
val_synth_n_ge_8: {set: "VAL, synthetic N >= 8", layouts: 74, median_pct: 1.417, mae_ff: 0.1628, rmse_ff: 0.7144}
val_synth_n_le_7: {set: "VAL, synthetic N <= 7", layouts: 29, median_pct: 2.422, mae_ff: 0.4819, rmse_ff: 0.8089}
val_transmon_pocket: {set: "VAL, TransmonPocket", layouts: 48, median_pct: 1.290, mae_ff: 0.2917, rmse_ff: 0.4039}
val_shielded_concentric: {set: "VAL, shielded concentric", layouts: 27, median_pct: 3.267, mae_ff: 1.655, rmse_ff: 2.298}
val_xmon: {set: "VAL, Xmon", layouts: 18, median_pct: 1.647, mae_ff: 0.6089, rmse_ff: 0.8233}
accuracy_lofo_note: "Held-out component family (leave-one-family-out over the five original families; VAL only, no test cohort read): the learned prior surrogate AND the networks retrained without the family, scored on that family's held-out VAL layouts (agents/fastlofo_p5/RESULTS_P5.md, row a, protocol PROTOCOL_P5.md frozen 2026-10-18; data_local/fastlofo_p5/metrics/family_table.csv). Layout-mean median relative error %, MAE and RMSE fF over mutual entries, 0.1 fF floor. Same layouts: physics prior alone C^R = R(1.5, 3) 1.554 % (0.166 / 0.205 fF); accurate recipe held out the same way 0.408 % (0.0494 / 0.0591 fF). Mac M5 Pro, 2 threads: fast 0.0835 s median per layout (p90 0.163 s), exact prior + networks 1.891 s (p90 9.53 s)."
accuracy_lofo:
lofo_pooled: {set: "held-out family, pooled VAL", layouts: 4071, median_pct: 3.113, mae_ff: 0.533, rmse_ff: 0.633}
lofo_macro: {set: "held-out family, macro of 5", layouts: 4071, median_pct: 5.413, mae_ff: 1.622, rmse_ff: 1.974}
lofo_interdigital: {set: "interdigital held out", layouts: 1931, median_pct: 2.550, mae_ff: 0.291, rmse_ff: 0.331}
lofo_paddle: {set: "paddle held out", layouts: 1592, median_pct: 2.444, mae_ff: 0.184, rmse_ff: 0.206}
lofo_concentric: {set: "concentric held out", layouts: 170, median_pct: 11.693, mae_ff: 6.431, rmse_ff: 7.358}
lofo_star: {set: "star held out", layouts: 202, median_pct: 1.761, mae_ff: 0.327, rmse_ff: 0.404}
lofo_multipad: {set: "multipad star held out", layouts: 176, median_pct: 8.619, mae_ff: 0.877, rmse_ff: 1.573}
accuracy_test_note: "TEST (paper_v4 Table tab:ladder-test and SM Table tab:ladder-cohort-b, column 'surrogate + networks'; read once): layout-mean median relative error %, MAE and RMSE fF over mutual entries, 0.1 fF floor. Zero-shot macro = mean of the nine component cohorts' layout means (1,878 layouts)."
accuracy_test:
infamily_test: {set: "in-family test (exp6-12 test split)", layouts: 4009, median_pct: 0.630, mae_ff: 0.1130, rmse_ff: 0.1414}
zeroshot_macro: {set: "zero-shot macro (9 component cohorts below)", layouts: 1878, median_pct: 3.576, mae_ff: 4.608, rmse_ff: 5.058}
newfam926: {set: "added families: Xmon, TransmonPocket, shielded concentric (926)", layouts: 926, median_pct: 1.879, mae_ff: 0.795, rmse_ff: 1.103}
twonet_test: {set: "two-net test", layouts: 147, median_pct: 2.453, mae_ff: 15.90, rmse_ff: 15.90}
random_sealed: {set: "random layouts, sealed", layouts: 98, median_pct: 2.136, mae_ff: 0.328, rmse_ff: 0.417}
synth_sealed: {set: "procedurally generated layouts, sealed", layouts: 107, median_pct: 1.449, mae_ff: 0.139, rmse_ff: 0.465}
p3rand36: {set: "held-out TransmonPocket + 3 couplers", layouts: 36, median_pct: 4.014, mae_ff: 1.045, rmse_ff: 2.169}
unusual: {set: "freeform geometries", layouts: 60, median_pct: 5.961, mae_ff: 2.288, rmse_ff: 2.952}
qm116: {set: "unseen qubits and couplers", layouts: 116, median_pct: 4.767, mae_ff: 1.358, rmse_ff: 2.230}
atlas: {set: "quantum-metal component atlas", layouts: 348, median_pct: 5.477, mae_ff: 19.23, rmse_ff: 19.37}
largen: {set: "large N, 8-12 conductors", layouts: 40, median_pct: 4.049, mae_ff: 0.384, rmse_ff: 0.920}
grid_qubit_1: {set: "one-qubit lattice, N = 15", layouts: 1, median_pct: 1.549, mae_ff: 0.548, rmse_ff: 1.649}
grid_qubit_5: {set: "five-qubit lattice, N = 35", layouts: 1, median_pct: 2.471, mae_ff: 0.206, rmse_ff: 0.751}
grid_qubit_9: {set: "nine-qubit lattice, N = 67", layouts: 1, median_pct: 1.408, mae_ff: 0.151, rmse_ff: 0.940}
latency_test_note: "paper_v4 Table tab:ladder-test: end to end per layout from the layout file, median / 90th percentile, every arm in the same pass on two CPUs of an x86 cluster node (AMD EPYC 7513 / 7542, Xeon Silver 4116)."
latency_test:
infamily_test: {set: "in-family test", layouts: 4009, median_s: 0.092, p90_s: 0.26}
zeroshot: {set: "zero-shot cohorts, pooled", layouts: 1878, median_s: 0.30, p90_s: 2.47}
all_test: {set: "sum over all test layouts", layouts: 5902, sum_s: 2635}
latency_note: "agents/learnedprior/STATUS.md (job 6501886, data_local/learnedprior/bench/final_t100): same run, exact prior v1 + the same networks 1.120 s median (p95 31.6 s, max 56.2 s); the surrogate step itself 5.6 ms median."
latency:
development: {set: "development: 14 TRAIN layouts, one per source, N <= 18 (job 6501886)", layouts: 14, total_median_s: 0.1224, total_p95_s: 0.5945, geometry_median_s: 0.1116, prior_median_s: 0.00556, network_median_s: 0.00317, machine: "GH200 node c02-13 CPU (aarch64, shared node), 2 threads, 3 interleaved invocations per layout, median over layouts"}
GNN-fast:
title: "GNN-fast: same network, no physics solve (much less accurate on unseen geometry)"
status: retired # 2026-10-18: superseded by GNN-LP (learned prior) at the same speed; kept, never deleted (restore: status available)
retired: "superseded 2026-10-18 by the fast tier GNN-LP (learned prior + the accurate tier's networks): same speed (test sum 2,635 vs 2,619 s), lower MAE and RMSE on every test cohort and lower median on all but the five- and nine-qubit lattices (2.47 vs 1.95 %, 1.41 vs 1.30 %)"
arm: direct_gnn
role: speed trade-off (paper_v4 solver-free arm; frozen 2026-10-16, FINALIZATION_V1)
ensemble: arithmetic mean of the three full Maxwell matrices
checkpoints:
- {path: data_local/geomfast/runs/p2d_direct_gnn_s0/model.pt, sha256: c3c0b8850d74984ebd6d1d42b63d9f8124bf9029627dda60e7aa0a1bdd9065c9}
- {path: data_local/geomfast/runs/p2d_direct_gnn_s1/model.pt, sha256: 3cc2e71de05e29a18212ad467c664d3219ff853c7522d45c1582b7af4185db25}
- {path: data_local/geomfast/runs/p2d_direct_gnn_s2/model.pt, sha256: 57bc28024fc6c5f0c5db424e313474cd64177b13fb9e3bc3d2044b1a01b8f691}
prior: null # no solve: base = training-median log mutual per pair kind, unbounded head
summary: "The GNN without the physics prior: reads the layout and the geometric embedding only. 5-12x faster, but far less accurate outside the training distribution."
warning: "GNN-fast runs no physics solve and is much less accurate on geometry unlike its training layouts: component atlas median 114.5 % (MAE 53.6 fF, RMSE 56.6 fF), unseen qubits and couplers 9.6 % (MAE 2.02 fF, RMSE 3.28 fF). Use GNN for any number you rely on."
accuracy:
infamily_test: {set: "in-family test (exp6-12 test split)", layouts: 4009, median_pct: 1.219, mae_ff: 0.1872, rmse_ff: 0.229}
unusual: {set: "freeform shapes (OOD-60)", layouts: 60, median_pct: 7.007, mae_ff: 3.358, rmse_ff: 4.731}
newfam926: {set: "added families: Xmon, TransmonPocket, shielded concentric (926 test designs; the GNN trained on 828 others of these families)", layouts: 926, median_pct: 2.381, mae_ff: 0.8933, rmse_ff: 1.194}
qm116: {set: "unseen qubits and couplers (qm116)", layouts: 116, median_pct: 9.572, mae_ff: 2.016, rmse_ff: 3.28}
p3rand36: {set: "held-out TransmonPockets (p3rand36)", layouts: 36, median_pct: 6.208, mae_ff: 1.515, rmse_ff: 3.127}
largen: {set: "large N, 8-12 conductors", layouts: 40, median_pct: 6.089, mae_ff: 0.5381, rmse_ff: 1.324}
atlas: {set: "quantum-metal component atlas", layouts: 348, median_pct: 114.5, mae_ff: 53.57, rmse_ff: 56.6}
twonet_test: {set: "two-net test", layouts: 147, median_pct: 2.724, mae_ff: 17.68, rmse_ff: 17.68}
random_sealed: {set: "random layouts, sealed", layouts: 110, median_pct: 3.282, mae_ff: 0.4278, rmse_ff: 0.5397}
synth_sealed: {set: "synthetic layouts, sealed (2-68 conductors)", layouts: 107, median_pct: 3.862, mae_ff: 0.413, rmse_ff: 1.271}
grid_qubit_1: {set: "grid unit, N = 15 (one layout)", layouts: 1, median_pct: 2.271, mae_ff: 1.163, rmse_ff: 3.212}
grid_qubit_5: {set: "grid cluster, N = 35 (one layout)", layouts: 1, median_pct: 1.948, mae_ff: 0.5229, rmse_ff: 2.736}
latency:
development: {set: "development: 14 TRAIN layouts, one per source, N <= 18 (job 6492904)", layouts: 14, total_median_s: 0.1246, total_p95_s: 0.6876, geometry_median_s: 0.1162, prior_median_s: 0.0, network_median_s: 0.00455, paired_speedup_vs_E3: 12.0, speedup_ci95: [6.58, 21.4], machine: "GH200 node c02-13 (aarch64), 2 threads, 3 interleaved invocations per layout, median over layouts"}
external: {set: "external: 9 label-free layouts, N = 2-68 (job 6493337)", layouts: 9, total_median_s: 1.817, total_p95_s: 6.641, geometry_median_s: 1.799, prior_median_s: 0.0, network_median_s: 0.0188, paired_speedup_vs_E3: 5.28, speedup_ci95: [4.0, 7.08], machine: "GH200 node c02-13 (aarch64), 2 threads, 3 interleaved invocations per layout, median over layouts"}
# ---------------------------------------------------------------- paper Space (paper_v4 PRX; user decision 2026-10-19; deploy/GNN_SPACE.md "Paper Space")
# Built by `python deploy/sync.py --paper` (dry run) / `--paper --push`; gated by deploy/tests/test_paper_space.py. One live selector per arm
# of the paper's ladder (tab:ladder-test) and routes (tab:routes) tables that runs from a layout file on a CPU; every other study
# checkpoint is in the private model repo paper_hub.checkpoints_repo. Numbers: layout-mean median relative error % (MAE / RMSE fF)
# over mutual entries, 0.1 fF floor; VAL = 4,608 validation layouts (development); test = the paper's test benchmark, read once
# (in family 4,009 layouts; zero-shot macro = mean of the nine component cohorts' layout means, 1,878 layouts); time = end to end per
# layout, median / 90th percentile, two CPUs of an x86 cluster node (tab:ladder-test) unless stated.
paper_hub:
space: shanto268/2026PaperModels # reused (same gradio SDK): its earlier contents are archived (archive below)
retired: {date: 2026-11-06, note: "author decision: no dedicated paper Space; the paper's models stay on HF as model repos only (SQuADDS/PlanarCapacitance-paper, SQuADDS/2026PaperModels-*). The Space itself is not deleted or modified; app_paper.py, sync.py --paper and its gate stay working, not deployed"}
title: Planar Capacitance (paper models)
private: true # Spaces stay private (CLAUDE.md rule 11)
space_hardware: cpu-basic # free tier: 2 vCPU, 16 GB RAM
default_arm: model
checkpoints_repo: shanto268/PlanarCap2026PaperCheckpoints # private model repo: every other checkpoint the paper reports + scripts
companion_space: shanto268/PlanarCapacitanceModel # the flagship Space (fast / balanced / accurate, trained on all data); was PlanarCapGNN until 2026-10-20
max_conductors: 70 # signal conductors per layout on this Space (the nine-qubit lattice has 66 + ground)
archive:
space: shanto268/2026PaperModels
commit: 6dc21890fa2d8569225ce8be97cb091b87465d0d # last commit of the earlier draft's 21-model Space (2026-10-10); restore = push this revision
local_copy: data_local/hf_archive/2026PaperModels_6dc21890 # snapshot_download of that revision (100 files, 64 MB)
note: "the earlier draft's model repos (SQuADDS/2026PaperModels-*) are untouched"
deployed: {commit: 53ba4502b9697bd65d69f249d71df65c475c863b, date: 2026-10-20, gate: deploy/tests/results/paper_space_gate_20261020.json,
live: deploy/tests/results/paper_space_live_20261020_examples.json} # 24 arms, 156 MB bundle, RUNNING on cpu-basic, private
text_update: {commit: 34091ebf7be6bbc7ebc37e392b7ae527805cbe97, date: 2026-10-20, gate: deploy/tests/results/paper_space_gate_flagshiplink_20261020.json} # README companion link -> flagship Space; outputs unchanged (48/48 live matrices equal)
docker: {commit: fa7caa49ddd11421b04b44ee2d03fbf4ba632d90, date: 2026-10-20, sdk: docker, bundle: deploy/build_paper_docker/2026PaperModels,
live: deploy/tests/results/paper_space_live_after_20261020.json, rollback: {commit: 34091ebf7be6bbc7ebc37e392b7ae527805cbe97, sdk: gradio,
local_copy: data_local/hf_archive/2026PaperModels_34091ebf}} # agents/embedspeed sped-up exact pipeline + baked numba cache; outputs bitwise = the Gradio build (12 live matrices)
farfield: {commit: e0695caed5075ed8df70ffb0933d8e51d6e3b4cd, date: 2026-10-20, rollback: fa7caa49ddd11421b04b44ee2d03fbf4ba632d90,
gate: deploy/tests/results/paper_space_gate_farfield_20261020.json, live: deploy/tests/results/paper_space_live_farfield_20261020.json} # stack 'corrected' = the corrected prior with substrate far field (agents/farfield/ADOPTION.md)
substrate: {commit: 37fe8236dca1a5a332a4097c5525f4f23e91b0f2, date: 2026-11-03, rollback: e0695caed5075ed8df70ffb0933d8e51d6e3b4cd,
gate: deploy/tests/results/paper_space_gate_substrate_20261103.json, live: deploy/tests/results/paper_space_live_substrate_20261103.json,
offstack: deploy/tests/results/paper_space_live_offstack_20261103.json} # calibrated substrate route (tl_squadds e60c67a): prior arms solve at the user's stack, learned-prior arms rescale + warn, solver-free refuses; rollback = the live farfield commit
prior_stacks: # process-wide prior settings (geomgnn.api.set_prior_stack), installed per request
corrected: &stack_corrected # since 2026-10-20: the corrected prior with substrate far field + compiled exact solver = the model's prior; also flagship_hub
title: "corrected prior with substrate far field"
version: v2m_sf_fc_dipfill_margin_slabfar # v2m + far pairs in different ground pockets: substrate (slab) dipole; same pocket: half-space dipole
patches: [priorv2m_sf] # VALUE-CHANGING: geomgnn/priorv2.py installed(policy='fc', dip=True, margin=True) + geomgnn/farfield.py arm 'b'
mode: exact_compiled
module_sha256: 1a0bdcb71500c7f11e63238afe531c18caabd5c9f6f7aaa73d9b7dcd0ad61f8c
farfield_sha256: 6c877b47c2894fe78a38ad4199b29e450b610c2c45a261c8cf4d53efc1912836 # scripts/farfield/farfield.py (agents/farfield), vendored verbatim
speed_package_sha256: 4fff680f35cb8b098de6c9b1b6fc55780064362ffc829bb47bf69a77454ce38a
value_changing_approved: "user 2026-10-17 (v2m, gnn_models.GNN); user 2026-10-20: adopted arm (b) of agents/farfield (substrate dipole for far pairs in different ground pockets; pre-declared rule met, agents/farfield/RESULTS.md; adoption agents/farfield/ADOPTION.md)"
speed_approved: "user 2026-10-17 (gnn_models.GNN)"
previous: corrected_halfspace # the same stack with the half-space far field (before 2026-10-20); process-wide override: TLSQ_FAR_FIELD=halfspace
corrected_halfspace: # the corrected prior before 2026-10-20 (v2m, half-space dipole for every far pair), for reproduction; not served by any arm
title: "corrected prior (half-space far field, before 2026-10-20)"
version: v2m_fc_dipfill_margin
patches: [priorv2m]
mode: exact_compiled
module_sha256: 1a0bdcb71500c7f11e63238afe531c18caabd5c9f6f7aaa73d9b7dcd0ad61f8c
speed_package_sha256: 4fff680f35cb8b098de6c9b1b6fc55780064362ffc829bb47bf69a77454ce38a
value_changing_approved: "user 2026-10-17 (gnn_models.GNN)"
speed_approved: "user 2026-10-17 (gnn_models.GNN)"
uncorrected: # prior v1, frozen solver (+ the value-preserving rejection-tail fix, bit-identical)
patches: [rejtail_fv]
version2: # prior version 2 (no window-margin widening), frozen solver: the prior of the 'retrained, version 2' ablation
patches: [priorv2]
module_sha256: 1a0bdcb71500c7f11e63238afe531c18caabd5c9f6f7aaa73d9b7dcd0ad61f8c
value_changing_approved: "paper ablation (SM tab:priorv2 'retrained, version 2'), served for reproduction only (paper Space, 2026-10-19)"
uncorrected_certified: # prior v1, certified solve (Rayleigh-Ritz windows with an error estimate)
patches: []
mode: certified
speed_package_sha256: 4fff680f35cb8b098de6c9b1b6fc55780064362ffc829bb47bf69a77454ce38a
speed_approved: "VAL tie on all 13 families (agents/fastprior_pkg/RESULTS.md); paper tab:routes (d)"
_paper_ckpt:
full_gnn: &ck_full_gnn
- {path: data_local/geomfast/runs/p2d_full_gnn_s0/model.pt, sha256: fcb5db5d8237436d546aca0e7463a2f1d4f2e1b18757b2a35eff646c03f3bc5c}
- {path: data_local/geomfast/runs/p2d_full_gnn_s1/model.pt, sha256: 9aef64adfaf99cad35da891a2329d43eac46b2768cd133b5313085891a4fed92}
- {path: data_local/geomfast/runs/p2d_full_gnn_s2/model.pt, sha256: 357c16539967404b0a7eea514ef3ff3dc763c764da192b219d919cb01f697669}
paper_arms:
# ---------------- (d) good physics
model:
title: "The model: extrapolated physics on the corrected prior + residual networks + closure, compiled solver"
paper_name: "model (corrected prior, compiled solver)"
route: "(d) good physics"
kind: gnn
arm: full_gnn
stack: corrected
ppg: [1.5, 3.0]
checkpoints: *ck_full_gnn
summary: "C^R = R(1.5, 3) of the corrected prior (two thin-film solves) corrected by the 3-seed network ensemble (89,858 parameters per seed, bounded residual +-0.2 in log) with a learned row-sum closure: the complete Maxwell matrix. Trained on 35,354 Q3D layouts of 13 families."
paper:
tables: [tab:ladder-test, tab:routes (d), tab:sota, tab:cohorts, tab:grid, tab:circuit]
val: {median_pct: 0.229, mae_ff: 0.0674, rmse_ff: 0.0792}
test_infamily: {median_pct: 0.177, mae_ff: 0.0203, rmse_ff: 0.0250}
test_zeroshot: {median_pct: 0.587, mae_ff: 1.092, rmse_ff: 1.152}
heldout_family: {median_pct: 0.408, mae_ff: 0.049, rmse_ff: 0.059}
time: "0.845 / 5.96 s in family, 1.76 / 18.1 s zero-shot (test, x86, 2 CPUs); 1.16 / 6.3 s (GH200)"
certified:
title: "The same networks, certified variational solve (uncorrected prior)"
paper_name: "the same, engineered solves: certified variational solve"
route: "(d) good physics"
kind: gnn
arm: full_gnn
stack: uncorrected_certified
ppg: [1.5, 3.0]
checkpoints: *ck_full_gnn
summary: "The model's networks on the uncorrected prior solved with Rayleigh-Ritz windows that carry an error estimate; the paper's speed comparison of the solver modes, scored on the validation set against the exact solve of the same prior."
paper:
tables: [tab:routes (d), S-tab:package]
val: {median_pct: 0.3369, mae_ff: 0.0772, rmse_ff: 0.1018, note: "against 0.3369 % (0.0750 / 0.0997 fF) with the exact solve of the same prior: a tie on every family"}
time: "0.83 / 3.4 s (GH200, 2 CPU threads, 14 timing layouts) against 1.16 / 6.3 s compiled"
networks-uncorrected:
title: "Networks on the uncorrected prior"
paper_name: "networks on the uncorrected prior"
route: "(d) good physics"
kind: gnn
arm: full_gnn
stack: uncorrected
ppg: [1.5, 3.0]
checkpoints: *ck_full_gnn
summary: "The model's networks (same weights) on the uncorrected prior (frozen solver); identical to the model for N <= 9 conductors."
paper:
tables: [tab:ladder-test, S-tab:priorv2 (frozen, original)]
val: {median_pct: 0.337, mae_ff: 0.0750, rmse_ff: 0.0997}
test_infamily: {median_pct: 0.177, mae_ff: 0.0203, rmse_ff: 0.0250}
test_zeroshot: {median_pct: 0.593, mae_ff: 1.099, rmse_ff: 1.180}
time: "0.857 / 7.91 s in family, 1.89 / 41.5 s zero-shot (test)"
physics:
title: "Extrapolated physics C^R = R(1.5, 3), corrected prior (no learned parameter)"
paper_name: "extrapolated physics R(1.5,3), corrected prior"
route: "(d) good physics"
kind: physics
stack: corrected
levels: [1.5, 3.0]
self_note: "Richardson-extrapolated self terms of the thin-film solve (no network); the ground self term depends on the chip boundary"
summary: "2 log C(3) - log C(1.5) of the thin-film method of moments on the corrected prior, every entry; zero fitted parameters (the gap share alpha is TRAIN-fitted once)."
paper:
tables: [tab:ladder-test, tab:routes (d)]
val: {median_pct: 1.508, mae_ff: 0.291, rmse_ff: 0.337}
test_infamily: {median_pct: 1.559, mae_ff: 0.1703, rmse_ff: 0.2114}
test_zeroshot: {median_pct: 0.982, mae_ff: 1.050, rmse_ff: 1.128}
heldout_family: {median_pct: 1.554, mae_ff: 0.166, rmse_ff: 0.205}
time: "0.752 / 5.70 s in family, 1.39 / 15.0 s zero-shot (test)"
physics-uncorrected:
title: "Extrapolated physics R(1.5, 3), uncorrected prior"
paper_name: "extrapolated physics R(1.5,3), uncorrected prior"
route: "(d) good physics"
kind: physics
stack: uncorrected
levels: [1.5, 3.0]
self_note: "Richardson-extrapolated self terms of the thin-film solve (no network)"
summary: "The same extrapolation on the uncorrected prior (frozen solver)."
paper:
tables: [tab:ladder-test, S-tab:multifid]
val: {median_pct: 1.615, mae_ff: 0.2993, rmse_ff: 0.3574}
test_infamily: {median_pct: 1.559, mae_ff: 0.1703, rmse_ff: 0.2114}
test_zeroshot: {median_pct: 0.991, mae_ff: 1.057, rmse_ff: 1.157}
time: "0.767 / 7.71 s in family, 1.48 / 39.5 s zero-shot (test)"
single-3:
title: "Single solve, 3 panels per gap (corrected prior)"
paper_name: "single solve, 3 ppg"
route: "(d) good physics"
kind: physics
stack: corrected
levels: [3.0]
self_note: "self terms of the thin-film solve (no network)"
summary: "One thin-film solve at 3 panels per gap, no extrapolation."
paper:
tables: [tab:ladder-test]
test_infamily: {median_pct: 2.577, mae_ff: 0.2547, rmse_ff: 0.2991}
test_zeroshot: {median_pct: 1.755, mae_ff: 3.736, rmse_ff: 3.847}
time: "0.491 / 4.01 s in family, 0.94 / 10.4 s zero-shot (test)"
retrained-corrected:
title: "Networks retrained on the corrected prior"
paper_name: "retrained, corrected"
route: "(d) good physics"
kind: gnn
arm: full_gnn
stack: corrected
ppg: [1.5, 3.0]
checkpoints:
- {path: data_local/priorv2/hpc/runs/p2d_v2m_full_gnn_s0/model.pt, sha256: 04721d7e7f5c6faa7aea352fe1f9bb2261c5cc6835d169957bd3a1a8e61164dd}
- {path: data_local/priorv2/hpc/runs/p2d_v2m_full_gnn_s1/model.pt, sha256: f2352aa25536776e85daecd6681da697497ce9474c3a2391baeb66ec27ea89a5}
- {path: data_local/priorv2/hpc/runs/p2d_v2m_full_gnn_s2/model.pt, sha256: 4257ff6b4a16f69151d04514b71e922dd2b3a998f4ef550693365fd5415fcd8f}
summary: "The model's recipe retrained from scratch on the corrected prior (ablation; the paper keeps the frozen networks, which take the whole gain)."
paper:
tables: [S-tab:priorv2 (retrained, corrected)]
val: {median_pct: 0.218, mae_ff: 0.0774, rmse_ff: 0.0888}
retrained-version2:
title: "Networks retrained on prior version 2"
paper_name: "retrained, version 2"
route: "(d) good physics"
kind: gnn
arm: full_gnn
stack: version2
ppg: [1.5, 3.0]
checkpoints:
- {path: data_local/priorv2/hpc/runs/p2d_v2_full_gnn_s0/model.pt, sha256: 0d436573e10a8d102ece6b58fb2a20d6b50e50f1feb653bc435229725f91c2b3}
- {path: data_local/priorv2/hpc/runs/p2d_v2_full_gnn_s1/model.pt, sha256: a22effd6b33a099bd4e633c188e1f428b626f4b51cbf660d68ab1c63d165a3f6}
- {path: data_local/priorv2/hpc/runs/p2d_v2_full_gnn_s2/model.pt, sha256: e6e93fd00d35836f5c18a960fef59e1cb450c9e52042bcd0fd531b892db9a518}
summary: "The model's recipe retrained on prior version 2 (the corrected prior without the window-margin widening; ablation)."
paper:
tables: [S-tab:priorv2 (retrained, version 2)]
val: {median_pct: 0.230, mae_ff: 0.0853, rmse_ff: 0.0997}
control-no-message-passing:
title: "Control: no message passing"
paper_name: "no-message-passing control"
route: "(d) good physics"
kind: gnn
arm: full_mlp
stack: uncorrected
ppg: [1.5, 3.0]
checkpoints:
- {path: data_local/geomfast/runs/p2d_full_mlp_s0/model.pt, sha256: 38be84e67ca58fe60d797f533e0fe7de355c7bf8e707d73268e19638677f94b6}
- {path: data_local/geomfast/runs/p2d_full_mlp_s1/model.pt, sha256: 37b2fefbac1c8191084464d56c3461687e945f8c76078cbbe3a690553a5ba757}
- {path: data_local/geomfast/runs/p2d_full_mlp_s2/model.pt, sha256: 7efe03578f314ae6516b3f6bd210e61caf25f77d57b05d06d3aef65e56747e9f}
summary: "The model's encoders and heads with no message-passing round (same data, loss, optimiser, seeds), on the uncorrected prior it was trained and scored on."
paper:
tables: [S-app:controls, S-fig:ladder-val]
val: {median_pct: 0.367, mae_ff: 0.084, rmse_ff: 0.109}
control-pair-only:
title: "Control: pair vector only"
paper_name: "pair-only control"
route: "(d) good physics"
kind: gnn
arm: pair_mlp
stack: uncorrected
ppg: [1.5, 3.0]
checkpoints:
- {path: data_local/geomfast/runs/p2d_pair_mlp_s0/model.pt, sha256: 67ee4de766ebdc20295534462929835437e3de437d08b54eb1d705d83d97915a}
- {path: data_local/geomfast/runs/p2d_pair_mlp_s1/model.pt, sha256: 2c84421056598c2d551e570f1e2e20ba55144b3c69cdbb9ee40d3170ab9ae4db}
- {path: data_local/geomfast/runs/p2d_pair_mlp_s2/model.pt, sha256: ddcd72c06be3b3cf6c5860cb4d57cba2e1c38888b6a497086381452382da46dc}
summary: "The no-message-passing control with the conductor vector reduced to the ground flag and the global vector removed: the matched ablation of the geometric embedding (uncorrected prior)."
paper:
tables: [S-app:controls, S-fig:ladder-val]
val: {median_pct: 0.417, mae_ff: 0.126, rmse_ff: 0.157}
prior-generation:
title: "Prior-generation model: extrapolated physics + per-kind offsets + bounded learned correction"
paper_name: "prior-generation model"
route: "baseline"
kind: final
final_model: E3
trust_rule_applies: true
summary: "The baseline of every results table: R(1.5, 3) of the uncorrected prior (alpha 0.52), per-entry-kind offsets and a +-3 % network correction (747,266 parameters per seed), mutual entries only (self terms from row sums). Accepts other substrates (half-space factor). Its trust rule is on by default (box below)."
paper:
tables: [tab:ladder-test, S-tab:earlier-read, tab:sota]
test_infamily: {median_pct: 0.124, mae_ff: 0.0170, rmse_ff: 0.0211}
test_zeroshot: {median_pct: 0.959, mae_ff: 1.914, rmse_ff: 2.014}
time: "not timed in the test pass"
prior-generation-fast:
title: "Prior-generation model, fast mode: R(0.75, 1.5) + per-kind offsets (no network)"
paper_name: "prior-generation model, fast mode"
route: "baseline"
kind: final
final_model: E1
summary: "The coarser Richardson pair of the uncorrected prior (alpha 0.52) plus the prior-generation model's per-kind offsets; no network."
paper:
tables: [Fig. sota (fast mode)]
note: "the paper shows it only in the k-curve of Fig. sota; the numbers under the model panel are the prior-generation evaluation (pooled over entries)"
# ---------------- (c) cheap physics
single-0.375:
title: "Single solve, 0.375 panels per gap (corrected prior)"
paper_name: "single solve, 0.375 ppg"
route: "(c) cheap physics"
kind: physics
stack: corrected
levels: [0.375]
self_note: "self terms of the thin-film solve (no network)"
summary: "One thin-film solve at 0.375 panels per gap."
paper:
tables: [tab:ladder-test]
test_infamily: {median_pct: 8.532, mae_ff: 0.7994, rmse_ff: 0.8871}
test_zeroshot: {median_pct: 7.578, mae_ff: 17.19, rmse_ff: 17.84}
time: "0.081 / 0.30 s in family, 0.11 / 0.73 s zero-shot (test)"
single-0.75:
title: "Single solve, 0.75 panels per gap (corrected prior)"
paper_name: "single solve, 0.75 ppg"
route: "(c) cheap physics"
kind: physics
stack: corrected
levels: [0.75]
self_note: "self terms of the thin-film solve (no network)"
summary: "One thin-film solve at 0.75 panels per gap."
paper:
tables: [tab:ladder-test]
test_infamily: {median_pct: 5.609, mae_ff: 0.5284, rmse_ff: 0.5918}
test_zeroshot: {median_pct: 4.777, mae_ff: 12.00, rmse_ff: 12.33}
time: "0.140 / 0.56 s in family, 0.21 / 1.93 s zero-shot (test)"
single-1.5:
title: "Single solve, 1.5 panels per gap (corrected prior)"
paper_name: "single solve, 1.5 ppg"
route: "(c) cheap physics"
kind: physics
stack: corrected
levels: [1.5]
self_note: "self terms of the thin-film solve (no network)"
summary: "One thin-film solve at 1.5 panels per gap."
paper:
tables: [tab:ladder-test]
test_infamily: {median_pct: 3.575, mae_ff: 0.3441, rmse_ff: 0.3928}
test_zeroshot: {median_pct: 2.871, mae_ff: 7.041, rmse_ff: 7.236}
time: "0.261 / 1.45 s in family, 0.42 / 4.81 s zero-shot (test)"
pair-0.375-0.75:
title: "Richardson pair R(0.375, 0.75) (corrected prior)"
paper_name: "R(0.375,0.75)"
route: "(c) cheap physics"
kind: physics
stack: corrected
levels: [0.375, 0.75]
self_note: "Richardson-extrapolated self terms of the thin-film solve (no network)"
summary: "2 log C(0.75) - log C(0.375): the cheapest Richardson pair."
paper:
tables: [tab:ladder-test, S-tab:multifid]
val: {median_pct: 2.773, mae_ff: 0.8157, rmse_ff: 0.8980, note: "validation on the uncorrected prior"}
test_infamily: {median_pct: 2.558, mae_ff: 0.2601, rmse_ff: 0.3125}
test_zeroshot: {median_pct: 3.740, mae_ff: 9.021, rmse_ff: 9.327}
time: "0.223 / 0.91 s in family, 0.32 / 2.79 s zero-shot (test)"
pair-0.75-1.5:
title: "Richardson pair R(0.75, 1.5) (corrected prior)"
paper_name: "R(0.75,1.5)"
route: "(c) cheap physics"
kind: physics
stack: corrected
levels: [0.75, 1.5]
self_note: "Richardson-extrapolated self terms of the thin-film solve (no network)"
summary: "2 log C(1.5) - log C(0.75): the coarser pair."
paper:
tables: [tab:ladder-test, S-tab:multifid]
val: {median_pct: 1.661, mae_ff: 0.4495, rmse_ff: 0.5077, note: "validation on the uncorrected prior"}
test_infamily: {median_pct: 1.508, mae_ff: 0.1663, rmse_ff: 0.2031}
test_zeroshot: {median_pct: 1.660, mae_ff: 2.450, rmse_ff: 2.575}
time: "0.403 / 2.05 s in family, 0.63 / 6.84 s zero-shot (test)"
cheap-2:
title: "Cheap-physics network, 2 levels (0.375, 0.75)"
paper_name: "cheap-physics network, 2 levels (0.375, 0.75)"
route: "(c) cheap physics"
kind: multifid
stack: corrected
levels: [0.375, 0.75]
checkpoints:
- {path: data_local/multifid/runs/L2_s0/model.pt, sha256: af43c30ac6fc2c34b12310b2dc1ccd91c6daf0342f1849f3601ff2659c6a8d1f}
- {path: data_local/multifid/runs/L2_s1/model.pt, sha256: 58d5815b8c8f3d170b04c89c26e86ca4586eae1b881c16fc8873c5bc8c01a45f}
- {path: data_local/multifid/runs/L2_s2/model.pt, sha256: a82145cb2ecc628db6c6c0328004a2c6a62d24c7aeb1d540ba20bdc23092e1fc}
summary: "Solves at 0.375 and 0.75 panels per gap and a residual network on their extrapolation (3 seeds; the variance head is not served)."
paper:
tables: [tab:ladder-test, tab:routes (c), S-tab:multifid]
val: {median_pct: 0.695, mae_ff: 0.1996, rmse_ff: 0.2397, note: "validation on the uncorrected prior"}
test_infamily: {median_pct: 0.480, mae_ff: 0.0510, rmse_ff: 0.0622}
test_zeroshot: {median_pct: 1.695, mae_ff: 2.591, rmse_ff: 2.779}
time: "0.308 / 1.13 s in family, 0.66 / 5.83 s zero-shot (test)"
cheap-3:
title: "Cheap-physics network, 3 levels (0.375, 0.75, 1.5)"
paper_name: "cheap-physics network, 3 levels (0.375, 0.75, 1.5)"
route: "(c) cheap physics"
kind: multifid
stack: corrected
levels: [0.375, 0.75, 1.5]
checkpoints:
- {path: data_local/multifid/runs/L3_s0/model.pt, sha256: 5aeb0c03f180ea3365ceadf9d30ea0eee559087e684cfc07d237cc2842084877}
- {path: data_local/multifid/runs/L3_s1/model.pt, sha256: e1bb3f6c09477bf434d1319d420ad0aa323f6239494bd29d740e650a3fb4fad0}
- {path: data_local/multifid/runs/L3_s2/model.pt, sha256: 100d0363aab3567324d0f42e6040161b132700148a56473e5d0de0446c83b9da}
summary: "Solves at 0.375, 0.75 and 1.5 panels per gap and a residual network (3 seeds; the variance head is not served)."
paper:
tables: [tab:ladder-test, tab:routes (c), S-tab:multifid]
val: {median_pct: 0.451, mae_ff: 0.1157, rmse_ff: 0.1445, note: "validation on the uncorrected prior"}
test_infamily: {median_pct: 0.275, mae_ff: 0.0307, rmse_ff: 0.0375}
test_zeroshot: {median_pct: 0.914, mae_ff: 1.711, rmse_ff: 1.795}
time: "0.575 / 2.63 s in family, 1.16 / 10.7 s zero-shot (test)"
cheap-2-exponent:
title: "Cheap-physics network, 2 levels, learned exponent"
paper_name: "2 levels, learned exponent"
route: "(c) cheap physics"
kind: multifid
stack: corrected
levels: [0.375, 0.75]
checkpoints:
- {path: data_local/multifid/runs/L2x_s0/model.pt, sha256: 48d83cdf96614f4bc3b2c9ae6d2765c1c985222ab7689e9d848737bc6200fe88}
- {path: data_local/multifid/runs/L2x_s1/model.pt, sha256: 718d4387fc42e79ef1c318c482496fa112b3985adf5be7546976dbf09cec6f5a}
- {path: data_local/multifid/runs/L2x_s2/model.pt, sha256: df0fe917bb6493ebdf41652006023375e75452a7bb18dff56409ca9b5bbc2319}
summary: "The 2-level network predicting a per-entry convergence exponent of the extrapolation instead of the fixed Richardson step."
paper:
tables: [S-tab:multifid]
val: {median_pct: 0.698, mae_ff: 0.2312, rmse_ff: 0.2683, note: "validation on the uncorrected prior; not test-scored"}
cheap-3-exponent:
title: "Cheap-physics network, 3 levels, learned exponent"
paper_name: "3 levels, learned exponent"
route: "(c) cheap physics"
kind: multifid
stack: corrected
levels: [0.375, 0.75, 1.5]
checkpoints:
- {path: data_local/multifid/runs/L3x_s0/model.pt, sha256: 9fa1d9814e7823eff1953337db9ee9a2752caf722fdfee817e28308056c705c9}
- {path: data_local/multifid/runs/L3x_s1/model.pt, sha256: 6a2485022c29af6494c69f344f7e49acf8cc422bd63d1e5d19de611a90aba6a1}
- {path: data_local/multifid/runs/L3x_s2/model.pt, sha256: d88cfaf5e988d85ed5e4408321c676eaf8bd39c8a50c4bcfa6240c3aebf34c4d}
summary: "The 3-level network predicting a per-entry convergence exponent."
paper:
tables: [S-tab:multifid]
val: {median_pct: 0.456, mae_ff: 0.1286, rmse_ff: 0.1573, note: "validation on the uncorrected prior; not test-scored"}
coarser-pair:
title: "The model's recipe on the coarser pair R(0.75, 1.5), retrained"
paper_name: "coarser pair + the model's recipe"
route: "(c) cheap physics"
kind: gnn
arm: full_gnn
stack: corrected
ppg: [0.75, 1.5]
checkpoints:
- {path: data_local/fastprior_p2/runs/p2_s0/model.pt, sha256: d9c01b20823a1c0b473949ad2b3f0977fc8c9f288b1d7737fb9e42f0ef2ee4ab}
- {path: data_local/fastprior_p2/runs/p2_s1/model.pt, sha256: 6dec390290f9d94af700bfd5e4e7e907a56c413d303d777b13eb9d2ce73e650c}
- {path: data_local/fastprior_p2/runs/p2_s2/model.pt, sha256: 0d43b0b5388e06fd49bf5c1b350e0aaa2a8a627d7b5d95972c0062affb1a24b0}
summary: "The model's networks retrained from scratch on R(0.75, 1.5) of the corrected prior: half the prior time; a trade-off point, never test-scored."
paper:
tables: [tab:routes (c), S-fig:ladder-val]
val: {median_pct: 0.343, mae_ff: 0.123, rmse_ff: 0.138, note: "against 0.229 % (0.067 / 0.079 fF) for the model; not test-scored (the physics pair alone: 1.508 % in family, 1.660 % zero-shot)"}
time: "prior 0.430 / 2.29 s against 0.809 / 6.61 s for the model's pair (x86, 2 CPUs, validation)"
# ---------------- (b) learned physics, (a) no physics
fast-model:
title: "The fast model: learned prior surrogate + the model's networks (families of its training data only)"
paper_name: "the fast model (learned prior surrogate + networks)"
route: "(b) learned physics"
kind: gnn
arm: full_gnn
stack: null
checkpoints: *ck_full_gnn
learned_prior:
name: surrogate_t100_norm
checkpoint: {path: data_local/learnedprior/runs/t100_norm/model.pt, sha256: 622fc374cc989bcc5569373787996169b8cde1254efdd118baa7252da0d788ce}
floor_ff: 1.7161304599611071e-09 # smallest positive prior mutual entry of the surrogate's training data (learnedprior/p2d_prepared TRAIN, both levels): the predicted priors are floored here before the networks' log (agents/embedspeed; grid_full underflowed to 0)
summary: "A learned surrogate (1.02M parameters, 385,593 generated layouts labelled by the prior alone + every family's training geometry) predicts the two priors; the model's networks correct them. No field solve."
warning: "Use the fast model only for the 13 component families of its training data (interdigital, paddle, concentric, star, multipad star, Xmon, TransmonPocket, shielded concentric, two-net, random, procedurally generated, synthetic campaign layouts). On a family held out of it, it is behind the physics it distils: 3.113 % (MAE 0.533 / RMSE 0.633 fF) against 1.554 % (0.166 / 0.205 fF). Use the model for a new component type."
paper:
tables: [tab:ladder-test, tab:routes (b), S-tab:sf-fast]
val: {median_pct: 0.846, mae_ff: 0.491, rmse_ff: 0.549}
test_infamily: {median_pct: 0.630, mae_ff: 0.1130, rmse_ff: 0.1414}
test_zeroshot: {median_pct: 3.576, mae_ff: 4.608, rmse_ff: 5.058, note: "families of its surrogate's training data"}
heldout_family: {median_pct: 3.113, mae_ff: 0.533, rmse_ff: 0.633}
time: "0.092 / 0.26 s in family, 0.30 / 2.47 s zero-shot (test); 0.122 / 0.471 s (GH200)"
solver-free:
title: "Solver-free network: the model's network with no prior, geometry only"
paper_name: "solver-free network"
route: "(a) no physics"
kind: gnn
arm: direct_gnn
stack: null
checkpoints:
- {path: data_local/geomfast/runs/p2d_direct_gnn_s0/model.pt, sha256: c3c0b8850d74984ebd6d1d42b63d9f8124bf9029627dda60e7aa0a1bdd9065c9}
- {path: data_local/geomfast/runs/p2d_direct_gnn_s1/model.pt, sha256: 3cc2e71de05e29a18212ad467c664d3219ff853c7522d45c1582b7af4185db25}
- {path: data_local/geomfast/runs/p2d_direct_gnn_s2/model.pt, sha256: 57bc28024fc6c5f0c5db424e313474cd64177b13fb9e3bc3d2044b1a01b8f691}
summary: "The model's network reading the geometric embedding only: base = training-median log mutual per pair kind, unbounded head."
warning: "No field solve: it interpolates in family and loses an order of magnitude outside it (zero-shot macro 17.29 %, MAE 8.934 / RMSE 9.972 fF). Use the model for any number you rely on."
paper:
tables: [tab:ladder-test, tab:routes (a), S-tab:multifid]
val: {median_pct: 1.486, mae_ff: 0.630, rmse_ff: 0.728}
test_infamily: {median_pct: 1.219, mae_ff: 0.1872, rmse_ff: 0.2290}
test_zeroshot: {median_pct: 17.29, mae_ff: 8.934, rmse_ff: 9.972}
time: "0.089 / 0.27 s in family, 0.30 / 2.48 s zero-shot (test); 0.124 s (GH200)"
# ---------------------------------------------------------------- flagship Space (v4; user decisions 2026-10-19/20; deploy/GNN_SPACE.md "Flagship Space")
# Built by `python deploy/sync.py --flagship` (dry run) / `--flagship --push`; gated by deploy/tests/test_flagship_space.py. Three tiers, one
# selector, each the best recipe per speed level RETRAINED ON ALL ON-STACK DATA (45,771 Q3D layouts: TRAIN + VAL + every read test layout
# passing label QC; agents/flagship_v4/{PROTOCOL,RESULTS}.md). These models therefore have NO held-out score of their own: `reference` = the
# paper's held-out numbers for the same recipe (paper_v4; trained on 35,354 TRAIN layouts), `b5` = the one fresh check (campaign450 B5, 39
# layouts of one generated family, N 3-4, never seen by any component; RESULTS.md section 5). Numbers: layout-mean median relative error %
# (MAE / RMSE fF) over mutual entries, 0.1 fF floor. Code path = the paper Space's (geomgnn.paper) through geomgnn.flagship.view.
flagship_hub:
public_link: SQuADDS/planar-capacitance # static Space, printed in the paper; opens the demo top-level (deploy/spaces/SQuADDS-planar-capacitance); never rename
deploy_20261114: # public-demo limits + short card (no paper model); Docker SYSTEM=spaces so HF OAuth works
commit: 142609ed7021861aea2be8ee5f51594b70e5918b # + text in the author's voice, Fast in the selector, Model not Tier, footer (was 4d44ef6)
live: deploy/tests/results/flagship_space_live_limits20261114.json # examples x tiers vs Mac <= 1.1e-3 rel; anon non-example refused; oauth_token call ran
deploy_20261115: # speed (phases A/B), outputs unchanged: concurrency groups, 2 threads, thickness floor 1 um, pinned deps (torch 2.14.1), background preload, slim bundle checkpoints, baked example results, prior-level cache
commit: f76c3bf6cfcd7c499049d7cc3d3ef0258d65a3e1
rollback: {commit: 142609ed7021861aea2be8ee5f51594b70e5918b, local_copy: data_local/hf_archive/PlanarCapacitanceModel_142609ed} # 102 files
prior_level_cache_mb: 256 # TLSQ_PRIOR_LEVEL_CACHE_MB, set in the Space app.py
bake: "18 entries (6 examples x 3 tiers), warm.py bake step 62.9 s, peak RSS 2328 MB; loaded at start (interdigital recomputes bit for bit)"
gate: deploy/tests/results/flagship_space_gate_speed_20261115.json # + demo limits 11/11, examples 9/9, crop 8/8, substrate 3/3, flagship 6/6
live: deploy/tests/results/flagship_space_live_speed_20261115.json # signed in (oauth_token): 12/12, examples bitwise = 2026-11-14 live, vs Mac <= 1.1e-3 rel (8.9e-3 fF)
offstack: deploy/tests/results/flagship_space_live_offstack_speed_20261115.json # bitwise = 2026-11-06 live
coldstart: {before: deploy/tests/results/flagship_space_coldstart_before_speed_20261115.json, after: deploy/tests/results/flagship_space_coldstart_speed_20261115.json}
deploy_20261116: # QCap rename (handoff/squadds_integration/RENAME_QCAP.md): QCap / QCap-Lite / QCap-Fast in the selector, results, /models ids, bake keys and API (model as text, any case); Space card with the new paper title; models and outputs unchanged
commit: 95e723d1f443ef0c758fbfab213248cbb799d0f0
rollback: {commit: f76c3bf6cfcd7c499049d7cc3d3ef0258d65a3e1, local_copy: data_local/hf_archive/PlanarCapacitanceModel_f76c3bf6}
gate: deploy/tests/results/flagship_space_gate_qcap_20261116.json # demo limits 11/11, examples 9/9, crop 8/8, substrate 3/3, ground detection 10/10, flagship 6/6
live: deploy/tests/results/flagship_space_live_qcap_20261116.json # signed in: 12/12, every matrix bitwise = 2026-11-15 live; bake loaded 18; aliases (qcap_fast) work, old names refused
offstack: deploy/tests/results/flagship_space_live_offstack_qcap_20261116.json # 6/6 matrices, routes bitwise = 2026-11-15; warnings equal apart from the names
deploy_20261116_launch: # public launch (author "go", 2026-11-16): About drops the private-preview sentence and the private-Space token comment; card title as confirmed by the author ("QCap: ... device layout"); models and outputs unchanged
commit: 6a50b4ce6b4ffb3bccbc446ea97d9992923d30de # Docker, cpu-basic, RUNNING 181 s after the upload
rollback: {commit: 95e723d1f443ef0c758fbfab213248cbb799d0f0, local_copy: data_local/hf_archive/PlanarCapacitanceModel_95e723d1} # 103 files
gate: deploy/tests/results/flagship_space_gate_launch_20261116.json # demo limits 11/11, examples 9/9, crop 8/8, substrate 3/3, ground detection 10/10, flagship 6/6 (B5 stored x86 worst 6.7e-7 / 5.6e-7 / 4.7e-6)
live: deploy/tests/results/flagship_space_live_launch_20261116.json # signed in (oauth_token): 12/12, every matrix bitwise = flagship_space_live_qcap_20261116.json; bake loaded 18; vs Mac <= 1.1e-3 rel (8.9e-3 fF)
anon: deploy/tests/results/flagship_space_live_launch_anon_20261116.json # no token: 4 examples answer, bitwise = the signed-in record; own upload refused with the sign-in message
deploy_20261116_logo: # QCap logo (handoff/squadds_integration/logo_qcap): compact QCap lockup beside the SQuADDS logo in the app header (title on its own line), QCap badge as the page icon (adaptive SVG link + 32 px PNG favicon_path); models and outputs unchanged
commit: ee162870236871d64b8e441158769ab7caae223f # Docker, cpu-basic, RUNNING 181 s after the upload
rollback: {commit: 6a50b4ce6b4ffb3bccbc446ea97d9992923d30de, local_copy: data_local/hf_archive/PlanarCapacitanceModel_6a50b4ce} # 103 files
gate: deploy/tests/results/flagship_space_gate_logo_20261116.json # demo limits 11/11, examples 9/9, crop 8/8, substrate 3/3, ground detection 10/10, flagship 6/6 (B5 stored x86 worst 6.7e-7 / 5.6e-7 / 4.7e-6)
live: deploy/tests/results/flagship_space_live_logo_20261116.json # signed in (oauth_token): 12/12, every matrix bitwise = flagship_space_live_launch_20261116.json; bake loaded 18; vs Mac <= 1.1e-3 rel (8.9e-3 fF)
anon: deploy/tests/results/flagship_space_live_logo_anon_20261116.json # no token: 4 examples answer, bitwise = the launch record; own upload refused with the sign-in message
deploy_20261116_logocard: # Space card only (handoff/hf_cards/body/PlanarCapacitanceModel.md): QCap horizontal lockup at the top (light / dark pair, HF dark:hidden classes), front matter thumbnail = assets/qcap_og.png (og:image checked); bundle adds assets/ (3 PNGs, sync.FLAGSHIP_ASSETS); app code and outputs unchanged
commit: 1f1a680dc2b0091d7effffa2a94fa8aa4b49de96 # Docker, cpu-basic, RUNNING 181 s after the upload
rollback: {commit: ee162870236871d64b8e441158769ab7caae223f, local_copy: data_local/hf_archive/PlanarCapacitanceModel_ee162870}
gate: deploy/tests/results/flagship_space_gate_logocard_20261116.json # demo limits 11/11, examples 9/9, crop 8/8, substrate 3/3, ground detection 10/10, flagship 6/6 (B5 stored x86 worst 6.7e-7 / 5.6e-7 / 4.7e-6)
live: deploy/tests/results/flagship_space_live_logocard_20261116.json # signed in: 12/12, every matrix bitwise = flagship_space_live_launch_20261116.json; bake loaded 18; vs Mac <= 1.1e-3 rel (8.9e-3 fF)
anon: deploy/tests/results/flagship_space_live_logocard_anon_20261116.json # no token: 4 examples bitwise = the launch record; own upload refused with the sign-in message
visibility_20261116: # flipped public with update_repo_settings(private=False), UTC; revision at the flip
SQuADDS/QCap: {utc: "2026-11-16T20:56:15Z", sha: e4087468129c641cde12f3ea8b2723998da305eb}
SQuADDS/SQuADDS_PlanarCapacitance: {utc: "2026-11-16T20:56:16Z", sha: 7ea40e91b1bc773ce74c2df6595809bd6f501a4d} # then 95a170d7 (MANIFEST.json without the internal repository name)
SQuADDS/SQuADDS_PlanarCapacitance-Synthetic: {utc: "2026-11-16T20:56:16Z", sha: 7319bc5d3f6bd9a5b597b1bc9f43c73022c50cb6}
shanto268/PlanarCapacitanceModel: {utc: "2026-11-16T20:56:17Z", sha: 6a50b4ce6b4ffb3bccbc446ea97d9992923d30de}
SQuADDS/planar-capacitance: {utc: "2026-11-16T20:56:18Z", sha: f905bdf1cc1558b0f13e24646214070bd17ec4ef} # static, opens https://shanto268-planarcapacitancemodel.hf.space top-level
still_private: [SQuADDS/PlanarCapacitance-paper, "SQuADDS/2026PaperModels-* (21 repos)", SQuADDS/SQuADDS_PlanarCapacitance-heldout, shanto268/SQuADDS-PlanarCapacitance-core] # checked after the flip (anonymous 401)
rollback: "update_repo_settings(repo_id, repo_type=..., private=True) per repo"
hf_cards_20261116_launch: # README-only (handoff/hf_cards/): QCap card drops "They go live there when the paper is published ..."; dataset card says the test labels are held out without naming the heldout repo; CHANGELOG / MANIFEST without internal repository names
SQuADDS/QCap: e4087468129c641cde12f3ea8b2723998da305eb # parent 3f5dbecd
SQuADDS/SQuADDS_PlanarCapacitance: 95a170d7d54ca48cd57dc4c187ab97799a75315c # 7ea40e91 (README + CHANGELOG, parent fb40d01a), then 95a170d7 (MANIFEST.json)
hf_cards_20261116_logo: # QCap logo on the Hugging Face pages (light / dark pair with HF's dark:hidden / hidden dark:block classes; verified rendered in both themes)
SQuADDS/QCap: {commit: b2217a5d431c2c89ce3e5c2a80c03d0aa76b1f35, rollback: e4087468129c641cde12f3ea8b2723998da305eb} # README (handoff/hf_cards/QCap.md): QCap horizontal lockup in place of the SQuADDS header image, "Part of SQuADDS" link; + assets/qcap_horizontal_{light,dark}.png
space: shanto268/PlanarCapacitanceModel # the URL is kept; its earlier v3 tiers are archived (archive below)
title: Planar Capacitance Model
private: false # public since 2026-11-16 20:56 UTC (author's launch "go"; visibility_20261116)
space_hardware: cpu-basic # free tier: 2 vCPU, 16 GB RAM
default_arm: accurate
max_conductors: 68 # signal conductors per layout: fewer than 70 with the ground (author 2026-11-14, geomgnn.access.MAX_NETS); the nine-qubit lattice has 66 + ground
paper_models: SQuADDS/PlanarCapacitance-paper # the paper's model (model repo); the paper Space is retired (paper_hub.retired)
checkpoints: data_local/flagship_v4/checkpoints.json # sha256 of every flagship checkpoint; sync.py refuses any other file
results: agents/flagship_v4/RESULTS.md
hf_cards_20261114: # README-only rewrite (re-pushed same day: /ml-paper/ link, then no links to the private paper model) (handoff/hf_cards/), private; package pins unchanged (manifest bb9d0ce / 5d4e094, dataset 13d46953)
SQuADDS/SQuADDS_PlanarCapacitance: c79953313874e9c116ca2f39a414d92025d18484 # physics_prior rename (24de414) + data/legacy_v1 removed
SQuADDS/SQuADDS_PlanarCapacitance-Synthetic: 3179ea7aed6f0ba92b549cc29aa645feea937513
SQuADDS/SQuADDS_PlanarCapacitance-heldout: c538385b3e8d48f4183b68d00e376a06762b0705
SQuADDS/PlanarCapacitance-flagship: a89f30850b638da1f693042b7d7cf14f20b19098 # + balanced/bal_v2m_sf_s{0,1,2}.pt (same sha256 as the Space) and SHA256SUMS
not_pushed: [SQuADDS/PlanarCapacitance-paper, shanto268/PlanarCapacitanceModel] # drafts in handoff/hf_cards/; author 2026-11-14: these stay private
model_repo: # rename 2026-11-16 (handoff/squadds_integration/RENAME_QCAP.md): flagship -> QCap; tier keys below stay internal
id: SQuADDS/QCap # was SQuADDS/PlanarCapacitance-flagship (old URL redirects)
folders_revision: 6dd46d33cd58788222dbf5478ca3550f2888f179 # accurate/ -> qcap/, balanced/ -> qcap-lite/, fast/ -> qcap-fast/; weights and sha256 unchanged
folders: {accurate: qcap/, balanced: qcap-lite/, fast: qcap-fast/}
names: {accurate: QCap, balanced: QCap-Lite, fast: QCap-Fast} # = flagship_tiers.*.paper_name; API ids qcap / qcap-fast (squadds_ml_api)
hf_cards_20261116: # README-only, QCap names (handoff/hf_cards/QCap.md, _links.md, _tiers.md); private repos stay private
SQuADDS/QCap: 25bdbb19ea3a17b2e86b08baea7b320e2d06d3a0
SQuADDS/SQuADDS_PlanarCapacitance: e85ca683991bd3368f8c40d128ae2e7861760dbd
SQuADDS/SQuADDS_PlanarCapacitance-Synthetic: d0058e32c863eee28a2060e87203c58527882320
SQuADDS/README: 007e7d48481eb78894d4f1a0beb2e54764325c4e # org card Space (public)
hf_cards_20261116_title_final: # README-only: title confirmed by the author (QCap: ... device layout); a revert in between (5b6ea8d7 / 84fa356b / 0ecc249c / dc03551a) is superseded
SQuADDS/QCap: 3f5dbecd53e54dcf4c6ee99c95e1b56ccda0d503
SQuADDS/SQuADDS_PlanarCapacitance: fb40d01a135f8d58e97e103e080a2852eab1ff60
SQuADDS/SQuADDS_PlanarCapacitance-Synthetic: 7319bc5d3f6bd9a5b597b1bc9f43c73022c50cb6
SQuADDS/README: 4d2ce9a80422c49a5de2c2d63c0147b2dd9b0247
hf_cards_20261116_title: # README-only: new paper title "QCap: a physics-corrected ... device layout" (package still pins QCap@25bdbb19; weights identical)
SQuADDS/QCap: bd51ff009a0680e2c715d330a6eb7c94c51260b7
SQuADDS/SQuADDS_PlanarCapacitance: 854744a16ae34aa8f66e77768ece39f6c81a0edf
SQuADDS/SQuADDS_PlanarCapacitance-Synthetic: 77df19d3c3e5edfd10c9c7df563ccef755341b7a
SQuADDS/README: 63ce1252d671cbc7b0ab12289d9a7299bab0299f
training_layouts: 45771
archive:
space: shanto268/PlanarCapacitanceModel
commit: 906bc45ae05ce9a1dac796a22c9aef481f079d40 # last commit of the earlier draft's v3 tiers (fast / coarse / fine, 2026-10-10); restore = push this revision
local_copy: data_local/hf_archive/PlanarCapacitanceModel_906bc45a # snapshot_download of that revision (79 files, 27 MB)
note: "the earlier draft's model repos (SQuADDS/2026PaperModels-*) are untouched; the old builder (sync.py without flags) skips this Space"
deployed: {commit: 8db73f7d1bcd7a5aeaf777dfd9f878afa054068a, date: 2026-10-20, gate: deploy/tests/results/flagship_space_gate_20261020.json,
live: deploy/tests/results/flagship_space_live_20261020.json} # 3 tiers, 80.7 MB bundle, RUNNING on cpu-basic, private (first push 79acb1d9, then the Space timings)
docker: {commit: e3e5a832e04ede8088bd91d7c0567cd68330ff51, first_push: c512e691e3117aea79062f89cc2aa1235a9706f0, date: 2026-10-20, sdk: docker, bundle: deploy/build_flagship_docker/PlanarCapacitanceModel,
live: deploy/tests/results/flagship_space_live_after_20261020.json, rollback: {commit: 8db73f7d1bcd7a5aeaf777dfd9f878afa054068a, sdk: gradio,
local_copy: data_local/hf_archive/PlanarCapacitanceModel_8db73f7d}} # agents/embedspeed: F64 embedding default (TLSQ_EMBEDDING_EXACT=1 = exact recipe) + baked numba cache
farfield: {commit: 6f16a03ab7d2f7a41e66261de3c98d5051960290, date: 2026-10-20, rollback: e3e5a832e04ede8088bd91d7c0567cd68330ff51,
gate: deploy/tests/results/flagship_space_gate_farfield_20261020.json, live: deploy/tests/results/flagship_space_live_farfield_20261020.json} # corrected prior with substrate far field + Accurate / Balanced retrained (agents/farfield/ADOPTION.md)
substrate: {commit: 4f5e6036f37b21cf9a24f6b5ad4762cbe1341809, date: 2026-11-03, rollback: 6f16a03ab7d2f7a41e66261de3c98d5051960290,
gate: deploy/tests/results/flagship_space_gate_substrate_20261103.json, live: deploy/tests/results/flagship_space_live_substrate_20261103.json,
offstack: deploy/tests/results/flagship_space_live_offstack_20261103.json} # calibrated substrate route (tl_squadds e60c67a; handoff/squadds_integration/reports/calibrated_substrate.md): Accurate solves the prior at the user's stack, Fast rescales + warns; rollback = the live farfield commit
ui: {commit: 334e8b524f6fb74d4cf56dd5f9718b594d525628, date: 2026-11-13, rollback: b749096deb5ae2d5a8af1dbb1927a8d182221844, source: tl_squadds 3dedf46,
live: deploy/tests/results/flagship_space_live_modelnames_20261113.json, offstack: deploy/tests/results/flagship_space_live_offstack_modelnames_20261113.json} # Fast warning quotes the measured off-stack errors and names the flagship model; physical-check text names models; models and numbers unchanged; rollback = the 2026-11-09 MetBrewer palette (live: flagship_space_live_palette_20261109.json)
prior_stacks:
corrected: *stack_corrected # the corrected prior with substrate far field + compiled exact solver (the same stack as the paper Space's model)
flagship_tiers:
accurate:
title: "QCap (~1 s): the physics prior from two coarse solves, extrapolated and corrected by the networks"
paper_name: "QCap" # display name = the API name (case-insensitive; handoff/squadds_integration/RENAME_QCAP.md)
route: "QCap"
kind: gnn
arm: full_gnn
embedding: {far_k: 64, d_far_um: 100.0} # agents/embedspeed Part A (PROTOCOL.md amendment 2): far-sample decimation F64 of the pair vectors, the fastest arm tying the exact embedding on validation 4,608 and the 1q / 5q / 9q lattices; the paper Space keeps the exact embedding
stack: corrected
ppg: [1.5, 3.0]
checkpoints:
- {name: acc_v2m_sf_s0, path: data_local/flagship_v4/hpc/runs/acc_v2m_sf_s0/model.pt, sha256: cc69c5818f5028c26177159b88239116520abbb290743d6a6c091657e8dbf1c5} # was acc_v2m_s0 56109a61 (half-space far field)
- {name: acc_v2m_sf_s1, path: data_local/flagship_v4/hpc/runs/acc_v2m_sf_s1/model.pt, sha256: 2f88606dbec5c599550d088213a4c37af3b3ad5134ebad2e60519fb534553a57} # was acc_v2m_s1 ae9c4be5 (half-space far field)
- {name: acc_v2m_sf_s2, path: data_local/flagship_v4/hpc/runs/acc_v2m_sf_s2/model.pt, sha256: d13bafc6b2eb7cadcd97a387d791c0210976db1a9d618f9acc1b614e3701daed} # was acc_v2m_s2 0841acc6 (half-space far field)
summary: "Two thin-film method-of-moments solves at two mesh sizes, extrapolated, then corrected entry by entry by a graph neural network (three seeds, 89,858 parameters each, with a bounded correction) and a learned closure for the self terms. The output is the complete Maxwell matrix. Trained on all 45,771 of our Q3D designs."
own: # this release's own numbers; its networks saw every row scored here except B5 (data_local/farfield_adopt/insample.json, acc_v2m_sf ensemble)
insample: {set: "in-sample, all 45,771 training layouts (TRAIN + VAL + the paper's test split; it was trained on them)", median_pct: 0.207, mae_ff: 0.0679, rmse_ff: 0.0770}
reference:
recipe: "the paper's version of this model (same architecture and prior stack at inference; its networks were trained on the 35,354 TRAIN layouts only, on the uncorrected prior; retrained on the corrected prior, as here, they score validation 0.218 % (0.0774 / 0.0888 fF)). These held-out numbers belong to the paper's version, not to this release, whose networks were retrained on every labelled design, the paper's test split included"
val: {median_pct: 0.229, mae_ff: 0.0674, rmse_ff: 0.0792}
test_infamily: {median_pct: 0.177, mae_ff: 0.0203, rmse_ff: 0.0250}
test_zeroshot: {median_pct: 0.587, mae_ff: 1.092, rmse_ff: 1.152}
heldout_family: {median_pct: 0.408, mae_ff: 0.049, rmse_ff: 0.059}
b5: {median_pct: 1.050, mae_ff: 0.132, rmse_ff: 0.188} # retrained 2026-10-20 (substrate far field); before 1.034 % (0.131 / 0.189 fF); B5 priors unchanged (N 3-4), seed-level network change
b5_physics: {name: "physics alone C^R = R(1.5, 3)", median_pct: 1.765, mae_ff: 0.275, rmse_ff: 0.376}
time_space: "interdigital 0.47 s, xmon 0.91 s, 1-qubit lattice (N 15) 19.0 s, 16-conductor layout 9.9 s, 9-qubit lattice (N 67) 154 s (server time, one request, Docker build 2026-10-20; a repeated request is answered from the cache in ~0.01 s)"
time_x86: "0.845 / 5.96 s in family, 1.75 / 18.1 s zero-shot (median / p90; x86 cluster, 2 CPUs); B5 0.99 s median (1.14 s on the 2026-10-20 re-run, Xeon E5-2640 v4)"
balanced:
title: "QCap-Lite (~0.5 s): the same recipe on two coarser solves"
paper_name: "QCap-Lite"
route: "QCap-Lite"
kind: gnn
arm: full_gnn
embedding: {far_k: 64, d_far_um: 100.0} # agents/embedspeed Part A (PROTOCOL.md amendment 2): far-sample decimation F64 of the pair vectors, the fastest arm tying the exact embedding on validation 4,608 and the 1q / 5q / 9q lattices; the paper Space keeps the exact embedding
stack: corrected
ppg: [0.75, 1.5]
checkpoints:
- {name: bal_v2m_sf_s0, path: data_local/flagship_v4/hpc/runs/bal_v2m_sf_s0/model.pt, sha256: aa742ab93b1e3ffcd86a753bda5cc4a68812acbe8140a6b0e311c11d1f36b4b2} # was bal_v2m_s0 50166723 (half-space far field)
- {name: bal_v2m_sf_s1, path: data_local/flagship_v4/hpc/runs/bal_v2m_sf_s1/model.pt, sha256: 486466473eba61c090f4e3fa7988bf880fc183570c5d02bd78f4c1356691f926} # was bal_v2m_s1 5d1f0ffb (half-space far field)
- {name: bal_v2m_sf_s2, path: data_local/flagship_v4/hpc/runs/bal_v2m_sf_s2/model.pt, sha256: 1b55ded04fd624585eeeccf2ab6e4a8e9354a9b96694131e1737dde43ad3ff58} # was bal_v2m_s2 dcd14c61 (half-space far field)
summary: "The QCap recipe on two coarser meshes, about half the solve time, with networks retrained for that pair on all 45,771 of our Q3D designs. About 2x faster than QCap at 1.5-1.7x its error."
own: # data_local/farfield_adopt/insample.json, bal_v2m_sf ensemble
insample: {set: "in-sample, all 45,771 training layouts (TRAIN + VAL + the paper's test split; it was trained on them)", median_pct: 0.300, mae_ff: 0.0872, rmse_ff: 0.0999}
reference:
recipe: "the paper's version: the coarser pair + the model's recipe, trained on the 35,354 TRAIN layouts only (validation only, never test-scored). These numbers belong to the paper's version, not to this release"
val: {median_pct: 0.343, mae_ff: 0.123, rmse_ff: 0.138}
b5: {median_pct: 2.044, mae_ff: 0.258, rmse_ff: 0.348} # retrained 2026-10-20 (substrate far field); before 1.753 % (0.236 / 0.308 fF); B5 priors unchanged (N 3-4), within the earlier seed range 1.79-2.07 %
b5_physics: {name: "physics alone R(0.75, 1.5)", median_pct: 2.681, mae_ff: 0.396, rmse_ff: 0.583}
time_space: "interdigital 0.29 s, xmon 0.51 s, 1-qubit lattice (N 15) 7.7 s, 16-conductor layout 2.4 s, 9-qubit lattice (N 67) 54 s (server time, one request, Docker build 2026-10-20; a repeated request is answered from the cache in ~0.01 s)"
time_x86: "0.490 / 2.31 s in family, 1.00 / 10.0 s zero-shot (median / p90; x86 cluster, 2 CPUs); B5 0.56 s median"
fast:
title: "QCap-Fast (~0.1 s): a learned prior replaces the solves (trained component families only)"
paper_name: "QCap-Fast"
route: "QCap-Fast"
kind: gnn
arm: full_gnn
embedding: {far_k: 64, d_far_um: 100.0} # agents/embedspeed Part A (PROTOCOL.md amendment 2): far-sample decimation F64 of the pair vectors, the fastest arm tying the exact embedding on validation 4,608 and the 1q / 5q / 9q lattices; the paper Space keeps the exact embedding
stack: null
checkpoints:
- {name: acc_v1_s0, path: data_local/flagship_v4/hpc/runs/acc_v1_s0/model.pt, sha256: 9dff20ade26619e7713c2d145dada65e4554a623780290673cfe9452cf9b64ba}
- {name: acc_v1_s1, path: data_local/flagship_v4/hpc/runs/acc_v1_s1/model.pt, sha256: f1ad2ca065dd650ccabcad73a3c6ad2ac1042f4f68bb828893bbdecc47411ffe}
- {name: acc_v1_s2, path: data_local/flagship_v4/hpc/runs/acc_v1_s2/model.pt, sha256: 8cbada211724df5cc05a7b4fdd232607de047b248a43993d3844bd68ff8060c4}
learned_prior:
name: sall_all_s0 # surrogate seed 0 by declaration (PROTOCOL.md section 3, amendment A1)
checkpoint: {name: sall_all_s0, path: data_local/flagship_v4/hpc/sur/sall_all_s0/model.pt, sha256: 132d6f389eda798d31bd5654a96e24cce18fa750134a9fb2a9725eafed339dac}
summary: "A learned prior (1.02M parameters) predicts the two physics solves from the geometry, and the networks correct it. There is no field solve, so the time is the geometric embedding. The learned prior was trained on the physics prior of every Q3D design plus 432,697 generated layouts solved by the prior alone, and the networks on all 45,771 of our Q3D designs."
warning: "Use QCap-Fast only for quick screening of the component families it was trained on (interdigital, paddle, concentric, star, multipad star, Xmon, TransmonPocket, shielded concentric and the generated layouts). On new designs of a trained family it was worse than the physics alone, 3.175 % (MAE 0.583 / RMSE 0.824 fF) against 1.765 % (0.275 / 0.376 fF). In the paper it was also worse than the physics alone on a component family held out of its training, 3.11 % (0.533 / 0.633 fF) against 1.55 % (0.166 / 0.205 fF). Use QCap for anything that matters."
own: # data_local/flagship_v4/scores/tables.json pooled.fast_sall_all_s0_ens (agents/flagship_v4/RESULTS.md section 2)
insample: {set: "in-sample, all 45,771 training layouts (TRAIN + VAL + the paper's test split; it was trained on them)", median_pct: 0.671, mae_ff: 0.353, rmse_ff: 0.395}
reference:
recipe: "the paper's version of the fast model (surrogate t100_norm trained on 385,593 generated layouts + the TRAIN geometry; networks trained on the 35,354 TRAIN layouts only). These held-out numbers belong to the paper's version, not to this release, which was trained on the paper's test split too"
val: {median_pct: 0.846, mae_ff: 0.491, rmse_ff: 0.549}
test_infamily: {median_pct: 0.630, mae_ff: 0.1130, rmse_ff: 0.1414}
test_zeroshot: {median_pct: 3.576, mae_ff: 4.608, rmse_ff: 5.058}
heldout_family: {median_pct: 3.113, mae_ff: 0.533, rmse_ff: 0.633}
heldout_family_physics: {median_pct: 1.554, mae_ff: 0.166, rmse_ff: 0.205}
val_same_surrogate: {median_pct: 0.722, mae_ff: 0.387, rmse_ff: 0.436, note: "this release's surrogate recipe trained without the validation layouts (P6 SUR-ALL) + the paper's TRAIN-only networks, held out (agents/surcov_p6/RESULTS_P6.md)"}
b5: {median_pct: 3.175, mae_ff: 0.583, rmse_ff: 0.824}
b5_physics: {name: "physics alone C^R = R(1.5, 3) (worse than the physics: QCap-Fast is behind it here)", median_pct: 1.765, mae_ff: 0.275, rmse_ff: 0.376}
time_space: "interdigital 0.036 s, xmon 0.11 s, 1-qubit lattice (N 15) 0.82 s, 16-conductor layout 0.49 s, 9-qubit lattice (N 67) 3.9 s (server time, one request, Docker build 2026-10-20; a repeated request is answered from the cache in ~0.01 s)"
time_x86: "0.091 / 0.274 s in family, 0.30 / 2.45 s zero-shot (median / p90; x86 cluster, 2 CPUs); B5 0.24 s median"
squadds_ml_api: # unified SQuADDS ML inference API (handoff/squadds_ml_api/DESIGN.md, source deploy/squadds_ml_api/, build.py)
space: SQuADDS/squadds-ml-inference-api # free cpu-basic (created 2026-04-10, before HF's org paywall; a rebuild stays free, checked 2026-11-14)
deployed: 2026-11-16
commit: 21f3f577d3385a721d2f5d02d706d17a86851bc3 # public launch (author "go" 2026-11-16): qcap / qcap-fast visibility public, card QCap section; squadds 25c5392 (squadds/ml = 0b3a12f); 59/59 local (tests adapted: hidden / gated checks on forced registries, public checks on the released one), --check resolves; live anonymous: /v1/health ok (squadds_commit 25c5392), legacy oracle 11/11, /v1/models lists qcap + qcap-fast, interdigital multipart predict qcap bitwise = examples/qcap.json, qcap-fast max 1.3e-5 fF (3.0e-6 rel, MAE 5.8e-6 / RMSE 8.6e-6 fF; example generated on arm64) (deploy/tests/results/squadds_ml_api_live_launch_20261116.json)
deploy_20261116: 8427cff1c825f4bd8e9350fb7934cb12511c0d68 # rollback of the launch deploy; squadds 7879eba (QCap rename): ids qcap / qcap-fast (hidden, 404 until launch), examples regenerated (matrices bitwise, revision 25bdbb19); 56/56 local, legacy oracle 11/11 live (deploy/tests/results/squadds_ml_api_live_qcap_20261116.json)
deploy_20261115: 8a0986e534f64165d2bbfe849412b05814f2f358 # rollback of the 2026-11-16 deploy; squadds ec4414d (lean/fast prior, bitwise), thread env 2 + torch threads, check_thickness; warm.py uses physics_prior (0dc991fe failed to build on the old name); 56/56 local, 11/11 legacy oracle live, capacitance models 404
previous: a4f1ca1bfd6f0087baa0ab493d3ae04438671321 # 2026-11-14 deploy (a97aa742 + later commits), squadds 04ebd59
rollback: legacy-olivia # tag on 01966faaa06432d425e06895f616db40e6cb858d (Olivia's single-model API); README-only test commit 7c985614
models: {transmon_cross_hamiltonian_inverse: public, qcap: public, qcap-fast: public} # ids renamed 2026-11-16 (were planar_capacitance_flagship / _fast); qcap / qcap-fast public since the launch deploy 21f3f577
collections: {models: SQuADDS/squadds-models-6af8b7da67a22df62af374c6, datasets: SQuADDS/squadds-datasets-6a65152b857f8efa5620880a}
gating: gated = HF token of a SQuADDS org member; needs Space secret HF_TOKEN (org read token) before the capacitance models go gated
# Single source of truth for every released name, link and paper number of the tl-squadds capacitance models.
# Everything else (tlsq_infer model loader, Space apps, model cards, paper_ml/deploy_links.tex, the sync workflow)
# reads this file. Renaming a repo or Space = edit one line here, then run `python deploy/sync.py` (dry run) /
# `python deploy/sync.py --push`.
#
# Paths are relative to the repository root. Metrics: median |relative error| (%) / MAE (fF) / RMSE (fF) over all
# mutual entries, as in the paper (paper_ml/floats/tab_candidates.tex, paper_ml/tables/cost_rows.tex).
version: 2 # 2: corrected prior with substrate far field (2026-10-20, agents/farfield/ADOPTION.md); 1: before
package_version: "0.1.0"
hub:
org: SQuADDS
private: true # create every repo private; flip deliberately when the paper is public
space_hardware: cpu-basic # free tier: 2 vCPU, 16 GB RAM
flagship_space: shanto268/PlanarCapacitanceModel # org Gradio Spaces need a paid HF plan; move to SQuADDS/ by editing this line
# ("flagship" here = the default Space; the tiers are "coarse" / "fine"; since
# 2026-10-10 the fine tier runs on the "flagship prior", PRIOR_SCHEME.md Part II)
flagship_title: Planar Capacitance Model
paper_models_space: shanto268/2026PaperModels
paper_models_title: 2026 Paper Models
collection: "Zero-shot planar capacitance models (2026)"
collection_slug: null # filled in by sync.py --push after the collection is created
code_url: null # public source repository of the training code, when released
paper:
short: "Zero-shot capacitance of planar superconducting layouts"
draft_dir: paper_ml
links_tex: paper_ml/deploy_links.tex
arxiv: null # set when posted; cards and Spaces pick it up
training_stack:
substrate: Si (cryogenic)
eps_r: 11.45
thickness_um: 350
metal: zero-thickness sheet (Q3D thin conductor)
layers: {ground: [1, 0], conductor: [1, 10]}
training_designs: 40153
training_families: [interdigital (CapNInterdigital), paddle, concentric, star, multi-pad star]
eval_sets:
never_seen:
name: never-seen components (Xmon, TransmonPocket, shielded concentric)
designs: 926
entries: all mutual entries
ood60:
name: OOD-60 (procedural near / intermediate / alien shapes)
designs: 60
blend:
weights_gnn_side: {signal-signal: 0.6, signal-ground: 0.7, pad: 0.8} # log-space; rest = charge-corrected prior
note: "pad = any entry involving the third or later signal body (file order), as in the paper"
kinds: invariant # deployment: order-invariant kinds (deploy/MODEL_SELECTION.md amendment 3); "published" = paper file-order kinds
weights_gnn_side_invariant: {conductor-conductor: 0.6, conductor-ground: 0.9} # LOFO on the 5 training families (scripts/blend_lofo.py --kinds invariant)
note_invariant: "conductor-conductor = any pair of floating conductors, conductor-ground = pairs with ground; independent of polygon order"
self_capacitance:
row_sum_eps: 0.0143 # C_ii = sum_j |C_ij| / (1 - eps) (IMPROVE_IDEAS.md, training median)
substrates: # dropdown of the Spaces; the 11 materials of the stack DoE + the training substrate
Si (cryogenic, training): 11.45
sapphire (c-plane, isotropic eq.): 10.34
GaAs: 12.9
fused silica: 3.8
Si (room temperature): 11.7
SiC: 9.7
MgO: 9.8
InP: 12.4
Ge: 16.0
quartz (crystalline): 4.5
diamond: 5.7
LaAlO3: 24.0
artifacts:
charge_gbm: deploy/artifacts/charge_gbm.joblib # built once by `python -m tlsq_infer.charge fit`
charge_gbm_train_cells: data_local/charge_p1/train_cells.parquet
stack_error: deploy/artifacts/stack_error.json # built by `python deploy/stack_error.py fit`
# per-entry error model (ERROR_MODEL.md): NOT ENABLED in the release (user decision 2026-10-08: no intervals until the
# pre-declared v2 is evaluated on the fresh atlas + >7-conductor labels). To enable, set e.g.
# error_model: {all-data-v3s: error_model_fast.joblib, all-data-v3-rich: error_model_accurate.joblib,
# pair-local-fast: error_model_pair-local-fast.joblib}
# (files in deploy/artifacts/, built by `python scripts/error_model.py fit`; tests: deploy/tests/test_error_model.py)
error_model: {}
defaults:
flagship_model: all-data-lite1500 # fast tier = flagship-fast, DEFAULT since 2026-10-10 (user decision; deploy/MODEL_SELECTION.md Results 4)
coarse_model: all-data-v3s # coarse tier (default until 2026-10-10; deploy/MODEL_SELECTION.md, amendment 1 rule)
flagship_accurate_model: all-data-flagship # fine tier (key name kept). 2026-10-10, user-approved: all-data-v3-rich -> all-data-flagship
# (same weights, flagship prior; PRIOR_SCHEME.md section 8.1, deploy/MODEL_SELECTION.md Results 3)
large_layout_model: all-data-flagship # (PRIOR_SCHEME.md Part II): layouts on which the capped coarse prior would
# resolve < 0.75 panels across the smallest gap (outside the training regime; the grids of
# GRID_SCALING.md returned zero entries) are predicted with this flagship-prior model instead
# (predict(large_layouts="route"), reported in model_used / warnings). In-regime layouts unchanged.
paper_space_model: pair-local-fast # this work's headline (paper abstract)
tiers: # shown in the flagship UI; keys = tier names (alias accurate -> fine accepted by predict);
fast: {latency_s: 1, median_pct: 2.67, mae_ff: 0.768, rmse_ff: 1.874, set: "212 held-out designs (OOD-60, random TransmonPocket-3, 6 new Qiskit Metal classes)",
note: "flagship-fast (default): cheap micro-MoM prior (<= 1,500 merged panels; core-graded windows and screened-dipole far pairs beyond 8 conductors) and a network retrained on it. Less accurate than the coarse and fine tiers (large-N 40: 4.96 % vs 2.93 % / 1.63 %), several times faster, and the only tier that runs multi-qubit chips. On this free CPU Space (2026-10-10): 0.2-0.3 s for the Xmon and TransmonPocket examples, 8.5 s for a 200-finger comb, 14 s for a 66-qubit-grid layout, about 4 min for a 780-conductor chip (18 Mac cores: 15 s). PRIOR_SCHEME.md Part III section 15.3."}
# held-out = OOD-60 + pocket3 random-36 + Qiskit Metal new 116, pooled mutual entries
coarse: {latency_s: 4, median_pct: 1.75, mae_ff: 0.534, rmse_ff: 1.173, set: "212 held-out designs (OOD-60, random TransmonPocket-3, 6 new Qiskit Metal classes)",
note: "On this free CPU Space: 1-5 s for single components (measured on the examples, 2026-10-10). Layouts whose coarse mesh would fall below the training resolution (e.g. the TransmonPocket example, long combs) are predicted with the fine tier instead (reported under model used)."}